Decomposing hot-path modules into focused submodules

Signed-off-by: Alexey <247128645+axkurcom@users.noreply.github.com>

docs/Architecture/API/API.md

@@ -13,7 +13,7 @@ API runtime is configured in `[server.api]`.
 | `listen` | `string` (`IP:PORT`) | `0.0.0.0:9091` | API bind address. |
 | `whitelist` | `CIDR[]` | `127.0.0.0/8` | Source IP allowlist. Empty list means allow all. |
 | `auth_header` | `string` | `""` | Exact value for `Authorization` header. Empty disables header auth. |
-| `request_body_limit_bytes` | `usize` | `65536` | Maximum request body size. Must be `> 0`. |
+| `request_body_limit_bytes` | `usize` | `65536` | Maximum request body size. Must be within `[1, 1048576]`. |
 | `minimal_runtime_enabled` | `bool` | `true` | Enables runtime snapshot endpoints requiring ME pool read-lock aggregation. |
 | `minimal_runtime_cache_ttl_ms` | `u64` | `1000` | Cache TTL for minimal snapshots. `0` disables cache; valid range is `[0, 60000]`. |
 | `runtime_edge_enabled` | `bool` | `false` | Enables runtime edge endpoints with cached aggregation payloads. |
@@ -26,7 +26,7 @@ API runtime is configured in `[server.api]`.
 
 Runtime validation for API config:
 - `server.api.listen` must be a valid `IP:PORT`.
-- `server.api.request_body_limit_bytes` must be `> 0`.
+- `server.api.request_body_limit_bytes` must be within `[1, 1048576]`.
 - `server.api.minimal_runtime_cache_ttl_ms` must be within `[0, 60000]`.
 - `server.api.runtime_edge_cache_ttl_ms` must be within `[0, 60000]`.
 - `server.api.runtime_edge_top_n` must be within `[1, 1000]`.
@@ -76,13 +76,14 @@ Requests are processed in this order:
 
 Notes:
 - Whitelist is evaluated against the direct TCP peer IP (`SocketAddr::ip`), without `X-Forwarded-For` support.
-- `Authorization` check is exact string equality against configured `auth_header`.
+- `Authorization` check is exact constant-time byte equality against configured `auth_header`.
 
 ## Endpoint Matrix
 
 | Method | Path | Body | Success | `data` contract |
 | --- | --- | --- | --- | --- |
 | `GET` | `/v1/health` | none | `200` | `HealthData` |
+| `GET` | `/v1/health/ready` | none | `200` or `503` | `HealthReadyData` |
 | `GET` | `/v1/system/info` | none | `200` | `SystemInfoData` |
 | `GET` | `/v1/runtime/gates` | none | `200` | `RuntimeGatesData` |
 | `GET` | `/v1/runtime/initialization` | none | `200` | `RuntimeInitializationData` |
@@ -102,13 +103,50 @@ Notes:
 | `GET` | `/v1/runtime/me-selftest` | none | `200` | `RuntimeMeSelftestData` |
 | `GET` | `/v1/runtime/connections/summary` | none | `200` | `RuntimeEdgeConnectionsSummaryData` |
 | `GET` | `/v1/runtime/events/recent` | none | `200` | `RuntimeEdgeEventsData` |
+| `GET` | `/v1/stats/users/active-ips` | none | `200` | `UserActiveIps[]` |
 | `GET` | `/v1/stats/users` | none | `200` | `UserInfo[]` |
 | `GET` | `/v1/users` | none | `200` | `UserInfo[]` |
-| `POST` | `/v1/users` | `CreateUserRequest` | `201` | `CreateUserResponse` |
+| `POST` | `/v1/users` | `CreateUserRequest` | `201` or `202` | `CreateUserResponse` |
 | `GET` | `/v1/users/{username}` | none | `200` | `UserInfo` |
-| `PATCH` | `/v1/users/{username}` | `PatchUserRequest` | `200` | `UserInfo` |
-| `DELETE` | `/v1/users/{username}` | none | `200` | `string` (deleted username) |
-| `POST` | `/v1/users/{username}/rotate-secret` | `RotateSecretRequest` or empty body | `404` | `ErrorResponse` (`not_found`, current runtime behavior) |
+| `PATCH` | `/v1/users/{username}` | `PatchUserRequest` | `200` or `202` | `UserInfo` |
+| `DELETE` | `/v1/users/{username}` | none | `200` or `202` | `DeleteUserResponse` |
+| `POST` | `/v1/users/{username}/rotate-secret` | `RotateSecretRequest` or empty body | `200` or `202` | `CreateUserResponse` |
+| `POST` | `/v1/users/{username}/reset-quota` | empty body | `200` | `ResetUserQuotaResponse` |
+
+## Endpoint Behavior
+
+| Endpoint | Function |
+| --- | --- |
+| `GET /v1/health` | Returns basic API liveness and current `read_only` flag. |
+| `GET /v1/health/ready` | Returns readiness based on admission state and upstream health; returns `503` when not ready. |
+| `GET /v1/system/info` | Returns binary/build metadata, process uptime, config path/hash, and reload counters. |
+| `GET /v1/runtime/gates` | Returns admission, ME readiness, fallback/reroute, and startup gate state. |
+| `GET /v1/runtime/initialization` | Returns startup progress, ME initialization status, and per-component timeline. |
+| `GET /v1/limits/effective` | Returns effective timeout, upstream, ME, unique-IP, and TCP policy values after config defaults/resolution. |
+| `GET /v1/security/posture` | Returns current API/security/telemetry posture flags. |
+| `GET /v1/security/whitelist` | Returns configured API whitelist CIDRs. |
+| `GET /v1/stats/summary` | Returns compact core counters and classed failure counters. |
+| `GET /v1/stats/zero/all` | Returns zero-cost core, upstream, ME, pool, and desync counters. |
+| `GET /v1/stats/upstreams` | Returns upstream zero counters and, when enabled/available, runtime upstream health rows. |
+| `GET /v1/stats/minimal/all` | Returns cached minimal ME writer/DC/runtime/network-path snapshot. |
+| `GET /v1/stats/me-writers` | Returns cached ME writer coverage and per-writer status rows. |
+| `GET /v1/stats/dcs` | Returns cached per-DC endpoint/writer/load status rows. |
+| `GET /v1/runtime/me_pool_state` | Returns active/warm/pending/draining generation state, writer contour/health, and refill state. |
+| `GET /v1/runtime/me_quality` | Returns ME lifecycle counters, route-drop counters, family states, drain gate, and per-DC RTT/coverage. |
+| `GET /v1/runtime/upstream_quality` | Returns upstream policy/counters plus runtime upstream health rows when available. |
+| `GET /v1/runtime/nat_stun` | Returns NAT/STUN runtime flags, configured/live STUN servers, reflection cache, and backoff. |
+| `GET /v1/runtime/me-selftest` | Returns ME self-test state for KDF, time skew, IP family, PID, and SOCKS BND observations. |
+| `GET /v1/runtime/connections/summary` | Returns runtime-edge connection totals and top-N users by connections/throughput. |
+| `GET /v1/runtime/events/recent` | Returns recent API/runtime event records with optional `limit` query. |
+| `GET /v1/stats/users/active-ips` | Returns users that currently have non-empty active source-IP lists. |
+| `GET /v1/stats/users` | Alias of `GET /v1/users`; returns disk-first user views with runtime lag flag. |
+| `GET /v1/users` | Returns disk-first user views sorted by username. |
+| `POST /v1/users` | Creates a user and returns the effective user view plus secret. |
+| `GET /v1/users/{username}` | Returns one disk-first user view or `404` when absent. |
+| `PATCH /v1/users/{username}` | Updates selected per-user fields with JSON Merge Patch semantics. |
+| `DELETE /v1/users/{username}` | Deletes one user and related per-user access-map entries. |
+| `POST /v1/users/{username}/rotate-secret` | Rotates one user's secret and returns the effective secret. |
+| `POST /v1/users/{username}/reset-quota` | Resets one user's runtime quota counter and persists quota state. |
 
 ## Common Error Codes
 
@@ -118,7 +156,7 @@ Notes:
 | `401` | `unauthorized` | Missing/invalid `Authorization` when `auth_header` is configured. |
 | `403` | `forbidden` | Source IP is not allowed by whitelist. |
 | `403` | `read_only` | Mutating endpoint called while `read_only=true`. |
-| `404` | `not_found` | Unknown route, unknown user, or unsupported sub-route (including current `rotate-secret` route). |
+| `404` | `not_found` | Unknown route, unknown user, or unsupported sub-route. |
 | `405` | `method_not_allowed` | Unsupported method for `/v1/users/{username}` route shape. |
 | `409` | `revision_conflict` | `If-Match` revision mismatch. |
 | `409` | `user_exists` | User already exists on create. |
@@ -132,11 +170,12 @@ Notes:
 | Case | Behavior |
 | --- | --- |
 | Path matching | Exact match on `req.uri().path()`. Query string does not affect route matching. |
-| Trailing slash | Not normalized. Example: `/v1/users/` is `404`. |
+| Trailing slash | Trimmed for route matching when path length is greater than 1. Example: `/v1/users/` matches `/v1/users`. |
 | Username route with extra slash | `/v1/users/{username}/...` is not treated as user route and returns `404`. |
 | `PUT /v1/users/{username}` | `405 method_not_allowed`. |
 | `POST /v1/users/{username}` | `404 not_found`. |
-| `POST /v1/users/{username}/rotate-secret` | `404 not_found` in current release due route matcher limitation. |
+| `POST /v1/users/{username}/rotate-secret/` | Trailing slash is trimmed and the route matches `rotate-secret`. |
+| `POST /v1/users/{username}/reset-quota/` | Trailing slash is trimmed and the route matches `reset-quota`. |
 
 ## Body and JSON Semantics
 
@@ -146,7 +185,7 @@ Notes:
 - Invalid JSON returns `400 bad_request` (`Invalid JSON body`).
 - `Content-Type` is not required for JSON parsing.
 - Unknown JSON fields are ignored by deserialization.
-- `PATCH` updates only provided fields and does not support explicit clearing of optional fields.
+- `PATCH` uses JSON Merge Patch semantics for optional per-user fields: omitted means unchanged, explicit `null` removes the config entry, and a non-null value sets it.
 - `If-Match` supports both quoted and unquoted values; surrounding whitespace is trimmed.
 
 ## Query Parameters
@@ -166,17 +205,21 @@ Notes:
 | `max_tcp_conns` | `usize` | no | Per-user concurrent TCP limit. |
 | `expiration_rfc3339` | `string` | no | RFC3339 expiration timestamp. |
 | `data_quota_bytes` | `u64` | no | Per-user traffic quota. |
+| `rate_limit_up_bps` | `u64` | no | Per-user upload rate limit in bytes per second. |
+| `rate_limit_down_bps` | `u64` | no | Per-user download rate limit in bytes per second. |
 | `max_unique_ips` | `usize` | no | Per-user unique source IP limit. |
 
 ### `PatchUserRequest`
 | Field | Type | Required | Description |
 | --- | --- | --- | --- |
 | `secret` | `string` | no | Exactly 32 hex chars. |
-| `user_ad_tag` | `string` | no | Exactly 32 hex chars. |
-| `max_tcp_conns` | `usize` | no | Per-user concurrent TCP limit. |
-| `expiration_rfc3339` | `string` | no | RFC3339 expiration timestamp. |
-| `data_quota_bytes` | `u64` | no | Per-user traffic quota. |
-| `max_unique_ips` | `usize` | no | Per-user unique source IP limit. |
+| `user_ad_tag` | `string|null` | no | Exactly 32 hex chars; `null` removes the per-user ad tag. |
+| `max_tcp_conns` | `usize|null` | no | Per-user concurrent TCP limit; `null` removes the per-user override. |
+| `expiration_rfc3339` | `string|null` | no | RFC3339 expiration timestamp; `null` removes the expiration. |
+| `data_quota_bytes` | `u64|null` | no | Per-user traffic quota; `null` removes the per-user quota. |
+| `rate_limit_up_bps` | `u64|null` | no | Per-user upload rate limit in bytes per second; `null` removes the upload direction limit. |
+| `rate_limit_down_bps` | `u64|null` | no | Per-user download rate limit in bytes per second; `null` removes the download direction limit. |
+| `max_unique_ips` | `usize|null` | no | Per-user unique source IP limit; `null` removes the per-user override. |
 
 ### `access.user_source_deny` via API
 - In current API surface, per-user deny-list is **not** exposed as a dedicated field in `CreateUserRequest` / `PatchUserRequest`.
@@ -198,7 +241,7 @@ bob = ["198.51.100.42/32"]
 | --- | --- | --- | --- |
 | `secret` | `string` | no | Exactly 32 hex chars. If missing, generated automatically. |
 
-Note: the request contract is defined, but the corresponding route currently returns `404` (see routing edge cases).
+An empty request body is accepted and generates a new secret automatically.
 
 ## Response Data Contracts
 
@@ -208,15 +251,33 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `status` | `string` | Always `"ok"`. |
 | `read_only` | `bool` | Mirrors current API `read_only` mode. |
 
+### `HealthReadyData`
+| Field | Type | Description |
+| --- | --- | --- |
+| `ready` | `bool` | `true` when admission is open and at least one upstream is healthy. |
+| `status` | `string` | `"ready"` or `"not_ready"`. |
+| `reason` | `string?` | `admission_closed` or `no_healthy_upstreams` when not ready. |
+| `admission_open` | `bool` | Current admission-gate state. |
+| `healthy_upstreams` | `usize` | Number of healthy upstream entries. |
+| `total_upstreams` | `usize` | Number of configured upstream entries. |
+
 ### `SummaryData`
 | Field | Type | Description |
 | --- | --- | --- |
 | `uptime_seconds` | `f64` | Process uptime in seconds. |
 | `connections_total` | `u64` | Total accepted client connections. |
 | `connections_bad_total` | `u64` | Failed/invalid client connections. |
+| `connections_bad_by_class` | `ClassCount[]` | Failed/invalid connections grouped by class. |
+| `handshake_failures_by_class` | `ClassCount[]` | Handshake failures grouped by class. |
 | `handshake_timeouts_total` | `u64` | Handshake timeout count. |
 | `configured_users` | `usize` | Number of configured users in config. |
 
+#### `ClassCount`
+| Field | Type | Description |
+| --- | --- | --- |
+| `class` | `string` | Failure class label. |
+| `total` | `u64` | Counter value for this class. |
+
 ### `SystemInfoData`
 | Field | Type | Description |
 | --- | --- | --- |
@@ -241,7 +302,12 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `conditional_cast_enabled` | `bool` | Whether conditional ME admission logic is enabled (`general.use_middle_proxy`). |
 | `me_runtime_ready` | `bool` | Current ME runtime readiness status used for conditional gate decisions. |
 | `me2dc_fallback_enabled` | `bool` | Whether ME -> direct fallback is enabled. |
+| `me2dc_fast_enabled` | `bool` | Whether fast ME -> direct fallback is enabled. |
 | `use_middle_proxy` | `bool` | Current transport mode preference. |
+| `route_mode` | `string` | Current route mode label from route runtime controller. |
+| `reroute_active` | `bool` | `true` when ME fallback currently routes new sessions to Direct-DC. |
+| `reroute_to_direct_at_epoch_secs` | `u64?` | Unix timestamp when current direct reroute began. |
+| `reroute_reason` | `string?` | `startup_direct_fallback`, `fast_not_ready_fallback`, or `strict_grace_fallback` while reroute is active. |
 | `startup_status` | `string` | Startup status (`pending`, `initializing`, `ready`, `failed`, `skipped`). |
 | `startup_stage` | `string` | Current startup stage identifier. |
 | `startup_progress_pct` | `f64` | Startup progress percentage (`0..100`). |
@@ -292,11 +358,13 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `upstream` | `EffectiveUpstreamLimits` | Effective upstream connect/retry limits. |
 | `middle_proxy` | `EffectiveMiddleProxyLimits` | Effective ME pool/floor/reconnect limits. |
 | `user_ip_policy` | `EffectiveUserIpPolicyLimits` | Effective unique-IP policy mode/window. |
+| `user_tcp_policy` | `EffectiveUserTcpPolicyLimits` | Effective per-user TCP connection policy. |
 
 #### `EffectiveTimeoutLimits`
 | Field | Type | Description |
 | --- | --- | --- |
 | `client_handshake_secs` | `u64` | Client handshake timeout. |
+| `client_first_byte_idle_secs` | `u64` | First-byte idle timeout before protocol classification. |
 | `tg_connect_secs` | `u64` | Upstream Telegram connect timeout. |
 | `client_keepalive_secs` | `u64` | Client keepalive interval. |
 | `client_ack_secs` | `u64` | ACK timeout. |
@@ -335,13 +403,20 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `writer_pick_mode` | `string` | Writer picker mode (`sorted_rr`, `p2c`). |
 | `writer_pick_sample_size` | `u8` | Candidate sample size for `p2c` picker mode. |
 | `me2dc_fallback` | `bool` | Effective ME -> direct fallback flag. |
+| `me2dc_fast` | `bool` | Effective fast fallback flag. |
 
 #### `EffectiveUserIpPolicyLimits`
 | Field | Type | Description |
 | --- | --- | --- |
+| `global_each` | `usize` | Global per-user unique-IP limit applied when no per-user override exists. |
 | `mode` | `string` | Unique-IP policy mode (`active_window`, `time_window`, `combined`). |
 | `window_secs` | `u64` | Time window length used by unique-IP policy. |
 
+#### `EffectiveUserTcpPolicyLimits`
+| Field | Type | Description |
+| --- | --- | --- |
+| `global_each` | `usize` | Global per-user concurrent TCP limit applied when no per-user override exists. |
+
 ### `SecurityPostureData`
 | Field | Type | Description |
 | --- | --- | --- |
@@ -445,6 +520,8 @@ Note: the request contract is defined, but the corresponding route currently ret
 | --- | --- | --- |
 | `counters` | `RuntimeMeQualityCountersData` | Key ME lifecycle/error counters. |
 | `route_drops` | `RuntimeMeQualityRouteDropData` | Route drop counters by reason. |
+| `family_states` | `RuntimeMeQualityFamilyStateData[]` | Per-family ME route/recovery state rows. |
+| `drain_gate` | `RuntimeMeQualityDrainGateData` | Current ME drain-gate decision state. |
 | `dc_rtt` | `RuntimeMeQualityDcRttData[]` | Per-DC RTT and writer coverage rows. |
 
 #### `RuntimeMeQualityCountersData`
@@ -466,6 +543,24 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `queue_full_base_total` | `u64` | Route drops in base-queue path. |
 | `queue_full_high_total` | `u64` | Route drops in high-priority queue path. |
 
+#### `RuntimeMeQualityFamilyStateData`
+| Field | Type | Description |
+| --- | --- | --- |
+| `family` | `string` | Address family label. |
+| `state` | `string` | Current family state label. |
+| `state_since_epoch_secs` | `u64` | Unix timestamp when current state began. |
+| `suppressed_until_epoch_secs` | `u64?` | Unix timestamp until suppression remains active. |
+| `fail_streak` | `u32` | Consecutive failure count. |
+| `recover_success_streak` | `u32` | Consecutive recovery success count. |
+
+#### `RuntimeMeQualityDrainGateData`
+| Field | Type | Description |
+| --- | --- | --- |
+| `route_quorum_ok` | `bool` | Whether route quorum condition allows drain. |
+| `redundancy_ok` | `bool` | Whether redundancy condition allows drain. |
+| `block_reason` | `string` | Current drain block reason label. |
+| `updated_at_epoch_secs` | `u64` | Unix timestamp of the latest gate update. |
+
 #### `RuntimeMeQualityDcRttData`
 | Field | Type | Description |
 | --- | --- | --- |
@@ -728,11 +823,24 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `uptime_seconds` | `f64` | Process uptime. |
 | `connections_total` | `u64` | Total accepted connections. |
 | `connections_bad_total` | `u64` | Failed/invalid connections. |
+| `connections_bad_by_class` | `ClassCount[]` | Failed/invalid connections grouped by class. |
+| `handshake_failures_by_class` | `ClassCount[]` | Handshake failures grouped by class. |
 | `handshake_timeouts_total` | `u64` | Handshake timeouts. |
+| `accept_permit_timeout_total` | `u64` | Listener admission permit acquisition timeouts. |
 | `configured_users` | `usize` | Configured user count. |
 | `telemetry_core_enabled` | `bool` | Core telemetry toggle. |
 | `telemetry_user_enabled` | `bool` | User telemetry toggle. |
 | `telemetry_me_level` | `string` | ME telemetry level (`off|normal|verbose`). |
+| `conntrack_control_enabled` | `bool` | Whether conntrack control is enabled by policy. |
+| `conntrack_control_available` | `bool` | Whether conntrack control backend is currently available. |
+| `conntrack_pressure_active` | `bool` | Current conntrack pressure flag. |
+| `conntrack_event_queue_depth` | `u64` | Current conntrack close-event queue depth. |
+| `conntrack_rule_apply_ok` | `bool` | Last conntrack rule application state. |
+| `conntrack_delete_attempt_total` | `u64` | Conntrack delete attempts. |
+| `conntrack_delete_success_total` | `u64` | Successful conntrack deletes. |
+| `conntrack_delete_not_found_total` | `u64` | Conntrack delete misses. |
+| `conntrack_delete_error_total` | `u64` | Conntrack delete errors. |
+| `conntrack_close_event_drop_total` | `u64` | Dropped conntrack close events. |
 
 #### `ZeroUpstreamData`
 | Field | Type | Description |
@@ -819,6 +927,24 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `route_drop_queue_full_total` | `u64` | Route drops due to full queue (total). |
 | `route_drop_queue_full_base_total` | `u64` | Route drops in base queue mode. |
 | `route_drop_queue_full_high_total` | `u64` | Route drops in high queue mode. |
+| `d2c_batches_total` | `u64` | ME D->C batch flushes. |
+| `d2c_batch_frames_total` | `u64` | ME D->C frames included in batches. |
+| `d2c_batch_bytes_total` | `u64` | ME D->C payload bytes included in batches. |
+| `d2c_flush_reason_queue_drain_total` | `u64` | Flushes caused by queue drain. |
+| `d2c_flush_reason_batch_frames_total` | `u64` | Flushes caused by frame-count batch limit. |
+| `d2c_flush_reason_batch_bytes_total` | `u64` | Flushes caused by byte-count batch limit. |
+| `d2c_flush_reason_max_delay_total` | `u64` | Flushes caused by max-delay budget. |
+| `d2c_flush_reason_ack_immediate_total` | `u64` | Flushes caused by immediate ACK policy. |
+| `d2c_flush_reason_close_total` | `u64` | Flushes caused by close path. |
+| `d2c_data_frames_total` | `u64` | ME D->C data frames. |
+| `d2c_ack_frames_total` | `u64` | ME D->C ACK frames. |
+| `d2c_payload_bytes_total` | `u64` | ME D->C payload bytes. |
+| `d2c_write_mode_coalesced_total` | `u64` | Coalesced D->C writes. |
+| `d2c_write_mode_split_total` | `u64` | Split D->C writes. |
+| `d2c_quota_reject_pre_write_total` | `u64` | D->C quota rejects before write. |
+| `d2c_quota_reject_post_write_total` | `u64` | D->C quota rejects after write. |
+| `d2c_frame_buf_shrink_total` | `u64` | D->C frame-buffer shrink operations. |
+| `d2c_frame_buf_shrink_bytes_total` | `u64` | Bytes released by D->C frame-buffer shrink operations. |
 | `socks_kdf_strict_reject_total` | `u64` | SOCKS KDF strict rejects. |
 | `socks_kdf_compat_fallback_total` | `u64` | SOCKS KDF compat fallbacks. |
 | `endpoint_quarantine_total` | `u64` | Endpoint quarantine activations. |
@@ -978,6 +1104,8 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `required_writers` | `usize` | Required writers based on current floor policy. |
 | `alive_writers` | `usize` | Writers currently alive. |
 | `coverage_pct` | `f64` | `alive_writers / required_writers * 100`. |
+| `fresh_alive_writers` | `usize` | Alive writers that match freshness requirements. |
+| `fresh_coverage_pct` | `f64` | `fresh_alive_writers / required_writers * 100`. |
 
 #### `MeWriterStatus`
 | Field | Type | Description |
@@ -992,6 +1120,12 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `bound_clients` | `usize` | Number of currently bound clients. |
 | `idle_for_secs` | `u64?` | Idle age in seconds if idle. |
 | `rtt_ema_ms` | `f64?` | RTT exponential moving average. |
+| `matches_active_generation` | `bool` | Whether this writer belongs to the active pool generation. |
+| `in_desired_map` | `bool` | Whether this writer's endpoint remains in desired topology. |
+| `allow_drain_fallback` | `bool` | Whether drain fallback is allowed for this writer. |
+| `drain_started_at_epoch_secs` | `u64?` | Unix timestamp when drain started. |
+| `drain_deadline_epoch_secs` | `u64?` | Unix timestamp of drain deadline. |
+| `drain_over_ttl` | `bool` | Whether drain has exceeded its TTL. |
 
 ### `DcStatusData`
 | Field | Type | Description |
@@ -1016,6 +1150,8 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `floor_capped` | `bool` | `true` when computed floor target was capped by active limits. |
 | `alive_writers` | `usize` | Alive writers in this DC. |
 | `coverage_pct` | `f64` | `alive_writers / required_writers * 100`. |
+| `fresh_alive_writers` | `usize` | Fresh alive writers in this DC. |
+| `fresh_coverage_pct` | `f64` | `fresh_alive_writers / required_writers * 100`. |
 | `rtt_ms` | `f64?` | Aggregated RTT for DC. |
 | `load` | `usize` | Active client sessions bound to this DC. |
 
@@ -1029,10 +1165,13 @@ Note: the request contract is defined, but the corresponding route currently ret
 | Field | Type | Description |
 | --- | --- | --- |
 | `username` | `string` | Username. |
+| `in_runtime` | `bool` | Whether current runtime config already contains this user. |
 | `user_ad_tag` | `string?` | Optional ad tag (32 hex chars). |
 | `max_tcp_conns` | `usize?` | Optional max concurrent TCP limit. |
 | `expiration_rfc3339` | `string?` | Optional expiration timestamp. |
 | `data_quota_bytes` | `u64?` | Optional data quota. |
+| `rate_limit_up_bps` | `u64?` | Optional upload rate limit in bytes per second. |
+| `rate_limit_down_bps` | `u64?` | Optional download rate limit in bytes per second. |
 | `max_unique_ips` | `usize?` | Optional unique IP limit. |
 | `current_connections` | `u64` | Current live connections. |
 | `active_unique_ips` | `usize` | Current active unique source IPs. |
@@ -1042,12 +1181,25 @@ Note: the request contract is defined, but the corresponding route currently ret
 | `total_octets` | `u64` | Total traffic octets for this user. |
 | `links` | `UserLinks` | Active connection links derived from current config. |
 
+### `UserActiveIps`
+| Field | Type | Description |
+| --- | --- | --- |
+| `username` | `string` | Username with at least one active tracked source IP. |
+| `active_ips` | `ip[]` | Active source IPs for this user. |
+
 #### `UserLinks`
 | Field | Type | Description |
 | --- | --- | --- |
 | `classic` | `string[]` | Active `tg://proxy` links for classic mode. |
 | `secure` | `string[]` | Active `tg://proxy` links for secure/DD mode. |
 | `tls` | `string[]` | Active `tg://proxy` links for EE-TLS mode (for each host+TLS domain). |
+| `tls_domains` | `TlsDomainLink[]` | Extra TLS-domain links as explicit domain/link pairs for `censorship.tls_domains`. |
+
+#### `TlsDomainLink`
+| Field | Type | Description |
+| --- | --- | --- |
+| `domain` | `string` | TLS domain represented by the link. |
+| `link` | `string` | `tg://proxy` link for this domain. |
 
 Link generation uses active config and enabled modes:
 - Link port is `general.links.public_port` when configured; otherwise `server.port`.
@@ -1067,13 +1219,27 @@ Link generation uses active config and enabled modes:
 | `user` | `UserInfo` | Created or updated user view. |
 | `secret` | `string` | Effective user secret. |
 
+### `DeleteUserResponse`
+| Field | Type | Description |
+| --- | --- | --- |
+| `username` | `string` | Deleted username. |
+| `in_runtime` | `bool` | `true` when runtime config still contains the user and hot-reload has not applied deletion yet. |
+
+### `ResetUserQuotaResponse`
+| Field | Type | Description |
+| --- | --- | --- |
+| `username` | `string` | User whose runtime quota counter was reset. |
+| `used_bytes` | `u64` | Current used bytes after reset; always `0` on success. |
+| `last_reset_epoch_secs` | `u64` | Unix timestamp of the reset operation. |
+
 ## Mutation Semantics
 
 | Endpoint | Notes |
 | --- | --- |
 | `POST /v1/users` | Creates user, validates config, then atomically updates only affected `access.*` TOML tables (`access.users` always, plus optional per-user tables present in request). |
-| `PATCH /v1/users/{username}` | Partial update of provided fields only. Missing fields remain unchanged. Current implementation persists full config document on success. |
-| `POST /v1/users/{username}/rotate-secret` | Currently returns `404` in runtime route matcher; request schema is reserved for intended behavior. |
+| `PATCH /v1/users/{username}` | Partial update of provided fields only. Missing fields remain unchanged; explicit `null` removes optional per-user entries. The write path updates only affected `access.*` TOML tables. |
+| `POST /v1/users/{username}/rotate-secret` | Replaces the user's secret with a provided valid 32-hex value or a generated value, then returns the effective secret in `CreateUserResponse`. |
+| `POST /v1/users/{username}/reset-quota` | Resets the runtime quota counter for the route username, persists quota state to `general.quota_state_path`, and does not modify user config. |
 | `DELETE /v1/users/{username}` | Deletes only specified user, removes this user from related optional `access.user_*` maps, blocks last-user deletion, and atomically updates only related `access.*` TOML tables. |
 
 All mutating endpoints:
@@ -1082,6 +1248,12 @@ All mutating endpoints:
 - Return new `revision` after successful write.
 - Use process-local mutation lock + atomic write (`tmp + rename`) for config persistence.
 
+Docker deployment note:
+- Mutating endpoints require `config.toml` to live inside a writable mounted directory.
+- Do not mount `config.toml` as a single bind-mounted file when API mutations are enabled; atomic `tmp + rename` writes can fail with `Device or resource busy`.
+- Mount the config directory instead, for example `./config:/etc/telemt:rw`, and start Telemt with `/etc/telemt/config.toml`.
+- A read-only single-file mount remains valid only for read-only deployments or when `[server.api].read_only=true`.
+
 Delete path cleanup guarantees:
 - Config cleanup removes only the requested username keys.
 - Runtime unique-IP cleanup removes only this user's limiter and tracked IP state.
@@ -1114,12 +1286,12 @@ Additional runtime endpoint behavior:
 ## ME Fallback Behavior Exposed Via API
 
 When `general.use_middle_proxy=true` and `general.me2dc_fallback=true`:
-- Startup does not block on full ME pool readiness; initialization can continue in background.
+- Startup opens Direct-DC routing first, then initializes ME in background and switches new sessions to Middle mode after ME readiness is observed.
 - Runtime initialization payload can expose ME stage `background_init` until pool becomes ready.
 - Admission/routing decision uses two readiness grace windows for "ME not ready" periods:
-  `80s` before first-ever readiness is observed (startup grace),
+  direct startup fallback before first-ever readiness is observed,
   `6s` after readiness has been observed at least once (runtime failover timeout).
-- While in fallback window breach, new sessions are routed via Direct-DC; when ME becomes ready, routing returns to Middle mode for new sessions.
+- While fallback is active, new sessions are routed via Direct-DC; when ME becomes ready, routing returns to Middle mode. Direct sessions affected by the cutover are closed with the existing staggered delay so clients reconnect through the current route.
 
 ## Serialization Rules
 
@@ -1148,5 +1320,4 @@ When `general.use_middle_proxy=true` and `general.me2dc_fallback=true`:
 
 ## Known Limitations (Current Release)
 
-- `POST /v1/users/{username}/rotate-secret` is currently unreachable in route matcher and returns `404`.
 - API runtime controls under `server.api` are documented as restart-required; hot-reload behavior for these fields is not strictly uniform in all change combinations.

docs/Config_params/CONFIG_PARAMS.en.md

@@ -98,7 +98,7 @@ This document lists all configuration keys accepted by `config.toml`.
 | [`middle_proxy_warm_standby`](#middle_proxy_warm_standby) | `usize` | `16` |
 | [`me_init_retry_attempts`](#me_init_retry_attempts) | `u32` | `0` |
 | [`me2dc_fallback`](#me2dc_fallback) | `bool` | `true` |
-| [`me2dc_fast`](#me2dc_fast) | `bool` | `false` |
+| [`me2dc_fast`](#me2dc_fast) | `bool` | `true` |
 | [`me_keepalive_enabled`](#me_keepalive_enabled) | `bool` | `true` |
 | [`me_keepalive_interval_secs`](#me_keepalive_interval_secs) | `u64` | `8` |
 | [`me_keepalive_jitter_secs`](#me_keepalive_jitter_secs) | `u64` | `2` |
@@ -392,7 +392,7 @@ This document lists all configuration keys accepted by `config.toml`.
     ```
 ## me2dc_fallback
   - **Constraints / validation**: `bool`.
-  - **Description**: Allows fallback from ME mode to direct DC when ME startup fails.
+  - **Description**: Allows Direct-DC fallback when ME is unavailable. With `use_middle_proxy = true`, startup opens Direct-DC routing first and moves new sessions to ME after ME readiness is observed.
   - **Example**:
 
     ```toml
@@ -401,14 +401,14 @@ This document lists all configuration keys accepted by `config.toml`.
     ```
 ## me2dc_fast
   - **Constraints / validation**: `bool`. Active only when `use_middle_proxy = true` and `me2dc_fallback = true`.
-  - **Description**: Fast ME->Direct fallback mode for new sessions.
+  - **Description**: Fast ME->Direct fallback mode for new sessions after ME was ready at least once. Initial direct-first startup fallback is controlled by `me2dc_fallback`.
   - **Example**:
 
     ```toml
     [general]
     use_middle_proxy = true
     me2dc_fallback = true
-    me2dc_fast = false
+    me2dc_fast = true
     ```
 ## me_keepalive_enabled
   - **Constraints / validation**: `bool`.
@@ -2352,6 +2352,7 @@ Note: This section also accepts the legacy alias `[server.admin_api]` (same sche
 | [`mask`](#mask) | `bool` | `true` |
 | [`mask_host`](#mask_host) | `String` | — |
 | [`mask_port`](#mask_port) | `u16` | `443` |
+| [`exclusive_mask`](#exclusive_mask) | `Map<String,String>` | `{}` |
 | [`mask_unix_sock`](#mask_unix_sock) | `String` | — |
 | [`fake_cert_len`](#fake_cert_len) | `usize` | `2048` |
 | [`tls_emulation`](#tls_emulation) | `bool` | `true` |
@@ -2459,6 +2460,18 @@ Note: This section also accepts the legacy alias `[server.admin_api]` (same sche
     [censorship]
     mask_port = 443
     ```
+## exclusive_mask
+  - **Constraints / validation**: TOML map. Keys must be SNI domain names. Values must be `host:port` with `port > 0`; IPv6 literals must be bracketed.
+  - **Description**: Per-SNI TCP mask targets for fallback traffic. When a TLS ClientHello SNI matches a key, Telemt relays that unauthenticated connection to the mapped target. Other fallback traffic keeps using the existing `mask_host`/`mask_port` or SNI-aware default masking behavior.
+  - **Example**:
+
+    ```toml
+    [censorship]
+    tls_domains = ["petrovich.ru", "bsi.bund.de", "telekom.com"]
+
+    [censorship.exclusive_mask]
+    "bsi.bund.de" = "127.0.0.1:443"
+    ```
 ## mask_unix_sock
   - **Constraints / validation**: `String` (optional).
     - Must not be empty when set.

docs/Config_params/CONFIG_PARAMS.ru.md

@@ -98,7 +98,7 @@
 | [`middle_proxy_warm_standby`](#middle_proxy_warm_standby) | `usize` | `16` |
 | [`me_init_retry_attempts`](#me_init_retry_attempts) | `u32` | `0` |
 | [`me2dc_fallback`](#me2dc_fallback) | `bool` | `true` |
-| [`me2dc_fast`](#me2dc_fast) | `bool` | `false` |
+| [`me2dc_fast`](#me2dc_fast) | `bool` | `true` |
 | [`me_keepalive_enabled`](#me_keepalive_enabled) | `bool` | `true` |
 | [`me_keepalive_interval_secs`](#me_keepalive_interval_secs) | `u64` | `8` |
 | [`me_keepalive_jitter_secs`](#me_keepalive_jitter_secs) | `u64` | `2` |
@@ -392,7 +392,7 @@
     ```
 ## me2dc_fallback
   - **Ограничения / валидация**: `bool`.
-  - **Описание**: Перейти из режима ME в режим прямого соединения (DC) в случае сбоя запуска ME.
+  - **Описание**: Разрешает fallback на прямой DC, когда ME недоступен. При `use_middle_proxy = true` запуск сначала открывает маршрутизацию через Direct-DC, а новые сеансы переводятся на ME после подтверждения готовности ME.
   - **Пример**:
 
     ```toml
@@ -401,14 +401,14 @@
     ```
 ## me2dc_fast
   - **Ограничения / валидация**: `bool`. Используется только, когда `use_middle_proxy = true` и `me2dc_fallback = true`.
-  - **Описание**: Режим для быстрого перехода между режимами ME->DC для новых сеансов.
+  - **Описание**: Быстрый fallback ME->Direct для новых сеансов после того, как ME уже был готов хотя бы один раз. Начальный direct-first fallback управляется `me2dc_fallback`.
   - **Пример**:
 
     ```toml
     [general]
     use_middle_proxy = true
     me2dc_fallback = true
-    me2dc_fast = false
+    me2dc_fast = true
     ```
 ## me_keepalive_enabled
   - **Ограничения / валидация**: `bool`.
@@ -2358,6 +2358,7 @@
 | [`mask`](#mask) | `bool` | `true` |
 | [`mask_host`](#mask_host) | `String` | — |
 | [`mask_port`](#mask_port) | `u16` | `443` |
+| [`exclusive_mask`](#exclusive_mask) | `Map<String,String>` | `{}` |
 | [`mask_unix_sock`](#mask_unix_sock) | `String` | — |
 | [`fake_cert_len`](#fake_cert_len) | `usize` | `2048` |
 | [`tls_emulation`](#tls_emulation) | `bool` | `true` |
@@ -2464,6 +2465,18 @@
     [censorship]
     mask_port = 443
     ```
+## exclusive_mask
+  - **Ограничения / валидация**: TOML map. Ключи должны быть доменами SNI. Значения должны иметь формат `host:port`, где `port > 0`; IPv6 literals должны быть в квадратных скобках.
+  - **Описание**: Per-SNI TCP targets для fallback-трафика. Если SNI в TLS ClientHello совпадает с ключом, Telemt проксирует это неаутентифицированное соединение на указанный target. Остальной fallback-трафик продолжает использовать существующий `mask_host`/`mask_port` или SNI-aware default masking behavior.
+  - **Пример**:
+
+    ```toml
+    [censorship]
+    tls_domains = ["petrovich.ru", "bsi.bund.de", "telekom.com"]
+
+    [censorship.exclusive_mask]
+    "bsi.bund.de" = "127.0.0.1:443"
+    ```
 ## mask_unix_sock
   - **Ограничения / валидация**: `String` (optional).
     - Значение не должно быть пустым, если задан.

docs/Quick_start/QUICK_START_GUIDE.en.md

@@ -254,6 +254,19 @@ docker compose down
 > - `docker-compose.yml` maps `./config.toml` to `/app/config.toml` (read-only)
 > - By default it publishes `443:443` and runs with dropped capabilities (only `NET_BIND_SERVICE` is added)
 > - If you really need host networking (usually only for some IPv6 setups) uncomment `network_mode: host`
+> - If you enable mutating Control API endpoints, mount a writable config directory instead of a single `config.toml` file. Telemt persists config changes with atomic `tmp + rename` writes, and a single bind-mounted file can fail with `Device or resource busy`.
+
+Example writable config mount for Control API mutations:
+```yaml
+services:
+  telemt:
+    working_dir: /run/telemt
+    volumes:
+      - ./config:/etc/telemt:rw
+    tmpfs:
+      - /run/telemt:rw,mode=1777,size=4m
+    command: /usr/local/bin/telemt /etc/telemt/config.toml
+```
 
 **Run without Compose**
 ```bash

src/api/config_store.rs

@@ -7,7 +7,7 @@ use hyper::header::IF_MATCH;
 use serde::Serialize;
 use sha2::{Digest, Sha256};
 
-use crate::config::ProxyConfig;
+use crate::config::{ProxyConfig, RateLimitBps};
 
 use super::model::ApiFailure;
 
@@ -18,6 +18,7 @@ pub(super) enum AccessSection {
     UserMaxTcpConns,
     UserExpirations,
     UserDataQuota,
+    UserRateLimits,
     UserMaxUniqueIps,
 }
 
@@ -29,6 +30,7 @@ impl AccessSection {
             Self::UserMaxTcpConns => "access.user_max_tcp_conns",
             Self::UserExpirations => "access.user_expirations",
             Self::UserDataQuota => "access.user_data_quota",
+            Self::UserRateLimits => "access.user_rate_limits",
             Self::UserMaxUniqueIps => "access.user_max_unique_ips",
         }
     }
@@ -169,6 +171,15 @@ fn render_access_section(cfg: &ProxyConfig, section: AccessSection) -> Result<St
                 .collect();
             serialize_table_body(&rows)?
         }
+        AccessSection::UserRateLimits => {
+            let rows: BTreeMap<String, RateLimitBps> = cfg
+                .access
+                .user_rate_limits
+                .iter()
+                .map(|(key, value)| (key.clone(), *value))
+                .collect();
+            serialize_rate_limit_body(&rows)?
+        }
         AccessSection::UserMaxUniqueIps => {
             let rows: BTreeMap<String, usize> = cfg
                 .access
@@ -197,6 +208,7 @@ fn access_section_is_empty(cfg: &ProxyConfig, section: AccessSection) -> bool {
         AccessSection::UserMaxTcpConns => cfg.access.user_max_tcp_conns.is_empty(),
         AccessSection::UserExpirations => cfg.access.user_expirations.is_empty(),
         AccessSection::UserDataQuota => cfg.access.user_data_quota.is_empty(),
+        AccessSection::UserRateLimits => cfg.access.user_rate_limits.is_empty(),
         AccessSection::UserMaxUniqueIps => cfg.access.user_max_unique_ips.is_empty(),
     }
 }
@@ -206,6 +218,28 @@ fn serialize_table_body<T: Serialize>(value: &T) -> Result<String, ApiFailure> {
         .map_err(|e| ApiFailure::internal(format!("failed to serialize access section: {}", e)))
 }
 
+fn serialize_rate_limit_body(rows: &BTreeMap<String, RateLimitBps>) -> Result<String, ApiFailure> {
+    let mut out = String::new();
+    for (key, value) in rows {
+        let key = serialize_toml_key(key)?;
+        out.push_str(&format!(
+            "{key} = {{ up_bps = {}, down_bps = {} }}\n",
+            value.up_bps, value.down_bps
+        ));
+    }
+    Ok(out)
+}
+
+fn serialize_toml_key(key: &str) -> Result<String, ApiFailure> {
+    let mut row = BTreeMap::new();
+    row.insert(key.to_string(), 0_u8);
+    let rendered = serialize_table_body(&row)?;
+    rendered
+        .split_once(" = ")
+        .map(|(key, _)| key.to_string())
+        .ok_or_else(|| ApiFailure::internal("failed to serialize TOML key"))
+}
+
 fn upsert_toml_table(source: &str, table_name: &str, replacement: &str) -> String {
     if let Some((start, end)) = find_toml_table_bounds(source, table_name) {
         let mut out = String::with_capacity(source.len() + replacement.len());
@@ -285,3 +319,26 @@ fn write_atomic_sync(path: &Path, contents: &str) -> std::io::Result<()> {
     }
     write_result
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn render_user_rate_limits_section() {
+        let mut cfg = ProxyConfig::default();
+        cfg.access.user_rate_limits.insert(
+            "alice".to_string(),
+            RateLimitBps {
+                up_bps: 1024,
+                down_bps: 2048,
+            },
+        );
+
+        let rendered = render_access_section(&cfg, AccessSection::UserRateLimits)
+            .expect("section must render");
+
+        assert!(rendered.starts_with("[access.user_rate_limits]\n"));
+        assert!(rendered.contains("alice = { up_bps = 1024, down_bps = 2048 }"));
+    }
+}

src/api/mod.rs

@@ -68,7 +68,9 @@ use runtime_zero::{
     build_limits_effective_data, build_runtime_gates_data, build_security_posture_data,
     build_system_info_data,
 };
-use users::{create_user, delete_user, patch_user, rotate_secret, users_from_config};
+use users::{
+    build_user_quota_list, create_user, delete_user, patch_user, rotate_secret, users_from_config,
+};
 
 const API_MAX_CONTROL_CONNECTIONS: usize = 1024;
 const API_HTTP_CONNECTION_TIMEOUT: Duration = Duration::from_secs(15);
@@ -504,6 +506,12 @@ async fn handle(
                 .await;
                 Ok(success_response(StatusCode::OK, users, revision))
             }
+            ("GET", "/v1/users/quota") => {
+                let revision = current_revision(&shared.config_path).await?;
+                let disk_cfg = load_config_from_disk(&shared.config_path).await?;
+                let data = build_user_quota_list(&disk_cfg, shared.stats.as_ref());
+                Ok(success_response(StatusCode::OK, data, revision))
+            }
             ("POST", "/v1/users") => {
                 if api_cfg.read_only {
                     return Ok(error_response(

src/api/model.rs

@@ -473,6 +473,8 @@ pub(super) struct UserInfo {
     pub(super) max_tcp_conns: Option<usize>,
     pub(super) expiration_rfc3339: Option<String>,
     pub(super) data_quota_bytes: Option<u64>,
+    pub(super) rate_limit_up_bps: Option<u64>,
+    pub(super) rate_limit_down_bps: Option<u64>,
     pub(super) max_unique_ips: Option<usize>,
     pub(super) current_connections: u64,
     pub(super) active_unique_ips: usize,
@@ -508,6 +510,19 @@ pub(super) struct ResetUserQuotaResponse {
     pub(super) last_reset_epoch_secs: u64,
 }
 
+#[derive(Serialize)]
+pub(super) struct UserQuotaListData {
+    pub(super) users: Vec<UserQuotaEntry>,
+}
+
+#[derive(Serialize)]
+pub(super) struct UserQuotaEntry {
+    pub(super) username: String,
+    pub(super) data_quota_bytes: u64,
+    pub(super) used_bytes: u64,
+    pub(super) last_reset_epoch_secs: u64,
+}
+
 #[derive(Deserialize)]
 pub(super) struct CreateUserRequest {
     pub(super) username: String,
@@ -516,6 +531,8 @@ pub(super) struct CreateUserRequest {
     pub(super) max_tcp_conns: Option<usize>,
     pub(super) expiration_rfc3339: Option<String>,
     pub(super) data_quota_bytes: Option<u64>,
+    pub(super) rate_limit_up_bps: Option<u64>,
+    pub(super) rate_limit_down_bps: Option<u64>,
     pub(super) max_unique_ips: Option<usize>,
 }
 
@@ -531,6 +548,10 @@ pub(super) struct PatchUserRequest {
     #[serde(default, deserialize_with = "patch_field")]
     pub(super) data_quota_bytes: Patch<u64>,
     #[serde(default, deserialize_with = "patch_field")]
+    pub(super) rate_limit_up_bps: Patch<u64>,
+    #[serde(default, deserialize_with = "patch_field")]
+    pub(super) rate_limit_down_bps: Patch<u64>,
+    #[serde(default, deserialize_with = "patch_field")]
     pub(super) max_unique_ips: Patch<usize>,
 }
 

src/api/patch.rs

@@ -114,7 +114,9 @@ mod tests {
             "secret": "00112233445566778899aabbccddeeff",
             "max_tcp_conns": 0,
             "max_unique_ips": null,
-            "data_quota_bytes": 1024
+            "data_quota_bytes": 1024,
+            "rate_limit_up_bps": 4096,
+            "rate_limit_down_bps": null
         }"#;
         let req: PatchUserRequest = serde_json::from_str(raw).expect("valid json");
         assert_eq!(
@@ -124,6 +126,8 @@ mod tests {
         assert!(matches!(req.max_tcp_conns, Patch::Set(0)));
         assert!(matches!(req.max_unique_ips, Patch::Remove));
         assert!(matches!(req.data_quota_bytes, Patch::Set(1024)));
+        assert!(matches!(req.rate_limit_up_bps, Patch::Set(4096)));
+        assert!(matches!(req.rate_limit_down_bps, Patch::Remove));
         assert!(matches!(req.expiration_rfc3339, Patch::Unchanged));
         assert!(matches!(req.user_ad_tag, Patch::Unchanged));
     }

src/api/runtime_zero.rs

@@ -178,6 +178,7 @@ pub(super) async fn build_runtime_gates_data(
     cfg: &ProxyConfig,
 ) -> RuntimeGatesData {
     let startup_summary = build_runtime_startup_summary(shared).await;
+    let startup_snapshot = shared.startup_tracker.snapshot().await;
     let route_state = shared.route_runtime.snapshot();
     let route_mode = route_state.mode.as_str();
     let fast_fallback_enabled =
@@ -191,7 +192,9 @@ pub(super) async fn build_runtime_gates_data(
         None
     };
     let reroute_reason = if reroute_active {
-        if fast_fallback_enabled {
+        if startup_snapshot.me.status.as_str() != "ready" {
+            Some("startup_direct_fallback")
+        } else if fast_fallback_enabled {
             Some("fast_not_ready_fallback")
         } else {
             Some("strict_grace_fallback")

src/api/users.rs

@@ -3,6 +3,7 @@ use std::net::IpAddr;
 use hyper::StatusCode;
 
 use crate::config::ProxyConfig;
+use crate::config::RateLimitBps;
 use crate::ip_tracker::UserIpTracker;
 use crate::stats::Stats;
 
@@ -13,8 +14,9 @@ use super::config_store::{
 };
 use super::model::{
     ApiFailure, CreateUserRequest, CreateUserResponse, PatchUserRequest, RotateSecretRequest,
-    TlsDomainLink, UserInfo, UserLinks, is_valid_ad_tag, is_valid_user_secret, is_valid_username,
-    parse_optional_expiration, parse_patch_expiration, random_user_secret,
+    TlsDomainLink, UserInfo, UserLinks, UserQuotaEntry, UserQuotaListData, is_valid_ad_tag,
+    is_valid_user_secret, is_valid_username, parse_optional_expiration, parse_patch_expiration,
+    random_user_secret,
 };
 use super::patch::Patch;
 
@@ -27,6 +29,8 @@ pub(super) async fn create_user(
     let touches_user_max_tcp_conns = body.max_tcp_conns.is_some();
     let touches_user_expirations = body.expiration_rfc3339.is_some();
     let touches_user_data_quota = body.data_quota_bytes.is_some();
+    let touches_user_rate_limits =
+        body.rate_limit_up_bps.is_some() || body.rate_limit_down_bps.is_some();
     let touches_user_max_unique_ips = body.max_unique_ips.is_some();
 
     if !is_valid_username(&body.username) {
@@ -91,6 +95,15 @@ pub(super) async fn create_user(
             .user_data_quota
             .insert(body.username.clone(), quota);
     }
+    if touches_user_rate_limits {
+        cfg.access.user_rate_limits.insert(
+            body.username.clone(),
+            RateLimitBps {
+                up_bps: body.rate_limit_up_bps.unwrap_or(0),
+                down_bps: body.rate_limit_down_bps.unwrap_or(0),
+            },
+        );
+    }
 
     let updated_limit = body.max_unique_ips;
     if let Some(limit) = updated_limit {
@@ -115,6 +128,9 @@ pub(super) async fn create_user(
     if touches_user_data_quota {
         touched_sections.push(AccessSection::UserDataQuota);
     }
+    if touches_user_rate_limits {
+        touched_sections.push(AccessSection::UserRateLimits);
+    }
     if touches_user_max_unique_ips {
         touched_sections.push(AccessSection::UserMaxUniqueIps);
     }
@@ -157,6 +173,8 @@ pub(super) async fn create_user(
                     .then_some(cfg.access.user_max_tcp_conns_global_each)),
             expiration_rfc3339: None,
             data_quota_bytes: None,
+            rate_limit_up_bps: body.rate_limit_up_bps.filter(|limit| *limit > 0),
+            rate_limit_down_bps: body.rate_limit_down_bps.filter(|limit| *limit > 0),
             max_unique_ips: updated_limit,
             current_connections: 0,
             active_unique_ips: 0,
@@ -181,6 +199,8 @@ pub(super) async fn patch_user(
     let touches_user_max_tcp_conns = !matches!(&body.max_tcp_conns, Patch::Unchanged);
     let touches_user_expirations = !matches!(&body.expiration_rfc3339, Patch::Unchanged);
     let touches_user_data_quota = !matches!(&body.data_quota_bytes, Patch::Unchanged);
+    let touches_user_rate_limits = !matches!(&body.rate_limit_up_bps, Patch::Unchanged)
+        || !matches!(&body.rate_limit_down_bps, Patch::Unchanged);
     let touches_user_max_unique_ips = !matches!(&body.max_unique_ips, Patch::Unchanged);
 
     if let Some(secret) = body.secret.as_ref()
@@ -253,6 +273,31 @@ pub(super) async fn patch_user(
             cfg.access.user_data_quota.insert(user.to_string(), quota);
         }
     }
+    if touches_user_rate_limits {
+        let mut rate_limit = cfg
+            .access
+            .user_rate_limits
+            .get(user)
+            .copied()
+            .unwrap_or_default();
+        match body.rate_limit_up_bps {
+            Patch::Unchanged => {}
+            Patch::Remove => rate_limit.up_bps = 0,
+            Patch::Set(limit) => rate_limit.up_bps = limit,
+        }
+        match body.rate_limit_down_bps {
+            Patch::Unchanged => {}
+            Patch::Remove => rate_limit.down_bps = 0,
+            Patch::Set(limit) => rate_limit.down_bps = limit,
+        }
+        if rate_limit.up_bps == 0 && rate_limit.down_bps == 0 {
+            cfg.access.user_rate_limits.remove(user);
+        } else {
+            cfg.access
+                .user_rate_limits
+                .insert(user.to_string(), rate_limit);
+        }
+    }
     // Capture how the per-user IP limit changed, so the in-memory ip_tracker
     // can be synced (set or removed) after the config is persisted.
     let max_unique_ips_change = match body.max_unique_ips {
@@ -288,6 +333,9 @@ pub(super) async fn patch_user(
     if touches_user_data_quota {
         touched_sections.push(AccessSection::UserDataQuota);
     }
+    if touches_user_rate_limits {
+        touched_sections.push(AccessSection::UserRateLimits);
+    }
     if touches_user_max_unique_ips {
         touched_sections.push(AccessSection::UserMaxUniqueIps);
     }
@@ -355,6 +403,7 @@ pub(super) async fn rotate_secret(
         AccessSection::UserMaxTcpConns,
         AccessSection::UserExpirations,
         AccessSection::UserDataQuota,
+        AccessSection::UserRateLimits,
         AccessSection::UserMaxUniqueIps,
     ];
     let revision =
@@ -414,6 +463,7 @@ pub(super) async fn delete_user(
     cfg.access.user_max_tcp_conns.remove(user);
     cfg.access.user_expirations.remove(user);
     cfg.access.user_data_quota.remove(user);
+    cfg.access.user_rate_limits.remove(user);
     cfg.access.user_max_unique_ips.remove(user);
 
     cfg.validate()
@@ -424,6 +474,7 @@ pub(super) async fn delete_user(
         AccessSection::UserMaxTcpConns,
         AccessSection::UserExpirations,
         AccessSection::UserDataQuota,
+        AccessSection::UserRateLimits,
         AccessSection::UserMaxUniqueIps,
     ];
     let revision =
@@ -485,6 +536,18 @@ pub(super) async fn users_from_config(
                 .get(&username)
                 .map(chrono::DateTime::<chrono::Utc>::to_rfc3339),
             data_quota_bytes: cfg.access.user_data_quota.get(&username).copied(),
+            rate_limit_up_bps: cfg
+                .access
+                .user_rate_limits
+                .get(&username)
+                .map(|limit| limit.up_bps)
+                .filter(|limit| *limit > 0),
+            rate_limit_down_bps: cfg
+                .access
+                .user_rate_limits
+                .get(&username)
+                .map(|limit| limit.down_bps)
+                .filter(|limit| *limit > 0),
             max_unique_ips: cfg
                 .access
                 .user_max_unique_ips
@@ -506,6 +569,33 @@ pub(super) async fn users_from_config(
     users
 }
 
+pub(super) fn build_user_quota_list(cfg: &ProxyConfig, stats: &Stats) -> UserQuotaListData {
+    let mut names = cfg.access.users.keys().cloned().collect::<Vec<_>>();
+    names.sort();
+
+    let snapshot = stats.user_quota_snapshot();
+    let mut users = Vec::with_capacity(names.len());
+    for username in names {
+        let Some(&data_quota_bytes) = cfg.access.user_data_quota.get(&username) else {
+            continue;
+        };
+        if data_quota_bytes == 0 {
+            continue;
+        }
+        let (used_bytes, last_reset_epoch_secs) = snapshot
+            .get(&username)
+            .map(|entry| (entry.used_bytes, entry.last_reset_epoch_secs))
+            .unwrap_or((0, 0));
+        users.push(UserQuotaEntry {
+            username,
+            data_quota_bytes,
+            used_bytes,
+            last_reset_epoch_secs,
+        });
+    }
+    UserQuotaListData { users }
+}
+
 fn empty_user_links() -> UserLinks {
     UserLinks {
         classic: Vec::new(),
@@ -758,6 +848,34 @@ mod tests {
         assert_eq!(alice.max_tcp_conns, None);
     }
 
+    #[tokio::test]
+    async fn users_from_config_reports_user_rate_limits() {
+        let mut cfg = ProxyConfig::default();
+        cfg.access.users.insert(
+            "alice".to_string(),
+            "0123456789abcdef0123456789abcdef".to_string(),
+        );
+        cfg.access.user_rate_limits.insert(
+            "alice".to_string(),
+            RateLimitBps {
+                up_bps: 1024,
+                down_bps: 0,
+            },
+        );
+
+        let stats = Stats::new();
+        let tracker = UserIpTracker::new();
+
+        let users = users_from_config(&cfg, &stats, &tracker, None, None, None).await;
+        let alice = users
+            .iter()
+            .find(|entry| entry.username == "alice")
+            .expect("alice must be present");
+
+        assert_eq!(alice.rate_limit_up_bps, Some(1024));
+        assert_eq!(alice.rate_limit_down_bps, None);
+    }
+
     #[tokio::test]
     async fn users_from_config_marks_runtime_membership_when_snapshot_is_provided() {
         let mut disk_cfg = ProxyConfig::default();
@@ -869,4 +987,68 @@ mod tests {
                 .any(|entry| entry.domain == "front-a.example.com")
         );
     }
+
+    #[test]
+    fn build_user_quota_list_skips_users_without_positive_quota_and_sorts_by_username() {
+        let mut cfg = ProxyConfig::default();
+        cfg.access.users.insert(
+            "alice".to_string(),
+            "0123456789abcdef0123456789abcdef".to_string(),
+        );
+        cfg.access.users.insert(
+            "bob".to_string(),
+            "fedcba9876543210fedcba9876543210".to_string(),
+        );
+        cfg.access.users.insert(
+            "carol".to_string(),
+            "aaaabbbbccccddddeeeeffff00001111".to_string(),
+        );
+        // alice has a positive quota and should be listed.
+        cfg.access
+            .user_data_quota
+            .insert("alice".to_string(), 1 << 20);
+        // bob has no quota entry at all (None) — should be skipped.
+        // carol has an explicit zero quota — should be skipped.
+        cfg.access.user_data_quota.insert("carol".to_string(), 0);
+
+        let stats = Stats::new();
+        // Charge some traffic against alice; carol gets traffic too but should
+        // still be filtered out by the quota check.
+        let alice_stats = stats.get_or_create_user_stats_handle("alice");
+        stats.quota_charge_post_write(&alice_stats, 4096);
+        let carol_stats = stats.get_or_create_user_stats_handle("carol");
+        stats.quota_charge_post_write(&carol_stats, 99);
+
+        let data = build_user_quota_list(&cfg, &stats);
+
+        assert_eq!(data.users.len(), 1);
+        let entry = &data.users[0];
+        assert_eq!(entry.username, "alice");
+        assert_eq!(entry.data_quota_bytes, 1 << 20);
+        assert_eq!(entry.used_bytes, 4096);
+        assert_eq!(entry.last_reset_epoch_secs, 0);
+    }
+
+    #[test]
+    fn build_user_quota_list_orders_multiple_users_by_username_ascending() {
+        let mut cfg = ProxyConfig::default();
+        for name in ["charlie", "alice", "bob"] {
+            cfg.access.users.insert(
+                name.to_string(),
+                "0123456789abcdef0123456789abcdef".to_string(),
+            );
+            cfg.access.user_data_quota.insert(name.to_string(), 1 << 30);
+        }
+
+        let stats = Stats::new();
+        let data = build_user_quota_list(&cfg, &stats);
+
+        let names: Vec<&str> = data.users.iter().map(|e| e.username.as_str()).collect();
+        assert_eq!(names, vec!["alice", "bob", "charlie"]);
+        for entry in &data.users {
+            assert_eq!(entry.used_bytes, 0);
+            assert_eq!(entry.last_reset_epoch_secs, 0);
+            assert_eq!(entry.data_quota_bytes, 1 << 30);
+        }
+    }
 }

src/config/hot_reload.rs

@@ -617,6 +617,7 @@ fn warn_non_hot_changes(old: &ProxyConfig, new: &ProxyConfig, non_hot_changed: b
         || old.censorship.mask != new.censorship.mask
         || old.censorship.mask_host != new.censorship.mask_host
         || old.censorship.mask_port != new.censorship.mask_port
+        || old.censorship.exclusive_mask != new.censorship.exclusive_mask
         || old.censorship.mask_unix_sock != new.censorship.mask_unix_sock
         || old.censorship.fake_cert_len != new.censorship.fake_cert_len
         || old.censorship.tls_emulation != new.censorship.tls_emulation

src/config/load.rs

@@ -31,6 +31,87 @@ fn is_valid_tls_domain_name(domain: &str) -> bool {
             .any(|ch| ch.is_whitespace() || matches!(ch, '/' | '\\'))
 }
 
+fn normalize_domain_to_ascii(domain: &str, field: &str) -> Result<String> {
+    let domain = domain.trim();
+    if !is_valid_tls_domain_name(domain) {
+        return Err(ProxyError::Config(format!(
+            "Invalid {field}: '{}'. Must be a valid domain name",
+            domain
+        )));
+    }
+
+    let parsed = url::Url::parse(&format!("https://{domain}/")).map_err(|error| {
+        ProxyError::Config(format!(
+            "Invalid {field}: '{}'. IDNA conversion failed: {error}",
+            domain
+        ))
+    })?;
+    let host = parsed.host_str().ok_or_else(|| {
+        ProxyError::Config(format!("Invalid {field}: '{}'. Host is empty", domain))
+    })?;
+    Ok(host.to_ascii_lowercase())
+}
+
+fn normalize_mask_host_to_ascii(host: &str, field: &str) -> Result<String> {
+    let host = host.trim();
+    if host.starts_with('[') && host.ends_with(']') {
+        let inner = &host[1..host.len() - 1];
+        let ip = inner.parse::<std::net::IpAddr>().map_err(|_| {
+            ProxyError::Config(format!(
+                "Invalid {field}: '{}'. IPv6 literal is invalid",
+                host
+            ))
+        })?;
+        return match ip {
+            std::net::IpAddr::V6(v6) => Ok(format!("[{v6}]")),
+            std::net::IpAddr::V4(v4) => Ok(v4.to_string()),
+        };
+    }
+    if let Ok(ip) = host.parse::<std::net::IpAddr>() {
+        return match ip {
+            std::net::IpAddr::V4(v4) => Ok(v4.to_string()),
+            std::net::IpAddr::V6(v6) => Ok(format!("[{v6}]")),
+        };
+    }
+
+    normalize_domain_to_ascii(host, field)
+}
+
+fn parse_exclusive_mask_target(target: &str) -> Option<(&str, u16)> {
+    let target = target.trim();
+    if target.is_empty() {
+        return None;
+    }
+
+    if target.starts_with('[') {
+        let end = target.find(']')?;
+        if target.get(end + 1..end + 2)? != ":" {
+            return None;
+        }
+        let host = &target[..=end];
+        let port = target[end + 2..].parse::<u16>().ok()?;
+        return (port > 0).then_some((host, port));
+    }
+
+    let (host, port) = target.rsplit_once(':')?;
+    if host.is_empty() || host.contains(':') {
+        return None;
+    }
+    let port = port.parse::<u16>().ok()?;
+    (port > 0).then_some((host, port))
+}
+
+fn normalize_exclusive_mask_target(target: &str, field: &str) -> Result<String> {
+    let (host, port) = parse_exclusive_mask_target(target).ok_or_else(|| {
+        ProxyError::Config(format!(
+            "Invalid {field}: '{}'. Expected host:port with port > 0",
+            target
+        ))
+    })?;
+    let host = normalize_mask_host_to_ascii(host, field)?;
+    Ok(format!("{host}:{port}"))
+}
+
 const TOP_LEVEL_CONFIG_KEYS: &[&str] = &[
     "general",
     "network",
@@ -291,6 +372,7 @@ const CENSORSHIP_CONFIG_KEYS: &[&str] = &[
     "mask",
     "mask_host",
     "mask_port",
+    "exclusive_mask",
     "mask_unix_sock",
     "fake_cert_len",
     "tls_emulation",
@@ -1887,10 +1969,8 @@ impl ProxyConfig {
             }
         }
 
-        // Validate tls_domain.
-        if config.censorship.tls_domain.is_empty() {
-            return Err(ProxyError::Config("tls_domain cannot be empty".to_string()));
-        }
+        config.censorship.tls_domain =
+            normalize_domain_to_ascii(&config.censorship.tls_domain, "censorship.tls_domain")?;
 
         // Validate mask_unix_sock.
         if let Some(ref sock_path) = config.censorship.mask_unix_sock {
@@ -1918,11 +1998,30 @@ impl ProxyConfig {
             }
         }
 
+        if let Some(mask_host) = config.censorship.mask_host.as_mut() {
+            *mask_host = normalize_mask_host_to_ascii(mask_host, "censorship.mask_host")?;
+        }
+
         // Default mask_host to tls_domain if not set and no unix socket configured.
         if config.censorship.mask_host.is_none() && config.censorship.mask_unix_sock.is_none() {
             config.censorship.mask_host = Some(config.censorship.tls_domain.clone());
         }
 
+        for (domain, target) in &config.censorship.exclusive_mask {
+            if !is_valid_tls_domain_name(domain) {
+                return Err(ProxyError::Config(format!(
+                    "Invalid censorship.exclusive_mask domain: '{}'. Must be a valid domain name",
+                    domain
+                )));
+            }
+            if parse_exclusive_mask_target(target).is_none() {
+                return Err(ProxyError::Config(format!(
+                    "Invalid censorship.exclusive_mask target for '{}': '{}'. Expected host:port with port > 0",
+                    domain, target
+                )));
+            }
+        }
+
         // Normalize optional TLS fetch scope: whitespace-only values disable scoped routing.
         config.censorship.tls_fetch_scope = config.censorship.tls_fetch_scope.trim().to_string();
 
@@ -1953,8 +2052,11 @@ impl ProxyConfig {
             let mut all = Vec::with_capacity(1 + config.censorship.tls_domains.len());
             all.push(config.censorship.tls_domain.clone());
             for d in std::mem::take(&mut config.censorship.tls_domains) {
-                if !d.is_empty() && !all.contains(&d) {
-                    all.push(d);
+                if !d.is_empty() {
+                    let domain = normalize_domain_to_ascii(&d, "censorship.tls_domains entry")?;
+                    if !all.contains(&domain) {
+                        all.push(domain);
+                    }
                 }
             }
             // keep primary as tls_domain; store remaining back to tls_domains
@@ -1963,6 +2065,31 @@ impl ProxyConfig {
             }
         }
 
+        let mut exclusive_mask = HashMap::with_capacity(config.censorship.exclusive_mask.len());
+        let mut exclusive_mask_targets =
+            HashMap::with_capacity(config.censorship.exclusive_mask.len());
+        for (domain, target) in std::mem::take(&mut config.censorship.exclusive_mask) {
+            let domain = normalize_domain_to_ascii(&domain, "censorship.exclusive_mask domain")?;
+            let target =
+                normalize_exclusive_mask_target(&target, "censorship.exclusive_mask target")?;
+            let Some((host, port)) = parse_exclusive_mask_target(&target) else {
+                return Err(ProxyError::Config(format!(
+                    "Invalid censorship.exclusive_mask target for '{}': '{}'. Expected host:port with port > 0",
+                    domain, target
+                )));
+            };
+            exclusive_mask_targets.insert(
+                domain.clone(),
+                ExclusiveMaskTarget {
+                    host: host.to_string(),
+                    port,
+                },
+            );
+            exclusive_mask.insert(domain, target);
+        }
+        config.censorship.exclusive_mask = exclusive_mask;
+        config.censorship.exclusive_mask_targets = exclusive_mask_targets;
+
         // Migration: prefer_ipv6 -> network.prefer.
         if config.general.prefer_ipv6 {
             if config.network.prefer == 4 {
@@ -2126,6 +2253,21 @@ impl ProxyConfig {
             }
         }
 
+        for (domain, target) in &self.censorship.exclusive_mask {
+            if !is_valid_tls_domain_name(domain) {
+                return Err(ProxyError::Config(format!(
+                    "Invalid censorship.exclusive_mask domain: '{}'. Must be a valid domain name",
+                    domain
+                )));
+            }
+            if parse_exclusive_mask_target(target).is_none() {
+                return Err(ProxyError::Config(format!(
+                    "Invalid censorship.exclusive_mask target for '{}': '{}'. Expected host:port with port > 0",
+                    domain, target
+                )));
+            }
+        }
+
         for (user, tag) in &self.access.user_ad_tags {
             let zeros = "00000000000000000000000000000000";
             if !is_valid_ad_tag(tag) {
@@ -2667,6 +2809,44 @@ mod tests {
         );
     }
 
+    #[test]
+    fn exclusive_mask_parses_domain_target_map() {
+        let cfg = load_config_from_temp_toml(
+            r#"
+            [general]
+            [network]
+            [server]
+            [access]
+            [censorship]
+            tls_domain = "weißbiergärten.de"
+            tls_domains = ["bürgeramt.de"]
+            [censorship.exclusive_mask]
+            "bürgeramt.de" = "rindfleischetikettierungsüberwachungsaufgabenübertragungsgesetz.de:443"
+            "ipv6.example" = "[::1]:443"
+            "#,
+        );
+
+        assert!(cfg.censorship.tls_domain.is_ascii());
+        assert!(cfg.censorship.tls_domain.contains("xn--"));
+        assert_eq!(cfg.censorship.tls_domains.len(), 1);
+        let normalized_extra = &cfg.censorship.tls_domains[0];
+        assert!(normalized_extra.is_ascii());
+        assert!(normalized_extra.contains("xn--"));
+
+        let normalized_target = cfg
+            .censorship
+            .exclusive_mask
+            .get(normalized_extra)
+            .expect("exclusive_mask key must match normalized tls_domains entry");
+        assert!(normalized_target.is_ascii());
+        assert!(normalized_target.contains("xn--"));
+        assert!(normalized_target.ends_with(":443"));
+        assert_eq!(
+            cfg.censorship.exclusive_mask.get("ipv6.example"),
+            Some(&"[::1]:443".to_string())
+        );
+    }
+
     #[test]
     fn api_gray_action_parses_and_defaults_to_drop() {
         let cfg_default: ProxyConfig = toml::from_str(

src/config/types.rs

@@ -21,8 +21,7 @@ pub enum LogLevel {
     #[default]
     Normal,
     /// Minimal output: only warnings and errors (warn + error).
-    /// Startup messages (config, DC connectivity, proxy links) are always shown
-    /// via info! before the filter is applied.
+    /// Proxy links may still be emitted through their dedicated target.
     Silent,
 }
 
@@ -1688,6 +1687,14 @@ impl Default for TlsFetchConfig {
     }
 }
 
+#[derive(Debug, Clone)]
+pub struct ExclusiveMaskTarget {
+    /// Target host after IDNA/IP normalization.
+    pub host: String,
+    /// TCP port for the selected target.
+    pub port: u16,
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct AntiCensorshipConfig {
     #[serde(default = "default_tls_domain")]
@@ -1719,6 +1726,14 @@ pub struct AntiCensorshipConfig {
     #[serde(default = "default_mask_port")]
     pub mask_port: u16,
 
+    /// Per-SNI TCP mask targets. Keys are SNI domains, values are `host:port`.
+    #[serde(default)]
+    pub exclusive_mask: HashMap<String, String>,
+
+    /// Parsed runtime cache for per-SNI TCP mask targets.
+    #[serde(skip)]
+    pub exclusive_mask_targets: HashMap<String, ExclusiveMaskTarget>,
+
     #[serde(default)]
     pub mask_unix_sock: Option<String>,
 
@@ -1842,6 +1857,8 @@ impl Default for AntiCensorshipConfig {
             mask: default_true(),
             mask_host: None,
             mask_port: default_mask_port(),
+            exclusive_mask: HashMap::new(),
+            exclusive_mask_targets: HashMap::new(),
             mask_unix_sock: None,
             fake_cert_len: default_fake_cert_len(),
             tls_emulation: true,

src/ip_tracker/admission.rs

@@ -0,0 +1,173 @@
+use super::*;
+
+impl UserIpTracker {
+    pub async fn set_limit_policy(&self, mode: UserMaxUniqueIpsMode, window_secs: u64) {
+        self.limit_mode
+            .store(Self::mode_to_u8(mode), Ordering::Relaxed);
+        self.limit_window_secs
+            .store(window_secs.max(1), Ordering::Relaxed);
+    }
+
+    pub async fn set_user_limit(&self, username: &str, max_ips: usize) {
+        self.max_ips.insert(username.to_string(), max_ips);
+    }
+
+    pub async fn remove_user_limit(&self, username: &str) {
+        self.max_ips.remove(username);
+    }
+
+    pub async fn load_limits(&self, default_limit: usize, limits: &HashMap<String, usize>) {
+        self.default_max_ips.store(default_limit, Ordering::Relaxed);
+        self.max_ips.clear();
+        for (username, limit) in limits {
+            self.max_ips.insert(username.clone(), *limit);
+        }
+    }
+
+    pub(super) fn prune_recent(
+        user_recent: &mut HashMap<IpAddr, Instant>,
+        now: Instant,
+        window: Duration,
+    ) -> usize {
+        if user_recent.is_empty() {
+            return 0;
+        }
+        let before = user_recent.len();
+        user_recent.retain(|_, seen_at| now.duration_since(*seen_at) <= window);
+        before.saturating_sub(user_recent.len())
+    }
+
+    pub async fn check_and_add(&self, username: &str, ip: IpAddr) -> Result<(), String> {
+        self.drain_cleanup_for_user(username).await;
+        self.maybe_compact_empty_users().await;
+        let limit = self.user_limit(username);
+        let mode = Self::mode_from_u8(self.limit_mode.load(Ordering::Relaxed));
+        let window = self.limit_window();
+        let now = Instant::now();
+
+        let shard_idx = Self::shard_idx(username);
+        let mut shard = self.shards[shard_idx].write().await;
+        let user_active = shard.active_ips.entry(username.to_string()).or_default();
+        let active_contains_ip = user_active.contains_key(&ip);
+        let active_len = user_active.len();
+        let user_recent = shard.recent_ips.entry(username.to_string()).or_default();
+        let pruned_recent_entries = Self::prune_recent(user_recent, now, window);
+        Self::decrement_counter(&self.recent_entry_count, pruned_recent_entries);
+        let recent_contains_ip = user_recent.contains_key(&ip);
+        let recent_len = user_recent.len();
+
+        if active_contains_ip {
+            if !recent_contains_ip
+                && !Self::try_increment_counter(&self.recent_entry_count, MAX_RECENT_IP_ENTRIES)
+            {
+                self.recent_cap_rejects.fetch_add(1, Ordering::Relaxed);
+                return Err(format!(
+                    "IP tracker recent entry cap reached: entries={}/{}",
+                    self.recent_entry_count.load(Ordering::Relaxed),
+                    MAX_RECENT_IP_ENTRIES
+                ));
+            }
+            let Some(count) = shard
+                .active_ips
+                .get_mut(username)
+                .and_then(|user_active| user_active.get_mut(&ip))
+            else {
+                return Err(format!(
+                    "IP tracker active entry unavailable for user '{username}'"
+                ));
+            };
+            *count = count.saturating_add(1);
+            if let Some(user_recent) = shard.recent_ips.get_mut(username) {
+                user_recent.insert(ip, now);
+            }
+            return Ok(());
+        }
+
+        let is_new_ip = !recent_contains_ip;
+
+        if let Some(limit) = limit {
+            let active_limit_reached = active_len >= limit;
+            let recent_limit_reached = recent_len >= limit && is_new_ip;
+            let deny = match mode {
+                UserMaxUniqueIpsMode::ActiveWindow => active_limit_reached,
+                UserMaxUniqueIpsMode::TimeWindow => recent_limit_reached,
+                UserMaxUniqueIpsMode::Combined => active_limit_reached || recent_limit_reached,
+            };
+
+            if deny {
+                return Err(format!(
+                    "IP limit reached for user '{}': active={}/{} recent={}/{} mode={:?}",
+                    username, active_len, limit, recent_len, limit, mode
+                ));
+            }
+        }
+
+        if !Self::try_increment_counter(&self.active_entry_count, MAX_ACTIVE_IP_ENTRIES) {
+            self.active_cap_rejects.fetch_add(1, Ordering::Relaxed);
+            return Err(format!(
+                "IP tracker active entry cap reached: entries={}/{}",
+                self.active_entry_count.load(Ordering::Relaxed),
+                MAX_ACTIVE_IP_ENTRIES
+            ));
+        }
+        let mut reserved_recent = false;
+        if is_new_ip {
+            if !Self::try_increment_counter(&self.recent_entry_count, MAX_RECENT_IP_ENTRIES) {
+                Self::decrement_counter(&self.active_entry_count, 1);
+                self.recent_cap_rejects.fetch_add(1, Ordering::Relaxed);
+                return Err(format!(
+                    "IP tracker recent entry cap reached: entries={}/{}",
+                    self.recent_entry_count.load(Ordering::Relaxed),
+                    MAX_RECENT_IP_ENTRIES
+                ));
+            }
+            reserved_recent = true;
+        }
+
+        let Some(user_active) = shard.active_ips.get_mut(username) else {
+            Self::decrement_counter(&self.active_entry_count, 1);
+            if reserved_recent {
+                Self::decrement_counter(&self.recent_entry_count, 1);
+            }
+            return Err(format!(
+                "IP tracker active entry unavailable for user '{username}'"
+            ));
+        };
+        if user_active.insert(ip, 1).is_some() {
+            Self::decrement_counter(&self.active_entry_count, 1);
+        }
+        let Some(user_recent) = shard.recent_ips.get_mut(username) else {
+            Self::decrement_counter(&self.active_entry_count, 1);
+            if reserved_recent {
+                Self::decrement_counter(&self.recent_entry_count, 1);
+            }
+            return Err(format!(
+                "IP tracker recent entry unavailable for user '{username}'"
+            ));
+        };
+        if user_recent.insert(ip, now).is_some() && reserved_recent {
+            Self::decrement_counter(&self.recent_entry_count, 1);
+        }
+        Ok(())
+    }
+
+    pub async fn remove_ip(&self, username: &str, ip: IpAddr) {
+        self.maybe_compact_empty_users().await;
+        let shard_idx = Self::shard_idx(username);
+        let mut shard = self.shards[shard_idx].write().await;
+        let mut removed_active_entries = 0usize;
+        if let Some(user_ips) = shard.active_ips.get_mut(username) {
+            if let Some(count) = user_ips.get_mut(&ip) {
+                if *count > 1 {
+                    *count -= 1;
+                } else if user_ips.remove(&ip).is_some() {
+                    removed_active_entries = 1;
+                }
+            }
+            if user_ips.is_empty() {
+                shard.active_ips.remove(username);
+            }
+        }
+        Self::decrement_counter(&self.active_entry_count, removed_active_entries);
+    }
+}

src/ip_tracker/cleanup.rs

@@ -0,0 +1,148 @@
+use super::*;
+
+impl UserIpTracker {
+    /// Queues a deferred active IP cleanup for a later async drain.
+    pub fn enqueue_cleanup(&self, user: String, ip: IpAddr) {
+        self.observe_cleanup_poison_for_tests();
+        let shard_idx = Self::shard_idx(&user);
+        let cleanup_shard = &self.cleanup_shards[shard_idx];
+        match cleanup_shard.queue.lock() {
+            Ok(mut queue) => {
+                let user_queue = queue.entry(user).or_default();
+                let count = user_queue.entry(ip).or_insert(0);
+                if *count == 0 {
+                    self.cleanup_queue_len.fetch_add(1, Ordering::Relaxed);
+                }
+                *count = count.saturating_add(1);
+                self.cleanup_deferred_releases
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            Err(poisoned) => {
+                let mut queue = poisoned.into_inner();
+                let user_queue = queue.entry(user.clone()).or_default();
+                let count = user_queue.entry(ip).or_insert(0);
+                if *count == 0 {
+                    self.cleanup_queue_len.fetch_add(1, Ordering::Relaxed);
+                }
+                *count = count.saturating_add(1);
+                self.cleanup_deferred_releases
+                    .fetch_add(1, Ordering::Relaxed);
+                cleanup_shard.queue.clear_poison();
+                tracing::warn!(
+                    "UserIpTracker cleanup_queue lock poisoned; recovered and enqueued IP cleanup for {} ({})",
+                    user,
+                    ip
+                );
+            }
+        }
+    }
+
+    #[cfg(test)]
+    pub(crate) fn cleanup_queue_len_for_tests(&self) -> usize {
+        self.cleanup_queue_len.load(Ordering::Relaxed) as usize
+    }
+
+    #[cfg(test)]
+    pub(crate) fn cleanup_queue_mutex_for_tests(
+        &self,
+    ) -> Arc<Mutex<HashMap<(String, IpAddr), usize>>> {
+        Arc::clone(&self.cleanup_queue_poison_probe)
+    }
+
+    pub(crate) async fn drain_cleanup_queue(&self) {
+        if self.cleanup_queue_len.load(Ordering::Relaxed) == 0 {
+            return;
+        }
+        for shard_idx in 0..USER_IP_TRACKER_SHARDS {
+            self.drain_cleanup_shard(shard_idx).await;
+        }
+    }
+
+    pub(super) async fn drain_cleanup_for_user(&self, user: &str) {
+        if self.cleanup_queue_len.load(Ordering::Relaxed) == 0 {
+            return;
+        }
+        let shard_idx = Self::shard_idx(user);
+        let cleanup_shard = &self.cleanup_shards[shard_idx];
+        let to_remove = match cleanup_shard.queue.lock() {
+            Ok(mut queue) => queue.remove(user).unwrap_or_default(),
+            Err(poisoned) => {
+                let mut queue = poisoned.into_inner();
+                let drained = queue.remove(user).unwrap_or_default();
+                cleanup_shard.queue.clear_poison();
+                drained
+            }
+        };
+        if to_remove.is_empty() {
+            return;
+        }
+        self.cleanup_queue_len
+            .fetch_sub(to_remove.len() as u64, Ordering::Relaxed);
+        let mut shard = self.shards[shard_idx].write().await;
+        let mut removed_active_entries = 0usize;
+        for (ip, pending_count) in to_remove {
+            removed_active_entries = removed_active_entries.saturating_add(
+                Self::apply_active_cleanup(&mut shard.active_ips, user, ip, pending_count),
+            );
+        }
+        Self::decrement_counter(&self.active_entry_count, removed_active_entries);
+    }
+
+    pub(super) async fn drain_cleanup_shard(&self, shard_idx: usize) {
+        let Ok(_drain_guard) = self.cleanup_drain_locks[shard_idx].try_lock() else {
+            return;
+        };
+
+        let cleanup_shard = &self.cleanup_shards[shard_idx];
+        let to_remove = {
+            match cleanup_shard.queue.lock() {
+                Ok(mut queue) => {
+                    if queue.is_empty() {
+                        return;
+                    }
+                    let mut drained =
+                        HashMap::with_capacity(queue.len().min(CLEANUP_DRAIN_BATCH_LIMIT));
+                    for _ in 0..CLEANUP_DRAIN_BATCH_LIMIT {
+                        let Some((user, ip, count)) = Self::pop_one_cleanup(&mut queue) else {
+                            break;
+                        };
+                        self.cleanup_queue_len.fetch_sub(1, Ordering::Relaxed);
+                        drained.insert((user, ip), count);
+                    }
+                    drained
+                }
+                Err(poisoned) => {
+                    let mut queue = poisoned.into_inner();
+                    if queue.is_empty() {
+                        cleanup_shard.queue.clear_poison();
+                        return;
+                    }
+                    let mut drained =
+                        HashMap::with_capacity(queue.len().min(CLEANUP_DRAIN_BATCH_LIMIT));
+                    for _ in 0..CLEANUP_DRAIN_BATCH_LIMIT {
+                        let Some((user, ip, count)) = Self::pop_one_cleanup(&mut queue) else {
+                            break;
+                        };
+                        self.cleanup_queue_len.fetch_sub(1, Ordering::Relaxed);
+                        drained.insert((user, ip), count);
+                    }
+                    cleanup_shard.queue.clear_poison();
+                    drained
+                }
+            }
+        };
+        drop(_drain_guard);
+        if to_remove.is_empty() {
+            return;
+        }
+
+        let mut shard = self.shards[shard_idx].write().await;
+        let mut removed_active_entries = 0usize;
+        for ((user, ip), pending_count) in to_remove {
+            removed_active_entries = removed_active_entries.saturating_add(
+                Self::apply_active_cleanup(&mut shard.active_ips, &user, ip, pending_count),
+            );
+        }
+        Self::decrement_counter(&self.active_entry_count, removed_active_entries);
+    }
+}

src/ip_tracker/snapshot.rs

@@ -0,0 +1,309 @@
+use super::*;
+
+impl UserIpTracker {
+    pub(super) async fn maybe_compact_empty_users(&self) {
+        const COMPACT_INTERVAL_SECS: u64 = 60;
+        let now_epoch_secs = Self::now_epoch_secs();
+        let last_compact_epoch_secs = self.last_compact_epoch_secs.load(Ordering::Relaxed);
+        if now_epoch_secs.saturating_sub(last_compact_epoch_secs) < COMPACT_INTERVAL_SECS {
+            return;
+        }
+        if self
+            .last_compact_epoch_secs
+            .compare_exchange(
+                last_compact_epoch_secs,
+                now_epoch_secs,
+                Ordering::AcqRel,
+                Ordering::Relaxed,
+            )
+            .is_err()
+        {
+            return;
+        }
+
+        let window = self.limit_window();
+        let now = Instant::now();
+        for shard_lock in self.shards.iter() {
+            let mut shard = shard_lock.write().await;
+            let mut pruned_recent_entries = 0usize;
+            for user_recent in shard.recent_ips.values_mut() {
+                pruned_recent_entries = pruned_recent_entries.saturating_add(Self::prune_recent(
+                    user_recent,
+                    now,
+                    window,
+                ));
+            }
+            Self::decrement_counter(&self.recent_entry_count, pruned_recent_entries);
+
+            let mut users = Vec::<String>::with_capacity(
+                shard
+                    .active_ips
+                    .len()
+                    .saturating_add(shard.recent_ips.len()),
+            );
+            users.extend(shard.active_ips.keys().cloned());
+            for user in shard.recent_ips.keys() {
+                if !shard.active_ips.contains_key(user) {
+                    users.push(user.clone());
+                }
+            }
+
+            for user in users {
+                let active_empty = shard
+                    .active_ips
+                    .get(&user)
+                    .map(|ips| ips.is_empty())
+                    .unwrap_or(true);
+                let recent_empty = shard
+                    .recent_ips
+                    .get(&user)
+                    .map(|ips| ips.is_empty())
+                    .unwrap_or(true);
+                if active_empty && recent_empty {
+                    shard.active_ips.remove(&user);
+                    shard.recent_ips.remove(&user);
+                }
+            }
+        }
+    }
+
+    pub async fn run_periodic_maintenance(self: Arc<Self>) {
+        let mut interval = tokio::time::interval(Duration::from_secs(1));
+        loop {
+            interval.tick().await;
+            self.drain_cleanup_queue().await;
+            self.maybe_compact_empty_users().await;
+        }
+    }
+
+    pub async fn memory_stats(&self) -> UserIpTrackerMemoryStats {
+        let cleanup_queue_len = self.cleanup_queue_len.load(Ordering::Relaxed) as usize;
+        let mut active_users = 0usize;
+        let mut recent_users = 0usize;
+        let mut active_entries = 0usize;
+        let mut recent_entries = 0usize;
+        for shard_lock in self.shards.iter() {
+            let shard = shard_lock.read().await;
+            active_users = active_users.saturating_add(shard.active_ips.len());
+            recent_users = recent_users.saturating_add(shard.recent_ips.len());
+            active_entries =
+                active_entries.saturating_add(shard.active_ips.values().map(HashMap::len).sum());
+            recent_entries =
+                recent_entries.saturating_add(shard.recent_ips.values().map(HashMap::len).sum());
+        }
+
+        UserIpTrackerMemoryStats {
+            active_users,
+            recent_users,
+            active_entries,
+            recent_entries,
+            cleanup_queue_len,
+            active_cap_rejects: self.active_cap_rejects.load(Ordering::Relaxed),
+            recent_cap_rejects: self.recent_cap_rejects.load(Ordering::Relaxed),
+            cleanup_deferred_releases: self.cleanup_deferred_releases.load(Ordering::Relaxed),
+        }
+    }
+
+    pub async fn get_recent_counts_for_users(&self, users: &[String]) -> HashMap<String, usize> {
+        self.drain_cleanup_queue().await;
+        self.get_recent_counts_for_users_snapshot(users).await
+    }
+
+    pub(crate) async fn get_recent_counts_for_users_snapshot(
+        &self,
+        users: &[String],
+    ) -> HashMap<String, usize> {
+        let window = self.limit_window();
+        let now = Instant::now();
+
+        let mut counts = HashMap::with_capacity(users.len());
+        for user in users {
+            let shard_idx = Self::shard_idx(user);
+            let shard = self.shards[shard_idx].read().await;
+            let count = if let Some(user_recent) = shard.recent_ips.get(user) {
+                user_recent
+                    .values()
+                    .filter(|seen_at| now.duration_since(**seen_at) <= window)
+                    .count()
+            } else {
+                0
+            };
+            counts.insert(user.clone(), count);
+        }
+        counts
+    }
+
+    pub async fn get_active_ips_for_users(&self, users: &[String]) -> HashMap<String, Vec<IpAddr>> {
+        self.drain_cleanup_queue().await;
+        let mut out = HashMap::with_capacity(users.len());
+        for user in users {
+            let shard_idx = Self::shard_idx(user);
+            let shard = self.shards[shard_idx].read().await;
+            let mut ips = shard
+                .active_ips
+                .get(user)
+                .map(|per_ip| per_ip.keys().copied().collect::<Vec<_>>())
+                .unwrap_or_else(Vec::new);
+            ips.sort();
+            out.insert(user.clone(), ips);
+        }
+        out
+    }
+
+    pub async fn get_recent_ips_for_users(&self, users: &[String]) -> HashMap<String, Vec<IpAddr>> {
+        self.drain_cleanup_queue().await;
+        let window = self.limit_window();
+        let now = Instant::now();
+
+        let mut out = HashMap::with_capacity(users.len());
+        for user in users {
+            let shard_idx = Self::shard_idx(user);
+            let shard = self.shards[shard_idx].read().await;
+            let mut ips = if let Some(user_recent) = shard.recent_ips.get(user) {
+                user_recent
+                    .iter()
+                    .filter(|(_, seen_at)| now.duration_since(**seen_at) <= window)
+                    .map(|(ip, _)| *ip)
+                    .collect::<Vec<_>>()
+            } else {
+                Vec::new()
+            };
+            ips.sort();
+            out.insert(user.clone(), ips);
+        }
+        out
+    }
+
+    pub async fn get_active_ip_count(&self, username: &str) -> usize {
+        self.drain_cleanup_queue().await;
+        let shard_idx = Self::shard_idx(username);
+        let shard = self.shards[shard_idx].read().await;
+        shard
+            .active_ips
+            .get(username)
+            .map(|ips| ips.len())
+            .unwrap_or(0)
+    }
+
+    pub async fn get_active_ips(&self, username: &str) -> Vec<IpAddr> {
+        self.drain_cleanup_queue().await;
+        let shard_idx = Self::shard_idx(username);
+        let shard = self.shards[shard_idx].read().await;
+        shard
+            .active_ips
+            .get(username)
+            .map(|ips| ips.keys().copied().collect())
+            .unwrap_or_else(Vec::new)
+    }
+
+    pub async fn get_stats(&self) -> Vec<(String, usize, usize)> {
+        self.drain_cleanup_queue().await;
+        self.get_stats_snapshot().await
+    }
+
+    pub(crate) async fn get_stats_snapshot(&self) -> Vec<(String, usize, usize)> {
+        let mut active_counts = Vec::new();
+        for shard_lock in self.shards.iter() {
+            let shard = shard_lock.read().await;
+            active_counts.extend(
+                shard
+                    .active_ips
+                    .iter()
+                    .map(|(username, user_ips)| (username.clone(), user_ips.len())),
+            );
+        }
+
+        let mut stats = Vec::with_capacity(active_counts.len());
+        for (username, active_count) in active_counts {
+            let limit = self.user_limit(&username).unwrap_or(0);
+            stats.push((username, active_count, limit));
+        }
+
+        stats.sort_by(|a, b| a.0.cmp(&b.0));
+        stats
+    }
+
+    pub async fn clear_user_ips(&self, username: &str) {
+        let shard_idx = Self::shard_idx(username);
+        let mut shard = self.shards[shard_idx].write().await;
+        let removed_active_entries = shard
+            .active_ips
+            .remove(username)
+            .map(|ips| ips.len())
+            .unwrap_or(0);
+        Self::decrement_counter(&self.active_entry_count, removed_active_entries);
+
+        let removed_recent_entries = shard
+            .recent_ips
+            .remove(username)
+            .map(|ips| ips.len())
+            .unwrap_or(0);
+        Self::decrement_counter(&self.recent_entry_count, removed_recent_entries);
+    }
+
+    pub async fn clear_all(&self) {
+        for shard_lock in self.shards.iter() {
+            let mut shard = shard_lock.write().await;
+            shard.active_ips.clear();
+            shard.recent_ips.clear();
+        }
+        self.active_entry_count.store(0, Ordering::Relaxed);
+        self.recent_entry_count.store(0, Ordering::Relaxed);
+        for cleanup_shard in self.cleanup_shards.iter() {
+            match cleanup_shard.queue.lock() {
+                Ok(mut queue) => queue.clear(),
+                Err(poisoned) => {
+                    poisoned.into_inner().clear();
+                    cleanup_shard.queue.clear_poison();
+                }
+            }
+        }
+        self.cleanup_queue_len.store(0, Ordering::Relaxed);
+    }
+
+    pub async fn is_ip_active(&self, username: &str, ip: IpAddr) -> bool {
+        self.drain_cleanup_queue().await;
+        let shard_idx = Self::shard_idx(username);
+        let shard = self.shards[shard_idx].read().await;
+        shard
+            .active_ips
+            .get(username)
+            .map(|ips| ips.contains_key(&ip))
+            .unwrap_or(false)
+    }
+
+    pub async fn get_user_limit(&self, username: &str) -> Option<usize> {
+        self.user_limit(username)
+    }
+
+    pub async fn format_stats(&self) -> String {
+        let stats = self.get_stats().await;
+
+        if stats.is_empty() {
+            return String::from("No active users");
+        }
+
+        let mut output = String::from("User IP Statistics:\n");
+        output.push_str("==================\n");
+
+        for (username, active_count, limit) in stats {
+            output.push_str(&format!(
+                "User: {:<20} Active IPs: {}/{}\n",
+                username,
+                active_count,
+                if limit > 0 {
+                    limit.to_string()
+                } else {
+                    "unlimited".to_string()
+                }
+            ));
+
+            let ips = self.get_active_ips(&username).await;
+            for ip in ips {
+                output.push_str(&format!("  - {}\n", ip));
+            }
+        }
+
+        output
+    }
+}

src/ip_tracker/tests.rs

@@ -0,0 +1,385 @@
+use super::*;
+use std::net::{IpAddr, Ipv4Addr, Ipv6Addr};
+use std::sync::atomic::Ordering;
+
+fn test_ipv4(oct1: u8, oct2: u8, oct3: u8, oct4: u8) -> IpAddr {
+    IpAddr::V4(Ipv4Addr::new(oct1, oct2, oct3, oct4))
+}
+
+fn test_ipv6() -> IpAddr {
+    IpAddr::V6(Ipv6Addr::new(0x2001, 0xdb8, 0, 0, 0, 0, 0, 1))
+}
+
+#[tokio::test]
+async fn test_basic_ip_limit() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 2).await;
+
+    let ip1 = test_ipv4(192, 168, 1, 1);
+    let ip2 = test_ipv4(192, 168, 1, 2);
+    let ip3 = test_ipv4(192, 168, 1, 3);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip2).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip3).await.is_err());
+
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 2);
+}
+
+#[tokio::test]
+async fn test_active_window_rejects_new_ip_and_keeps_existing_session() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 1).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::ActiveWindow, 30)
+        .await;
+
+    let ip1 = test_ipv4(10, 10, 10, 1);
+    let ip2 = test_ipv4(10, 10, 10, 2);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.is_ip_active("test_user", ip1).await);
+    assert!(tracker.check_and_add("test_user", ip2).await.is_err());
+
+    // Existing session remains active; only new unique IP is denied.
+    assert!(tracker.is_ip_active("test_user", ip1).await);
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 1);
+}
+
+#[tokio::test]
+async fn test_reconnection_from_same_ip() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 2).await;
+
+    let ip1 = test_ipv4(192, 168, 1, 1);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 1);
+}
+
+#[tokio::test]
+async fn test_same_ip_disconnect_keeps_active_while_other_session_alive() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 2).await;
+
+    let ip1 = test_ipv4(192, 168, 1, 1);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 1);
+
+    tracker.remove_ip("test_user", ip1).await;
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 1);
+
+    tracker.remove_ip("test_user", ip1).await;
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 0);
+}
+
+#[tokio::test]
+async fn test_ip_removal() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 2).await;
+
+    let ip1 = test_ipv4(192, 168, 1, 1);
+    let ip2 = test_ipv4(192, 168, 1, 2);
+    let ip3 = test_ipv4(192, 168, 1, 3);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip2).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip3).await.is_err());
+
+    tracker.remove_ip("test_user", ip1).await;
+
+    assert!(tracker.check_and_add("test_user", ip3).await.is_ok());
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 2);
+}
+
+#[tokio::test]
+async fn test_no_limit() {
+    let tracker = UserIpTracker::new();
+
+    let ip1 = test_ipv4(192, 168, 1, 1);
+    let ip2 = test_ipv4(192, 168, 1, 2);
+    let ip3 = test_ipv4(192, 168, 1, 3);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip2).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip3).await.is_ok());
+
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 3);
+}
+
+#[tokio::test]
+async fn test_multiple_users() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("user1", 2).await;
+    tracker.set_user_limit("user2", 1).await;
+
+    let ip1 = test_ipv4(192, 168, 1, 1);
+    let ip2 = test_ipv4(192, 168, 1, 2);
+
+    assert!(tracker.check_and_add("user1", ip1).await.is_ok());
+    assert!(tracker.check_and_add("user1", ip2).await.is_ok());
+
+    assert!(tracker.check_and_add("user2", ip1).await.is_ok());
+    assert!(tracker.check_and_add("user2", ip2).await.is_err());
+}
+
+#[tokio::test]
+async fn test_ipv6_support() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 2).await;
+
+    let ipv4 = test_ipv4(192, 168, 1, 1);
+    let ipv6 = test_ipv6();
+
+    assert!(tracker.check_and_add("test_user", ipv4).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ipv6).await.is_ok());
+
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 2);
+}
+
+#[tokio::test]
+async fn test_get_active_ips() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 3).await;
+
+    let ip1 = test_ipv4(192, 168, 1, 1);
+    let ip2 = test_ipv4(192, 168, 1, 2);
+
+    tracker.check_and_add("test_user", ip1).await.unwrap();
+    tracker.check_and_add("test_user", ip2).await.unwrap();
+
+    let active_ips = tracker.get_active_ips("test_user").await;
+    assert_eq!(active_ips.len(), 2);
+    assert!(active_ips.contains(&ip1));
+    assert!(active_ips.contains(&ip2));
+}
+
+#[tokio::test]
+async fn test_stats() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("user1", 3).await;
+    tracker.set_user_limit("user2", 2).await;
+
+    let ip1 = test_ipv4(192, 168, 1, 1);
+    let ip2 = test_ipv4(192, 168, 1, 2);
+
+    tracker.check_and_add("user1", ip1).await.unwrap();
+    tracker.check_and_add("user2", ip2).await.unwrap();
+
+    let stats = tracker.get_stats().await;
+    assert_eq!(stats.len(), 2);
+
+    assert!(stats.iter().any(|(name, _, _)| name == "user1"));
+    assert!(stats.iter().any(|(name, _, _)| name == "user2"));
+}
+
+#[tokio::test]
+async fn test_clear_user_ips() {
+    let tracker = UserIpTracker::new();
+    let ip1 = test_ipv4(192, 168, 1, 1);
+
+    tracker.check_and_add("test_user", ip1).await.unwrap();
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 1);
+
+    tracker.clear_user_ips("test_user").await;
+    assert_eq!(tracker.get_active_ip_count("test_user").await, 0);
+}
+
+#[tokio::test]
+async fn test_is_ip_active() {
+    let tracker = UserIpTracker::new();
+    let ip1 = test_ipv4(192, 168, 1, 1);
+    let ip2 = test_ipv4(192, 168, 1, 2);
+
+    tracker.check_and_add("test_user", ip1).await.unwrap();
+
+    assert!(tracker.is_ip_active("test_user", ip1).await);
+    assert!(!tracker.is_ip_active("test_user", ip2).await);
+}
+
+#[tokio::test]
+async fn test_load_limits_from_config() {
+    let tracker = UserIpTracker::new();
+
+    let mut config_limits = HashMap::new();
+    config_limits.insert("user1".to_string(), 5);
+    config_limits.insert("user2".to_string(), 3);
+
+    tracker.load_limits(0, &config_limits).await;
+
+    assert_eq!(tracker.get_user_limit("user1").await, Some(5));
+    assert_eq!(tracker.get_user_limit("user2").await, Some(3));
+    assert_eq!(tracker.get_user_limit("user3").await, None);
+}
+
+#[tokio::test]
+async fn test_load_limits_replaces_previous_map() {
+    let tracker = UserIpTracker::new();
+
+    let mut first = HashMap::new();
+    first.insert("user1".to_string(), 2);
+    first.insert("user2".to_string(), 3);
+    tracker.load_limits(0, &first).await;
+
+    let mut second = HashMap::new();
+    second.insert("user2".to_string(), 5);
+    tracker.load_limits(0, &second).await;
+
+    assert_eq!(tracker.get_user_limit("user1").await, None);
+    assert_eq!(tracker.get_user_limit("user2").await, Some(5));
+}
+
+#[tokio::test]
+async fn test_global_each_limit_applies_without_user_override() {
+    let tracker = UserIpTracker::new();
+    tracker.load_limits(2, &HashMap::new()).await;
+
+    let ip1 = test_ipv4(172, 16, 0, 1);
+    let ip2 = test_ipv4(172, 16, 0, 2);
+    let ip3 = test_ipv4(172, 16, 0, 3);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip2).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip3).await.is_err());
+    assert_eq!(tracker.get_user_limit("test_user").await, Some(2));
+}
+
+#[tokio::test]
+async fn test_user_override_wins_over_global_each_limit() {
+    let tracker = UserIpTracker::new();
+    let mut limits = HashMap::new();
+    limits.insert("test_user".to_string(), 1);
+    tracker.load_limits(3, &limits).await;
+
+    let ip1 = test_ipv4(172, 17, 0, 1);
+    let ip2 = test_ipv4(172, 17, 0, 2);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip2).await.is_err());
+    assert_eq!(tracker.get_user_limit("test_user").await, Some(1));
+}
+
+#[tokio::test]
+async fn test_time_window_mode_blocks_recent_ip_churn() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 1).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::TimeWindow, 30)
+        .await;
+
+    let ip1 = test_ipv4(10, 0, 0, 1);
+    let ip2 = test_ipv4(10, 0, 0, 2);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    tracker.remove_ip("test_user", ip1).await;
+    assert!(tracker.check_and_add("test_user", ip2).await.is_err());
+}
+
+#[tokio::test]
+async fn test_combined_mode_enforces_active_and_recent_limits() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 1).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::Combined, 30)
+        .await;
+
+    let ip1 = test_ipv4(10, 0, 1, 1);
+    let ip2 = test_ipv4(10, 0, 1, 2);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    assert!(tracker.check_and_add("test_user", ip2).await.is_err());
+
+    tracker.remove_ip("test_user", ip1).await;
+    assert!(tracker.check_and_add("test_user", ip2).await.is_err());
+}
+
+#[tokio::test]
+async fn test_time_window_expires() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 1).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::TimeWindow, 1)
+        .await;
+
+    let ip1 = test_ipv4(10, 1, 0, 1);
+    let ip2 = test_ipv4(10, 1, 0, 2);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    tracker.remove_ip("test_user", ip1).await;
+    assert!(tracker.check_and_add("test_user", ip2).await.is_err());
+
+    tokio::time::sleep(Duration::from_millis(1100)).await;
+    assert!(tracker.check_and_add("test_user", ip2).await.is_ok());
+}
+
+#[tokio::test]
+async fn test_memory_stats_reports_queue_and_entry_counts() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 4).await;
+    let ip1 = test_ipv4(10, 2, 0, 1);
+    let ip2 = test_ipv4(10, 2, 0, 2);
+
+    tracker.check_and_add("test_user", ip1).await.unwrap();
+    tracker.check_and_add("test_user", ip2).await.unwrap();
+    tracker.enqueue_cleanup("test_user".to_string(), ip1);
+
+    let snapshot = tracker.memory_stats().await;
+    assert_eq!(snapshot.active_users, 1);
+    assert_eq!(snapshot.recent_users, 1);
+    assert_eq!(snapshot.active_entries, 2);
+    assert_eq!(snapshot.recent_entries, 2);
+    assert_eq!(snapshot.cleanup_queue_len, 1);
+}
+
+#[tokio::test]
+async fn test_compact_prunes_stale_recent_entries() {
+    let tracker = UserIpTracker::new();
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::TimeWindow, 1)
+        .await;
+
+    let stale_user = "stale-user".to_string();
+    let stale_ip = test_ipv4(10, 3, 0, 1);
+    {
+        let shard_idx = UserIpTracker::shard_idx(&stale_user);
+        let mut shard = tracker.shards[shard_idx].write().await;
+        shard
+            .recent_ips
+            .entry(stale_user.clone())
+            .or_insert_with(HashMap::new)
+            .insert(stale_ip, Instant::now() - Duration::from_secs(5));
+    }
+
+    tracker.last_compact_epoch_secs.store(0, Ordering::Relaxed);
+    tracker
+        .check_and_add("trigger-user", test_ipv4(10, 3, 0, 2))
+        .await
+        .unwrap();
+
+    let shard_idx = UserIpTracker::shard_idx(&stale_user);
+    let shard = tracker.shards[shard_idx].read().await;
+    let stale_exists = shard
+        .recent_ips
+        .get(&stale_user)
+        .map(|ips| ips.contains_key(&stale_ip))
+        .unwrap_or(false);
+    assert!(!stale_exists);
+}
+
+#[tokio::test]
+async fn test_time_window_allows_same_ip_reconnect() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("test_user", 1).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::TimeWindow, 1)
+        .await;
+
+    let ip1 = test_ipv4(10, 4, 0, 1);
+
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+    tracker.remove_ip("test_user", ip1).await;
+    assert!(tracker.check_and_add("test_user", ip1).await.is_ok());
+}

src/maestro/admission.rs

@@ -1,7 +1,7 @@
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
-use tokio::sync::watch;
+use tokio::sync::{RwLock, watch};
 use tracing::{info, warn};
 
 use crate::config::ProxyConfig;
@@ -14,24 +14,32 @@ const RUNTIME_FALLBACK_AFTER: Duration = Duration::from_secs(6);
 pub(crate) async fn configure_admission_gate(
     config: &Arc<ProxyConfig>,
     me_pool: Option<Arc<MePool>>,
+    me_pool_runtime: Arc<RwLock<Option<Arc<MePool>>>>,
     route_runtime: Arc<RouteRuntimeController>,
     admission_tx: &watch::Sender<bool>,
     config_rx: watch::Receiver<Arc<ProxyConfig>>,
     me_ready_rx: watch::Receiver<u64>,
 ) {
     if config.general.use_middle_proxy {
-        if let Some(pool) = me_pool.as_ref() {
-            let initial_ready = pool.admission_ready_conditional_cast().await;
+        if me_pool.is_some() || config.general.me2dc_fallback {
+            let initial_pool = match me_pool.as_ref() {
+                Some(pool) => Some(pool.clone()),
+                None => me_pool_runtime.read().await.clone(),
+            };
+            let initial_ready = match initial_pool.as_ref() {
+                Some(pool) => pool.admission_ready_conditional_cast().await,
+                None => false,
+            };
             let mut fallback_enabled = config.general.me2dc_fallback;
             let mut fast_fallback_enabled = fallback_enabled && config.general.me2dc_fast;
             let (initial_gate_open, initial_route_mode, initial_fallback_reason) = if initial_ready
             {
                 (true, RelayRouteMode::Middle, None)
-            } else if fast_fallback_enabled {
+            } else if fallback_enabled {
                 (
                     true,
                     RelayRouteMode::Direct,
-                    Some("fast_not_ready_fallback"),
+                    Some("startup_direct_fallback"),
                 )
             } else {
                 (false, RelayRouteMode::Middle, None)
@@ -49,7 +57,8 @@ pub(crate) async fn configure_admission_gate(
                 warn!("Conditional-admission gate: closed / ME pool is NOT ready)");
             }
 
-            let pool_for_gate = pool.clone();
+            let mut pool_for_gate = initial_pool;
+            let pool_runtime_for_gate = me_pool_runtime.clone();
             let admission_tx_gate = admission_tx.clone();
             let route_runtime_gate = route_runtime.clone();
             let mut config_rx_gate = config_rx.clone();
@@ -83,12 +92,27 @@ pub(crate) async fn configure_admission_gate(
                         }
                         _ = tokio::time::sleep(Duration::from_millis(admission_poll_ms)) => {}
                     }
-                    let ready = pool_for_gate.admission_ready_conditional_cast().await;
+                    if pool_for_gate.is_none() {
+                        pool_for_gate = pool_runtime_for_gate.read().await.clone();
+                    }
+                    let ready = match pool_for_gate.as_ref() {
+                        Some(pool) => pool.admission_ready_conditional_cast().await,
+                        None => false,
+                    };
                     let now = Instant::now();
                     let (next_gate_open, next_route_mode, next_fallback_reason) = if ready {
                         ready_observed = true;
                         not_ready_since = None;
+                        if let Some(pool) = pool_for_gate.as_ref() {
+                            pool.set_runtime_ready(true);
+                        }
                         (true, RelayRouteMode::Middle, None)
+                    } else if fallback_enabled && !ready_observed {
+                        (
+                            true,
+                            RelayRouteMode::Direct,
+                            Some("startup_direct_fallback"),
+                        )
                     } else if fast_fallback_enabled {
                         (
                             true,
@@ -122,7 +146,14 @@ pub(crate) async fn configure_admission_gate(
                                 );
                             } else {
                                 let fallback_reason = next_fallback_reason.unwrap_or("unknown");
-                                if fallback_reason == "strict_grace_fallback" {
+                                if fallback_reason == "startup_direct_fallback" {
+                                    warn!(
+                                        target_mode = route_mode.as_str(),
+                                        cutover_generation = snapshot.generation,
+                                        fallback_reason,
+                                        "ME pool not-ready during startup; routing new sessions via Direct-DC"
+                                    );
+                                } else if fallback_reason == "strict_grace_fallback" {
                                     let fallback_after = if ready_observed {
                                         RUNTIME_FALLBACK_AFTER
                                     } else {

src/maestro/helpers.rs

@@ -15,6 +15,13 @@ use crate::transport::middle_proxy::{
     save_proxy_config_cache,
 };
 
+const MAESTRO_COLOR: &str = "\x1b[92m";
+const COLOR_RESET: &str = "\x1b[0m";
+
+pub(crate) fn print_maestro_line(message: impl AsRef<str>) {
+    eprintln!("{MAESTRO_COLOR}MAESTRO{COLOR_RESET}: {}", message.as_ref());
+}
+
 pub(crate) fn resolve_runtime_config_path(
     config_path_cli: &str,
     startup_cwd: &Path,
@@ -501,7 +508,7 @@ mod tests {
 }
 
 pub(crate) fn print_proxy_links(host: &str, port: u16, config: &ProxyConfig) {
-    info!(target: "telemt::links", "--- Proxy Links ({}) ---", host);
+    print_maestro_line(format!("Proxy links ({host})"));
     for user_name in config
         .general
         .links
@@ -509,20 +516,16 @@ pub(crate) fn print_proxy_links(host: &str, port: u16, config: &ProxyConfig) {
         .resolve_users(&config.access.users)
     {
         if let Some(secret) = config.access.users.get(user_name) {
-            info!(target: "telemt::links", "User: {}", user_name);
+            print_maestro_line(format!("User: {user_name}"));
             if config.general.modes.classic {
-                info!(
-                    target: "telemt::links",
-                    "  Classic: tg://proxy?server={}&port={}&secret={}",
-                    host, port, secret
-                );
+                print_maestro_line(format!(
+                    "Classic: tg://proxy?server={host}&port={port}&secret={secret}"
+                ));
             }
             if config.general.modes.secure {
-                info!(
-                    target: "telemt::links",
-                    "  DD:      tg://proxy?server={}&port={}&secret=dd{}",
-                    host, port, secret
-                );
+                print_maestro_line(format!(
+                    "DD: tg://proxy?server={host}&port={port}&secret=dd{secret}"
+                ));
             }
             if config.general.modes.tls {
                 let mut domains = Vec::with_capacity(1 + config.censorship.tls_domains.len());
@@ -535,18 +538,15 @@ pub(crate) fn print_proxy_links(host: &str, port: u16, config: &ProxyConfig) {
 
                 for domain in domains {
                     let domain_hex = hex::encode(&domain);
-                    info!(
-                        target: "telemt::links",
-                        "  EE-TLS:  tg://proxy?server={}&port={}&secret=ee{}{}",
-                        host, port, secret, domain_hex
-                    );
+                    print_maestro_line(format!(
+                        "EE-TLS: tg://proxy?server={host}&port={port}&secret=ee{secret}{domain_hex}"
+                    ));
                 }
             }
         } else {
             warn!(target: "telemt::links", "User '{}' in show_link not found", user_name);
         }
     }
-    info!(target: "telemt::links", "------------------------");
 }
 
 pub(crate) async fn write_beobachten_snapshot(path: &str, payload: &str) -> std::io::Result<()> {

src/maestro/listeners.rs

@@ -6,7 +6,7 @@ use std::time::Duration;
 use tokio::net::TcpListener;
 #[cfg(unix)]
 use tokio::net::UnixListener;
-use tokio::sync::{Semaphore, watch};
+use tokio::sync::{RwLock, Semaphore, watch};
 use tracing::{debug, error, info, warn};
 
 use crate::config::{ProxyConfig, RstOnCloseMode};
@@ -63,6 +63,7 @@ pub(crate) async fn bind_listeners(
     buffer_pool: Arc<BufferPool>,
     rng: Arc<SecureRandom>,
     me_pool: Option<Arc<MePool>>,
+    me_pool_runtime: Arc<RwLock<Option<Arc<MePool>>>>,
     route_runtime: Arc<RouteRuntimeController>,
     tls_cache: Option<Arc<TlsFrontCache>>,
     ip_tracker: Arc<UserIpTracker>,
@@ -236,6 +237,7 @@ pub(crate) async fn bind_listeners(
         let buffer_pool = buffer_pool.clone();
         let rng = rng.clone();
         let me_pool = me_pool.clone();
+        let me_pool_runtime = me_pool_runtime.clone();
         let route_runtime = route_runtime.clone();
         let tls_cache = tls_cache.clone();
         let ip_tracker = ip_tracker.clone();
@@ -298,6 +300,7 @@ pub(crate) async fn bind_listeners(
                         let buffer_pool = buffer_pool.clone();
                         let rng = rng.clone();
                         let me_pool = me_pool.clone();
+                        let me_pool_runtime = me_pool_runtime.clone();
                         let route_runtime = route_runtime.clone();
                         let tls_cache = tls_cache.clone();
                         let ip_tracker = ip_tracker.clone();
@@ -307,7 +310,8 @@ pub(crate) async fn bind_listeners(
 
                         tokio::spawn(async move {
                             let _permit = permit;
-                            if let Err(e) = crate::proxy::client::handle_client_stream_with_shared(
+                            if let Err(e) =
+                                crate::proxy::client::handle_client_stream_with_shared_and_pool_runtime(
                                 stream,
                                 fake_peer,
                                 config,
@@ -317,6 +321,7 @@ pub(crate) async fn bind_listeners(
                                 buffer_pool,
                                 rng,
                                 me_pool,
+                                Some(me_pool_runtime),
                                 route_runtime,
                                 tls_cache,
                                 ip_tracker,
@@ -367,6 +372,7 @@ pub(crate) fn spawn_tcp_accept_loops(
     buffer_pool: Arc<BufferPool>,
     rng: Arc<SecureRandom>,
     me_pool: Option<Arc<MePool>>,
+    me_pool_runtime: Arc<RwLock<Option<Arc<MePool>>>>,
     route_runtime: Arc<RouteRuntimeController>,
     tls_cache: Option<Arc<TlsFrontCache>>,
     ip_tracker: Arc<UserIpTracker>,
@@ -383,6 +389,7 @@ pub(crate) fn spawn_tcp_accept_loops(
         let buffer_pool = buffer_pool.clone();
         let rng = rng.clone();
         let me_pool = me_pool.clone();
+        let me_pool_runtime = me_pool_runtime.clone();
         let route_runtime = route_runtime.clone();
         let tls_cache = tls_cache.clone();
         let ip_tracker = ip_tracker.clone();
@@ -449,6 +456,7 @@ pub(crate) fn spawn_tcp_accept_loops(
                         let buffer_pool = buffer_pool.clone();
                         let rng = rng.clone();
                         let me_pool = me_pool.clone();
+                        let me_pool_runtime = me_pool_runtime.clone();
                         let route_runtime = route_runtime.clone();
                         let tls_cache = tls_cache.clone();
                         let ip_tracker = ip_tracker.clone();
@@ -470,6 +478,7 @@ pub(crate) fn spawn_tcp_accept_loops(
                                 buffer_pool,
                                 rng,
                                 me_pool,
+                                Some(me_pool_runtime),
                                 route_runtime,
                                 tls_cache,
                                 ip_tracker,

src/maestro/mod.rs

@@ -36,10 +36,10 @@ use crate::network::probe::{decide_network_capabilities, log_probe_result, run_p
 use crate::proxy::route_mode::{RelayRouteMode, RouteRuntimeController};
 use crate::proxy::shared_state::ProxySharedState;
 use crate::startup::{
-    COMPONENT_API_BOOTSTRAP, COMPONENT_CONFIG_LOAD, COMPONENT_ME_POOL_CONSTRUCT,
-    COMPONENT_ME_POOL_INIT_STAGE1, COMPONENT_ME_PROXY_CONFIG_V4, COMPONENT_ME_PROXY_CONFIG_V6,
-    COMPONENT_ME_SECRET_FETCH, COMPONENT_NETWORK_PROBE, COMPONENT_TRACING_INIT, StartupMeStatus,
-    StartupTracker,
+    COMPONENT_API_BOOTSTRAP, COMPONENT_CONFIG_LOAD, COMPONENT_DC_CONNECTIVITY_PING,
+    COMPONENT_ME_CONNECTIVITY_PING, COMPONENT_ME_POOL_CONSTRUCT, COMPONENT_ME_POOL_INIT_STAGE1,
+    COMPONENT_ME_PROXY_CONFIG_V4, COMPONENT_ME_PROXY_CONFIG_V6, COMPONENT_ME_SECRET_FETCH,
+    COMPONENT_NETWORK_PROBE, COMPONENT_TRACING_INIT, StartupMeStatus, StartupTracker,
 };
 use crate::stats::beobachten::BeobachtenStore;
 use crate::stats::telemetry::TelemetryPolicy;
@@ -47,7 +47,9 @@ use crate::stats::{ReplayChecker, Stats};
 use crate::stream::BufferPool;
 use crate::transport::UpstreamManager;
 use crate::transport::middle_proxy::MePool;
-use helpers::{parse_cli, resolve_runtime_base_dir, resolve_runtime_config_path};
+use helpers::{
+    parse_cli, print_maestro_line, resolve_runtime_base_dir, resolve_runtime_config_path,
+};
 
 #[cfg(unix)]
 use crate::daemon::{DaemonOptions, PidFile, drop_privileges};
@@ -325,7 +327,9 @@ async fn run_telemt_core(
         config.general.log_level.clone()
     };
 
-    let (filter_layer, filter_handle) = reload::Layer::new(EnvFilter::new("info"));
+    let initial_filter_spec = runtime_tasks::log_filter_spec(has_rust_log, &effective_log_level);
+    let (filter_layer, filter_handle) =
+        reload::Layer::new(EnvFilter::new(initial_filter_spec.clone()));
     startup_tracker
         .start_component(
             COMPONENT_TRACING_INIT,
@@ -356,7 +360,7 @@ async fn run_telemt_core(
                 destination: log_destination,
                 disable_colors: true,
             };
-            let (_, guard) = crate::logging::init_logging(&logging_opts, "info");
+            let (_, guard) = crate::logging::init_logging(&logging_opts, &initial_filter_spec);
             _logging_guard = Some(guard);
         }
         crate::logging::LogDestination::File { .. } => {
@@ -365,7 +369,7 @@ async fn run_telemt_core(
                 destination: log_destination,
                 disable_colors: true,
             };
-            let (_, guard) = crate::logging::init_logging(&logging_opts, "info");
+            let (_, guard) = crate::logging::init_logging(&logging_opts, &initial_filter_spec);
             _logging_guard = Some(guard);
         }
     }
@@ -377,7 +381,7 @@ async fn run_telemt_core(
         )
         .await;
 
-    info!("Telemt MTProxy v{}", env!("CARGO_PKG_VERSION"));
+    print_maestro_line(format!("Telemt MTProxy v{}", env!("CARGO_PKG_VERSION")));
     info!("Log level: {}", effective_log_level);
     if config.general.disable_colors {
         info!("Colors: disabled");
@@ -461,12 +465,13 @@ async fn run_telemt_core(
 
     let (api_config_tx, api_config_rx) = watch::channel(Arc::new(config.clone()));
     let (detected_ips_tx, detected_ips_rx) = watch::channel((None::<IpAddr>, None::<IpAddr>));
-    let initial_admission_open = !config.general.use_middle_proxy;
+    let initial_direct_first = config.general.use_middle_proxy && config.general.me2dc_fallback;
+    let initial_admission_open = !config.general.use_middle_proxy || initial_direct_first;
     let (admission_tx, admission_rx) = watch::channel(initial_admission_open);
-    let initial_route_mode = if config.general.use_middle_proxy {
-        RelayRouteMode::Middle
-    } else {
+    let initial_route_mode = if !config.general.use_middle_proxy || initial_direct_first {
         RelayRouteMode::Direct
+    } else {
+        RelayRouteMode::Middle
     };
     let route_runtime = Arc::new(RouteRuntimeController::new(initial_route_mode));
     let api_me_pool = Arc::new(RwLock::new(None::<Arc<MePool>>));
@@ -602,8 +607,9 @@ async fn run_telemt_core(
     let me_init_retry_attempts = config.general.me_init_retry_attempts;
     if use_middle_proxy && !decision.ipv4_me && !decision.ipv6_me {
         if me2dc_fallback {
-            warn!("No usable IP family for Middle Proxy detected; falling back to direct DC");
-            use_middle_proxy = false;
+            warn!(
+                "No usable IP family for Middle Proxy detected; Direct-DC startup fallback is active while ME init retries continue"
+            );
         } else {
             warn!(
                 "No usable IP family for Middle Proxy detected; me2dc_fallback=false, ME init retries stay active"
@@ -665,23 +671,34 @@ async fn run_telemt_core(
     }
 
     let (me_ready_tx, me_ready_rx) = watch::channel(0_u64);
+    let direct_first_startup = use_middle_proxy && me2dc_fallback;
 
-    let me_pool: Option<Arc<MePool>> = me_startup::initialize_me_pool(
-        use_middle_proxy,
-        &config,
-        &decision,
-        &probe,
-        &startup_tracker,
-        upstream_manager.clone(),
-        rng.clone(),
-        stats.clone(),
-        api_me_pool.clone(),
-        me_ready_tx.clone(),
-    )
-    .await;
+    let me_pool: Option<Arc<MePool>> = if direct_first_startup {
+        None
+    } else {
+        me_startup::initialize_me_pool(
+            use_middle_proxy,
+            &config,
+            &decision,
+            &probe,
+            &startup_tracker,
+            upstream_manager.clone(),
+            rng.clone(),
+            stats.clone(),
+            api_me_pool.clone(),
+            me_ready_tx.clone(),
+        )
+        .await
+    };
 
     // If ME failed to initialize, force direct-only mode.
-    if me_pool.is_some() {
+    if direct_first_startup {
+        startup_tracker.set_transport_mode("direct").await;
+        startup_tracker.set_degraded(true).await;
+        info!(
+            "Transport: Direct DC startup fallback active; Middle-End bootstrap continues in background"
+        );
+    } else if me_pool.is_some() {
         startup_tracker.set_transport_mode("middle_proxy").await;
         startup_tracker.set_degraded(false).await;
         info!("Transport: Middle-End Proxy - all DC-over-RPC");
@@ -719,18 +736,33 @@ async fn run_telemt_core(
         config.access.cidr_rate_limits.clone(),
     );
 
-    connectivity::run_startup_connectivity(
-        &config,
-        &me_pool,
-        rng.clone(),
-        &startup_tracker,
-        upstream_manager.clone(),
-        prefer_ipv6,
-        &decision,
-        process_started_at,
-        api_me_pool.clone(),
-    )
-    .await;
+    if direct_first_startup {
+        startup_tracker
+            .skip_component(
+                COMPONENT_ME_CONNECTIVITY_PING,
+                Some("deferred by direct-first startup".to_string()),
+            )
+            .await;
+        startup_tracker
+            .skip_component(
+                COMPONENT_DC_CONNECTIVITY_PING,
+                Some("background health checks active".to_string()),
+            )
+            .await;
+    } else {
+        connectivity::run_startup_connectivity(
+            &config,
+            &me_pool,
+            rng.clone(),
+            &startup_tracker,
+            upstream_manager.clone(),
+            prefer_ipv6,
+            &decision,
+            process_started_at,
+            api_me_pool.clone(),
+        )
+        .await;
+    }
 
     let runtime_watches = runtime_tasks::spawn_runtime_tasks(
         &config,
@@ -758,9 +790,72 @@ async fn run_telemt_core(
     let detected_ip_v4 = runtime_watches.detected_ip_v4;
     let detected_ip_v6 = runtime_watches.detected_ip_v6;
 
+    if direct_first_startup {
+        let config_bg = config.clone();
+        let decision_bg = decision.clone();
+        let probe_bg = probe.clone();
+        let startup_tracker_bg = startup_tracker.clone();
+        let upstream_manager_bg = upstream_manager.clone();
+        let rng_bg = rng.clone();
+        let stats_bg = stats.clone();
+        let api_me_pool_bg = api_me_pool.clone();
+        let me_ready_tx_bg = me_ready_tx.clone();
+        let config_rx_bg = config_rx.clone();
+        tokio::spawn(async move {
+            let mut bootstrap_attempt: u32 = 0;
+            loop {
+                bootstrap_attempt = bootstrap_attempt.saturating_add(1);
+                let pool = me_startup::initialize_me_pool(
+                    true,
+                    config_bg.as_ref(),
+                    &decision_bg,
+                    &probe_bg,
+                    &startup_tracker_bg,
+                    upstream_manager_bg.clone(),
+                    rng_bg.clone(),
+                    stats_bg.clone(),
+                    api_me_pool_bg.clone(),
+                    me_ready_tx_bg.clone(),
+                )
+                .await;
+                if let Some(pool) = pool {
+                    runtime_tasks::spawn_middle_proxy_runtime_tasks(
+                        config_bg.as_ref(),
+                        config_rx_bg,
+                        pool,
+                        rng_bg,
+                        me_ready_tx_bg,
+                    );
+                    break;
+                }
+                if me_init_retry_attempts > 0 && bootstrap_attempt >= me_init_retry_attempts {
+                    break;
+                }
+                tokio::time::sleep(Duration::from_secs(2)).await;
+            }
+        });
+
+        let startup_tracker_ready = startup_tracker.clone();
+        let api_me_pool_ready = api_me_pool.clone();
+        let mut me_ready_rx_transport = me_ready_tx.subscribe();
+        tokio::spawn(async move {
+            if me_ready_rx_transport.changed().await.is_ok() {
+                if let Some(pool) = api_me_pool_ready.read().await.as_ref() {
+                    pool.set_runtime_ready(true);
+                }
+                startup_tracker_ready
+                    .set_transport_mode("middle_proxy")
+                    .await;
+                startup_tracker_ready.set_degraded(false).await;
+                info!("Transport: Middle-End Proxy restored for new sessions");
+            }
+        });
+    }
+
     admission::configure_admission_gate(
         &config,
         me_pool.clone(),
+        api_me_pool.clone(),
         route_runtime.clone(),
         &admission_tx,
         config_rx.clone(),
@@ -789,6 +884,7 @@ async fn run_telemt_core(
         buffer_pool.clone(),
         rng.clone(),
         me_pool.clone(),
+        api_me_pool.clone(),
         route_runtime.clone(),
         tls_cache.clone(),
         ip_tracker.clone(),
@@ -843,6 +939,7 @@ async fn run_telemt_core(
         buffer_pool.clone(),
         rng.clone(),
         me_pool.clone(),
+        api_me_pool.clone(),
         route_runtime.clone(),
         tls_cache.clone(),
         ip_tracker.clone(),

src/maestro/runtime_tasks.rs

@@ -257,45 +257,7 @@ pub(crate) async fn spawn_runtime_tasks(
     });
 
     if let Some(pool) = me_pool {
-        let reinit_trigger_capacity = config.general.me_reinit_trigger_channel.max(1);
-        let (reinit_tx, reinit_rx) = mpsc::channel::<MeReinitTrigger>(reinit_trigger_capacity);
-
-        let pool_clone_sched = pool.clone();
-        let rng_clone_sched = rng.clone();
-        let config_rx_clone_sched = config_rx.clone();
-        let me_ready_tx_sched = me_ready_tx.clone();
-        tokio::spawn(async move {
-            crate::transport::middle_proxy::me_reinit_scheduler(
-                pool_clone_sched,
-                rng_clone_sched,
-                config_rx_clone_sched,
-                reinit_rx,
-                me_ready_tx_sched,
-            )
-            .await;
-        });
-
-        let pool_clone = pool.clone();
-        let config_rx_clone = config_rx.clone();
-        let reinit_tx_updater = reinit_tx.clone();
-        tokio::spawn(async move {
-            crate::transport::middle_proxy::me_config_updater(
-                pool_clone,
-                config_rx_clone,
-                reinit_tx_updater,
-            )
-            .await;
-        });
-
-        let config_rx_clone_rot = config_rx.clone();
-        let reinit_tx_rotation = reinit_tx.clone();
-        tokio::spawn(async move {
-            crate::transport::middle_proxy::me_rotation_task(
-                config_rx_clone_rot,
-                reinit_tx_rotation,
-            )
-            .await;
-        });
+        spawn_middle_proxy_runtime_tasks(config, config_rx.clone(), pool, rng, me_ready_tx);
     }
 
     RuntimeWatches {
@@ -306,19 +268,58 @@ pub(crate) async fn spawn_runtime_tasks(
     }
 }
 
+pub(crate) fn spawn_middle_proxy_runtime_tasks(
+    config: &ProxyConfig,
+    config_rx: watch::Receiver<Arc<ProxyConfig>>,
+    pool: Arc<MePool>,
+    rng: Arc<SecureRandom>,
+    me_ready_tx: watch::Sender<u64>,
+) {
+    let reinit_trigger_capacity = config.general.me_reinit_trigger_channel.max(1);
+    let (reinit_tx, reinit_rx) = mpsc::channel::<MeReinitTrigger>(reinit_trigger_capacity);
+
+    let pool_clone_sched = pool.clone();
+    let rng_clone_sched = rng.clone();
+    let config_rx_clone_sched = config_rx.clone();
+    let me_ready_tx_sched = me_ready_tx.clone();
+    tokio::spawn(async move {
+        crate::transport::middle_proxy::me_reinit_scheduler(
+            pool_clone_sched,
+            rng_clone_sched,
+            config_rx_clone_sched,
+            reinit_rx,
+            me_ready_tx_sched,
+        )
+        .await;
+    });
+
+    let pool_clone = pool.clone();
+    let config_rx_clone = config_rx.clone();
+    let reinit_tx_updater = reinit_tx.clone();
+    tokio::spawn(async move {
+        crate::transport::middle_proxy::me_config_updater(
+            pool_clone,
+            config_rx_clone,
+            reinit_tx_updater,
+        )
+        .await;
+    });
+
+    let config_rx_clone_rot = config_rx.clone();
+    let reinit_tx_rotation = reinit_tx.clone();
+    tokio::spawn(async move {
+        crate::transport::middle_proxy::me_rotation_task(config_rx_clone_rot, reinit_tx_rotation)
+            .await;
+    });
+}
+
 pub(crate) async fn apply_runtime_log_filter(
     has_rust_log: bool,
     effective_log_level: &LogLevel,
     filter_handle: reload::Handle<EnvFilter, tracing_subscriber::Registry>,
     mut log_level_rx: watch::Receiver<LogLevel>,
 ) {
-    let runtime_filter = if has_rust_log {
-        EnvFilter::from_default_env()
-    } else if matches!(effective_log_level, LogLevel::Silent) {
-        EnvFilter::new("warn,telemt::links=info")
-    } else {
-        EnvFilter::new(effective_log_level.to_filter_str())
-    };
+    let runtime_filter = EnvFilter::new(log_filter_spec(has_rust_log, effective_log_level));
     filter_handle
         .reload(runtime_filter)
         .expect("Failed to switch log filter");
@@ -329,7 +330,7 @@ pub(crate) async fn apply_runtime_log_filter(
                 break;
             }
             let level = log_level_rx.borrow_and_update().clone();
-            let new_filter = tracing_subscriber::EnvFilter::new(level.to_filter_str());
+            let new_filter = tracing_subscriber::EnvFilter::new(log_filter_spec(false, &level));
             if let Err(e) = filter_handle.reload(new_filter) {
                 tracing::error!("config reload: failed to update log filter: {}", e);
             }
@@ -337,6 +338,17 @@ pub(crate) async fn apply_runtime_log_filter(
     });
 }
 
+pub(crate) fn log_filter_spec(has_rust_log: bool, effective_log_level: &LogLevel) -> String {
+    if has_rust_log {
+        std::env::var("RUST_LOG")
+            .unwrap_or_else(|_| effective_log_level.to_filter_str().to_string())
+    } else if matches!(effective_log_level, LogLevel::Silent) {
+        "warn,telemt::links=info".to_string()
+    } else {
+        effective_log_level.to_filter_str().to_string()
+    }
+}
+
 pub(crate) async fn spawn_metrics_if_configured(
     config: &Arc<ProxyConfig>,
     startup_tracker: &Arc<StartupTracker>,

src/metrics.rs

@@ -726,6 +726,37 @@ async fn render_metrics(
         }
     );
 
+    let _ = writeln!(
+        out,
+        "# HELP telemt_route_cutover_parked_current Sessions currently parked in route cutover stagger delay"
+    );
+    let _ = writeln!(out, "# TYPE telemt_route_cutover_parked_current gauge");
+    let _ = writeln!(
+        out,
+        "telemt_route_cutover_parked_current{{route=\"direct\"}} {}",
+        stats.get_route_cutover_parked_direct_current()
+    );
+    let _ = writeln!(
+        out,
+        "telemt_route_cutover_parked_current{{route=\"middle\"}} {}",
+        stats.get_route_cutover_parked_middle_current()
+    );
+    let _ = writeln!(
+        out,
+        "# HELP telemt_route_cutover_parked_total Sessions parked in route cutover stagger delay"
+    );
+    let _ = writeln!(out, "# TYPE telemt_route_cutover_parked_total counter");
+    let _ = writeln!(
+        out,
+        "telemt_route_cutover_parked_total{{route=\"direct\"}} {}",
+        stats.get_route_cutover_parked_direct_total()
+    );
+    let _ = writeln!(
+        out,
+        "telemt_route_cutover_parked_total{{route=\"middle\"}} {}",
+        stats.get_route_cutover_parked_middle_total()
+    );
+
     let _ = writeln!(
         out,
         "# HELP telemt_quota_refund_bytes_total Reserved quota bytes returned before commit"

src/proxy/client.rs

@@ -11,6 +11,7 @@ use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite};
 use tokio::net::TcpStream;
+use tokio::sync::RwLock;
 use tokio::time::timeout;
 use tracing::{debug, warn};
 
@@ -452,7 +453,50 @@ where
 }
 
 #[allow(clippy::too_many_arguments)]
+#[allow(dead_code)]
 pub async fn handle_client_stream_with_shared<S>(
+    stream: S,
+    peer: SocketAddr,
+    config: Arc<ProxyConfig>,
+    stats: Arc<Stats>,
+    upstream_manager: Arc<UpstreamManager>,
+    replay_checker: Arc<ReplayChecker>,
+    buffer_pool: Arc<BufferPool>,
+    rng: Arc<SecureRandom>,
+    me_pool: Option<Arc<MePool>>,
+    route_runtime: Arc<RouteRuntimeController>,
+    tls_cache: Option<Arc<TlsFrontCache>>,
+    ip_tracker: Arc<UserIpTracker>,
+    beobachten: Arc<BeobachtenStore>,
+    shared: Arc<ProxySharedState>,
+    proxy_protocol_enabled: bool,
+) -> Result<()>
+where
+    S: AsyncRead + AsyncWrite + Unpin + Send + 'static,
+{
+    handle_client_stream_with_shared_and_pool_runtime(
+        stream,
+        peer,
+        config,
+        stats,
+        upstream_manager,
+        replay_checker,
+        buffer_pool,
+        rng,
+        me_pool,
+        None,
+        route_runtime,
+        tls_cache,
+        ip_tracker,
+        beobachten,
+        shared,
+        proxy_protocol_enabled,
+    )
+    .await
+}
+
+#[allow(clippy::too_many_arguments)]
+pub async fn handle_client_stream_with_shared_and_pool_runtime<S>(
     mut stream: S,
     peer: SocketAddr,
     config: Arc<ProxyConfig>,
@@ -462,6 +506,7 @@ pub async fn handle_client_stream_with_shared<S>(
     buffer_pool: Arc<BufferPool>,
     rng: Arc<SecureRandom>,
     me_pool: Option<Arc<MePool>>,
+    me_pool_runtime: Option<Arc<RwLock<Option<Arc<MePool>>>>>,
     route_runtime: Arc<RouteRuntimeController>,
     tls_cache: Option<Arc<TlsFrontCache>>,
     ip_tracker: Arc<UserIpTracker>,
@@ -731,6 +776,7 @@ where
                 RunningClientHandler::handle_authenticated_static_with_shared(
                     crypto_reader, crypto_writer, success,
                     upstream_manager, stats, config, buffer_pool, rng, me_pool,
+                    me_pool_runtime,
                     route_runtime.clone(),
                     local_addr, real_peer, ip_tracker.clone(),
                     shared.clone(),
@@ -791,6 +837,7 @@ where
                     buffer_pool,
                     rng,
                     me_pool,
+                    me_pool_runtime,
                     route_runtime.clone(),
                     local_addr,
                     real_peer,
@@ -846,6 +893,7 @@ pub struct RunningClientHandler {
     buffer_pool: Arc<BufferPool>,
     rng: Arc<SecureRandom>,
     me_pool: Option<Arc<MePool>>,
+    me_pool_runtime: Option<Arc<RwLock<Option<Arc<MePool>>>>>,
     route_runtime: Arc<RouteRuntimeController>,
     tls_cache: Option<Arc<TlsFrontCache>>,
     ip_tracker: Arc<UserIpTracker>,
@@ -891,6 +939,7 @@ impl ClientHandler {
             buffer_pool,
             rng,
             me_pool,
+            None,
             route_runtime,
             tls_cache,
             ip_tracker,
@@ -915,6 +964,7 @@ impl ClientHandler {
         buffer_pool: Arc<BufferPool>,
         rng: Arc<SecureRandom>,
         me_pool: Option<Arc<MePool>>,
+        me_pool_runtime: Option<Arc<RwLock<Option<Arc<MePool>>>>>,
         route_runtime: Arc<RouteRuntimeController>,
         tls_cache: Option<Arc<TlsFrontCache>>,
         ip_tracker: Arc<UserIpTracker>,
@@ -938,6 +988,7 @@ impl ClientHandler {
             buffer_pool,
             rng,
             me_pool,
+            me_pool_runtime,
             route_runtime,
             tls_cache,
             ip_tracker,
@@ -1345,6 +1396,7 @@ impl RunningClientHandler {
                 buffer_pool,
                 self.rng,
                 self.me_pool,
+                self.me_pool_runtime,
                 self.route_runtime.clone(),
                 local_addr,
                 peer,
@@ -1429,6 +1481,7 @@ impl RunningClientHandler {
                 buffer_pool,
                 self.rng,
                 self.me_pool,
+                self.me_pool_runtime,
                 self.route_runtime.clone(),
                 local_addr,
                 peer,
@@ -1472,6 +1525,7 @@ impl RunningClientHandler {
             buffer_pool,
             rng,
             me_pool,
+            None,
             route_runtime,
             local_addr,
             peer_addr,
@@ -1491,6 +1545,7 @@ impl RunningClientHandler {
         buffer_pool: Arc<BufferPool>,
         rng: Arc<SecureRandom>,
         me_pool: Option<Arc<MePool>>,
+        me_pool_runtime: Option<Arc<RwLock<Option<Arc<MePool>>>>>,
         route_runtime: Arc<RouteRuntimeController>,
         local_addr: SocketAddr,
         peer_addr: SocketAddr,
@@ -1521,15 +1576,29 @@ impl RunningClientHandler {
 
         let route_snapshot = route_runtime.snapshot();
         let session_id = rng.u64();
-        let relay_result = if config.general.use_middle_proxy
+        let selected_me_pool = if config.general.use_middle_proxy
             && matches!(route_snapshot.mode, RelayRouteMode::Middle)
         {
             if let Some(ref pool) = me_pool {
+                Some(pool.clone())
+            } else if let Some(pool_runtime) = me_pool_runtime.as_ref() {
+                pool_runtime.read().await.clone()
+            } else {
+                None
+            }
+        } else {
+            None
+        };
+
+        let relay_result = if config.general.use_middle_proxy
+            && matches!(route_snapshot.mode, RelayRouteMode::Middle)
+        {
+            if let Some(pool) = selected_me_pool {
                 handle_via_middle_proxy(
                     client_reader,
                     client_writer,
                     success,
-                    pool.clone(),
+                    pool,
                     stats.clone(),
                     config,
                     buffer_pool,

src/proxy/direct_relay.rs

@@ -358,6 +358,7 @@ where
                 delay_ms = delay.as_millis() as u64,
                 "Cutover affected direct session, closing client connection"
             );
+            let _cutover_park_lease = stats.acquire_direct_cutover_park_lease();
             tokio::time::sleep(delay).await;
             break Err(ProxyError::RouteSwitched);
         }

src/proxy/masking.rs

@@ -47,6 +47,12 @@ struct CopyOutcome {
     ended_by_eof: bool,
 }
 
+#[derive(Clone, Copy)]
+struct MaskTcpTarget<'a> {
+    host: &'a str,
+    port: u16,
+}
+
 async fn copy_with_idle_timeout<R, W>(
     reader: &mut R,
     writer: &mut W,
@@ -331,7 +337,9 @@ async fn wait_mask_outcome_budget(started: Instant, config: &ProxyConfig) {
 
 #[cfg(test)]
 mod tls_domain_mask_host_tests {
-    use super::{mask_host_for_initial_data, matching_tls_domain_for_sni};
+    use super::{
+        mask_host_for_initial_data, mask_tcp_target_for_initial_data, matching_tls_domain_for_sni,
+    };
     use crate::config::ProxyConfig;
 
     fn client_hello_with_sni(sni_host: &str) -> Vec<u8> {
@@ -410,6 +418,25 @@ mod tls_domain_mask_host_tests {
 
         assert_eq!(mask_host_for_initial_data(&config, &initial_data), "b.com");
     }
+
+    #[test]
+    fn exclusive_mask_target_overrides_only_matching_sni() {
+        let mut config = config_with_tls_domains();
+        config
+            .censorship
+            .exclusive_mask
+            .insert("b.com".to_string(), "origin-b.example:8443".to_string());
+        let b_initial_data = client_hello_with_sni("B.COM");
+        let c_initial_data = client_hello_with_sni("c.com");
+
+        let b_target = mask_tcp_target_for_initial_data(&config, &b_initial_data);
+        let c_target = mask_tcp_target_for_initial_data(&config, &c_initial_data);
+
+        assert_eq!(b_target.host, "origin-b.example");
+        assert_eq!(b_target.port, 8443);
+        assert_eq!(c_target.host, "c.com");
+        assert_eq!(c_target.port, config.censorship.mask_port);
+    }
 }
 
 /// Detect client type based on initial data
@@ -458,7 +485,92 @@ fn matching_tls_domain_for_sni<'a>(config: &'a ProxyConfig, sni: &str) -> Option
     None
 }
 
+fn parse_exclusive_mask_target(target: &str) -> Option<MaskTcpTarget<'_>> {
+    let target = target.trim();
+    if target.is_empty() {
+        return None;
+    }
+
+    if target.starts_with('[') {
+        let end = target.find(']')?;
+        if target.get(end + 1..end + 2)? != ":" {
+            return None;
+        }
+        let port = target[end + 2..].parse::<u16>().ok()?;
+        return (port > 0).then_some(MaskTcpTarget {
+            host: &target[..=end],
+            port,
+        });
+    }
+
+    let (host, port) = target.rsplit_once(':')?;
+    if host.is_empty() || host.contains(':') {
+        return None;
+    }
+    let port = port.parse::<u16>().ok()?;
+    (port > 0).then_some(MaskTcpTarget { host, port })
+}
+
+fn exclusive_mask_target_for_sni<'a>(
+    config: &'a ProxyConfig,
+    sni: &str,
+) -> Option<MaskTcpTarget<'a>> {
+    if let Some(target) = config.censorship.exclusive_mask_targets.get(sni) {
+        return Some(MaskTcpTarget {
+            host: target.host.as_str(),
+            port: target.port,
+        });
+    }
+    if let Some(target) = config.censorship.exclusive_mask.get(sni) {
+        return parse_exclusive_mask_target(target);
+    }
+
+    if sni.bytes().any(|byte| byte.is_ascii_uppercase()) {
+        let normalized_sni = sni.to_ascii_lowercase();
+        if let Some(target) = config
+            .censorship
+            .exclusive_mask_targets
+            .get(&normalized_sni)
+        {
+            return Some(MaskTcpTarget {
+                host: target.host.as_str(),
+                port: target.port,
+            });
+        }
+        if let Some(target) = config.censorship.exclusive_mask.get(&normalized_sni) {
+            return parse_exclusive_mask_target(target);
+        }
+    }
+
+    None
+}
+
+#[cfg(test)]
 fn mask_host_for_initial_data<'a>(config: &'a ProxyConfig, initial_data: &[u8]) -> &'a str {
+    mask_tcp_target_for_initial_data(config, initial_data).host
+}
+
+#[cfg(test)]
+fn mask_tcp_target_for_initial_data<'a>(
+    config: &'a ProxyConfig,
+    initial_data: &[u8],
+) -> MaskTcpTarget<'a> {
+    let sni = tls::extract_sni_from_client_hello(initial_data);
+    if let Some(target) = sni
+        .as_deref()
+        .and_then(|sni| exclusive_mask_target_for_sni(config, sni))
+    {
+        return target;
+    }
+
+    default_mask_tcp_target_for_initial_data(config, initial_data, sni.as_deref())
+}
+
+fn default_mask_tcp_target_for_initial_data<'a>(
+    config: &'a ProxyConfig,
+    initial_data: &[u8],
+    sni: Option<&str>,
+) -> MaskTcpTarget<'a> {
     let configured_mask_host = config
         .censorship
         .mask_host
@@ -466,13 +578,25 @@ fn mask_host_for_initial_data<'a>(config: &'a ProxyConfig, initial_data: &[u8])
         .unwrap_or(&config.censorship.tls_domain);
 
     if !configured_mask_host.eq_ignore_ascii_case(&config.censorship.tls_domain) {
-        return configured_mask_host;
+        return MaskTcpTarget {
+            host: configured_mask_host,
+            port: config.censorship.mask_port,
+        };
     }
 
-    tls::extract_sni_from_client_hello(initial_data)
-        .as_deref()
+    let extracted_sni = if sni.is_none() {
+        tls::extract_sni_from_client_hello(initial_data)
+    } else {
+        None
+    };
+    let host = sni
+        .or(extracted_sni.as_deref())
         .and_then(|sni| matching_tls_domain_for_sni(config, sni))
-        .unwrap_or(configured_mask_host)
+        .unwrap_or(configured_mask_host);
+    MaskTcpTarget {
+        host,
+        port: config.censorship.mask_port,
+    }
 }
 
 fn canonical_ip(ip: IpAddr) -> IpAddr {
@@ -770,9 +894,16 @@ pub async fn handle_bad_client<R, W>(
         return;
     }
 
+    let client_sni = tls::extract_sni_from_client_hello(initial_data);
+    let exclusive_tcp_target = client_sni
+        .as_deref()
+        .and_then(|sni| exclusive_mask_target_for_sni(config, sni));
+
     // Connect via Unix socket or TCP
     #[cfg(unix)]
-    if let Some(ref sock_path) = config.censorship.mask_unix_sock {
+    if exclusive_tcp_target.is_none()
+        && let Some(ref sock_path) = config.censorship.mask_unix_sock
+    {
         let outcome_started = Instant::now();
         let connect_started = Instant::now();
         debug!(
@@ -849,8 +980,11 @@ pub async fn handle_bad_client<R, W>(
         return;
     }
 
-    let mask_host = mask_host_for_initial_data(config, initial_data);
-    let mask_port = config.censorship.mask_port;
+    let mask_target = exclusive_tcp_target.unwrap_or_else(|| {
+        default_mask_tcp_target_for_initial_data(config, initial_data, client_sni.as_deref())
+    });
+    let mask_host = mask_target.host;
+    let mask_port = mask_target.port;
 
     // Fail closed when fallback points at our own listener endpoint.
     // Self-referential masking can create recursive proxy loops under

src/proxy/middle_relay/c2me.rs

@@ -0,0 +1,104 @@
+use super::*;
+
+pub(in crate::proxy::middle_relay) enum C2MeCommand {
+    Data {
+        payload: PooledBuffer,
+        flags: u32,
+        _permit: OwnedSemaphorePermit,
+    },
+    Close,
+}
+
+pub(super) fn should_yield_c2me_sender(sent_since_yield: usize, has_backlog: bool) -> bool {
+    has_backlog && sent_since_yield >= C2ME_SENDER_FAIRNESS_BUDGET
+}
+
+pub(super) fn c2me_payload_permits(payload_len: usize) -> u32 {
+    payload_len
+        .max(1)
+        .div_ceil(C2ME_QUEUED_BYTE_PERMIT_UNIT)
+        .min(u32::MAX as usize) as u32
+}
+
+pub(super) fn c2me_queued_permit_budget(channel_capacity: usize, frame_limit: usize) -> usize {
+    channel_capacity
+        .saturating_mul(C2ME_QUEUED_PERMITS_PER_SLOT)
+        .max(c2me_payload_permits(frame_limit) as usize)
+        .max(1)
+}
+
+pub(super) async fn acquire_c2me_payload_permit(
+    semaphore: &Arc<Semaphore>,
+    payload_len: usize,
+    send_timeout: Option<Duration>,
+    stats: &Stats,
+) -> Result<OwnedSemaphorePermit> {
+    let permits = c2me_payload_permits(payload_len);
+    let acquire = semaphore.clone().acquire_many_owned(permits);
+    match send_timeout {
+        Some(send_timeout) => match timeout(send_timeout, acquire).await {
+            Ok(Ok(permit)) => Ok(permit),
+            Ok(Err(_)) => Err(ProxyError::Proxy("ME sender byte budget closed".into())),
+            Err(_) => {
+                stats.increment_me_c2me_send_timeout_total();
+                Err(ProxyError::Proxy("ME sender byte budget timeout".into()))
+            }
+        },
+        None => acquire
+            .await
+            .map_err(|_| ProxyError::Proxy("ME sender byte budget closed".into())),
+    }
+}
+
+pub(super) async fn enqueue_c2me_command_in(
+    shared: &ProxySharedState,
+    tx: &mpsc::Sender<C2MeCommand>,
+    cmd: C2MeCommand,
+    send_timeout: Option<Duration>,
+    stats: &Stats,
+) -> std::result::Result<(), mpsc::error::SendError<C2MeCommand>> {
+    match tx.try_send(cmd) {
+        Ok(()) => Ok(()),
+        Err(mpsc::error::TrySendError::Closed(cmd)) => Err(mpsc::error::SendError(cmd)),
+        Err(mpsc::error::TrySendError::Full(cmd)) => {
+            stats.increment_me_c2me_send_full_total();
+            stats.increment_me_c2me_send_high_water_total();
+            note_relay_pressure_event_in(shared);
+            // Cooperative yield reduces burst catch-up when the per-conn queue is near saturation.
+            if tx.capacity() <= C2ME_SOFT_PRESSURE_MIN_FREE_SLOTS {
+                tokio::task::yield_now().await;
+            }
+            let reserve_result = match send_timeout {
+                Some(send_timeout) => match timeout(send_timeout, tx.reserve()).await {
+                    Ok(result) => result,
+                    Err(_) => {
+                        stats.increment_me_c2me_send_timeout_total();
+                        return Err(mpsc::error::SendError(cmd));
+                    }
+                },
+                None => tx.reserve().await,
+            };
+            match reserve_result {
+                Ok(permit) => {
+                    permit.send(cmd);
+                    Ok(())
+                }
+                Err(_) => {
+                    stats.increment_me_c2me_send_timeout_total();
+                    Err(mpsc::error::SendError(cmd))
+                }
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+pub(crate) async fn enqueue_c2me_command(
+    tx: &mpsc::Sender<C2MeCommand>,
+    cmd: C2MeCommand,
+    send_timeout: Option<Duration>,
+    stats: &Stats,
+) -> std::result::Result<(), mpsc::error::SendError<C2MeCommand>> {
+    let shared = ProxySharedState::new();
+    enqueue_c2me_command_in(shared.as_ref(), tx, cmd, send_timeout, stats).await
+}

src/proxy/middle_relay/d2c.rs

@@ -0,0 +1,456 @@
+use super::*;
+
+#[derive(Clone, Copy)]
+pub(super) struct MeD2cFlushPolicy {
+    pub(super) max_frames: usize,
+    pub(super) max_bytes: usize,
+    pub(super) max_delay: Duration,
+    pub(super) ack_flush_immediate: bool,
+    pub(super) quota_soft_overshoot_bytes: u64,
+    pub(super) frame_buf_shrink_threshold_bytes: usize,
+}
+
+impl MeD2cFlushPolicy {
+    pub(super) fn from_config(config: &ProxyConfig) -> Self {
+        Self {
+            max_frames: config
+                .general
+                .me_d2c_flush_batch_max_frames
+                .max(ME_D2C_FLUSH_BATCH_MAX_FRAMES_MIN),
+            max_bytes: config
+                .general
+                .me_d2c_flush_batch_max_bytes
+                .max(ME_D2C_FLUSH_BATCH_MAX_BYTES_MIN),
+            max_delay: Duration::from_micros(config.general.me_d2c_flush_batch_max_delay_us),
+            ack_flush_immediate: config.general.me_d2c_ack_flush_immediate,
+            quota_soft_overshoot_bytes: config.general.me_quota_soft_overshoot_bytes,
+            frame_buf_shrink_threshold_bytes: config
+                .general
+                .me_d2c_frame_buf_shrink_threshold_bytes
+                .max(4096),
+        }
+    }
+}
+
+pub(super) fn classify_me_d2c_flush_reason(
+    flush_immediately: bool,
+    batch_frames: usize,
+    max_frames: usize,
+    batch_bytes: usize,
+    max_bytes: usize,
+    max_delay_fired: bool,
+) -> MeD2cFlushReason {
+    if flush_immediately {
+        return MeD2cFlushReason::AckImmediate;
+    }
+    if batch_frames >= max_frames {
+        return MeD2cFlushReason::BatchFrames;
+    }
+    if batch_bytes >= max_bytes {
+        return MeD2cFlushReason::BatchBytes;
+    }
+    if max_delay_fired {
+        return MeD2cFlushReason::MaxDelay;
+    }
+    MeD2cFlushReason::QueueDrain
+}
+
+pub(super) fn observe_me_d2c_flush_event(
+    stats: &Stats,
+    reason: MeD2cFlushReason,
+    batch_frames: usize,
+    batch_bytes: usize,
+    flush_duration_us: Option<u64>,
+) {
+    stats.increment_me_d2c_flush_reason(reason);
+    if batch_frames > 0 || batch_bytes > 0 {
+        stats.increment_me_d2c_batches_total();
+        stats.add_me_d2c_batch_frames_total(batch_frames as u64);
+        stats.add_me_d2c_batch_bytes_total(batch_bytes as u64);
+        stats.observe_me_d2c_batch_frames(batch_frames as u64);
+        stats.observe_me_d2c_batch_bytes(batch_bytes as u64);
+    }
+    if let Some(duration_us) = flush_duration_us {
+        stats.observe_me_d2c_flush_duration_us(duration_us);
+    }
+}
+
+pub(super) enum MeWriterResponseOutcome {
+    Continue {
+        frames: usize,
+        bytes: usize,
+        flush_immediately: bool,
+    },
+    Close,
+}
+
+#[cfg(test)]
+pub(crate) async fn process_me_writer_response<W>(
+    response: MeResponse,
+    client_writer: &mut CryptoWriter<W>,
+    proto_tag: ProtoTag,
+    rng: &SecureRandom,
+    frame_buf: &mut Vec<u8>,
+    stats: &Stats,
+    user: &str,
+    quota_user_stats: Option<&UserStats>,
+    quota_limit: Option<u64>,
+    quota_soft_overshoot_bytes: u64,
+    bytes_me2c: &AtomicU64,
+    conn_id: u64,
+    ack_flush_immediate: bool,
+    batched: bool,
+) -> Result<MeWriterResponseOutcome>
+where
+    W: AsyncWrite + Unpin + Send + 'static,
+{
+    process_me_writer_response_with_traffic_lease(
+        response,
+        client_writer,
+        proto_tag,
+        rng,
+        frame_buf,
+        stats,
+        user,
+        quota_user_stats,
+        quota_limit,
+        quota_soft_overshoot_bytes,
+        None,
+        &CancellationToken::new(),
+        bytes_me2c,
+        conn_id,
+        ack_flush_immediate,
+        batched,
+    )
+    .await
+}
+
+pub(crate) async fn process_me_writer_response_with_traffic_lease<W>(
+    response: MeResponse,
+    client_writer: &mut CryptoWriter<W>,
+    proto_tag: ProtoTag,
+    rng: &SecureRandom,
+    frame_buf: &mut Vec<u8>,
+    stats: &Stats,
+    user: &str,
+    quota_user_stats: Option<&UserStats>,
+    quota_limit: Option<u64>,
+    quota_soft_overshoot_bytes: u64,
+    traffic_lease: Option<&Arc<TrafficLease>>,
+    cancel: &CancellationToken,
+    bytes_me2c: &AtomicU64,
+    conn_id: u64,
+    ack_flush_immediate: bool,
+    batched: bool,
+) -> Result<MeWriterResponseOutcome>
+where
+    W: AsyncWrite + Unpin + Send + 'static,
+{
+    match response {
+        MeResponse::Data { flags, data, .. } => {
+            if batched {
+                trace!(conn_id, bytes = data.len(), flags, "ME->C data (batched)");
+            } else {
+                trace!(conn_id, bytes = data.len(), flags, "ME->C data");
+            }
+            let data_len = data.len() as u64;
+            if let (Some(limit), Some(user_stats)) = (quota_limit, quota_user_stats) {
+                let soft_limit = quota_soft_cap(limit, quota_soft_overshoot_bytes);
+                match reserve_user_quota_with_yield(
+                    user_stats, data_len, soft_limit, stats, cancel, None,
+                )
+                .await
+                {
+                    Ok(_) => {}
+                    Err(MiddleQuotaReserveError::LimitExceeded) => {
+                        stats.increment_me_d2c_quota_reject_total(MeD2cQuotaRejectStage::PreWrite);
+                        return Err(ProxyError::DataQuotaExceeded {
+                            user: user.to_string(),
+                        });
+                    }
+                    Err(MiddleQuotaReserveError::Contended) => {
+                        return Err(ProxyError::Proxy(
+                            "ME D->C quota reservation contended".into(),
+                        ));
+                    }
+                    Err(MiddleQuotaReserveError::Cancelled) => {
+                        return Err(ProxyError::Proxy(
+                            "ME D->C quota reservation cancelled".into(),
+                        ));
+                    }
+                    Err(MiddleQuotaReserveError::DeadlineExceeded) => {
+                        return Err(ProxyError::Proxy(
+                            "ME D->C quota reservation deadline exceeded".into(),
+                        ));
+                    }
+                }
+            }
+            wait_for_traffic_budget_or_cancel(
+                traffic_lease,
+                RateDirection::Down,
+                data_len,
+                cancel,
+                stats,
+                None,
+            )
+            .await?;
+
+            let write_mode = match write_client_payload(
+                client_writer,
+                proto_tag,
+                flags,
+                &data,
+                rng,
+                frame_buf,
+                cancel,
+            )
+            .await
+            {
+                Ok(mode) => mode,
+                Err(err) => {
+                    if quota_limit.is_some() {
+                        stats.add_quota_write_fail_bytes_total(data_len);
+                        stats.increment_quota_write_fail_events_total();
+                    }
+                    return Err(err);
+                }
+            };
+
+            bytes_me2c.fetch_add(data_len, Ordering::Relaxed);
+            if let Some(user_stats) = quota_user_stats {
+                stats.add_user_octets_to_handle(user_stats, data_len);
+            } else {
+                stats.add_user_octets_to(user, data_len);
+            }
+            stats.increment_me_d2c_data_frames_total();
+            stats.add_me_d2c_payload_bytes_total(data_len);
+            stats.increment_me_d2c_write_mode(write_mode);
+
+            Ok(MeWriterResponseOutcome::Continue {
+                frames: 1,
+                bytes: data.len(),
+                flush_immediately: false,
+            })
+        }
+        MeResponse::Ack(confirm) => {
+            if batched {
+                trace!(conn_id, confirm, "ME->C quickack (batched)");
+            } else {
+                trace!(conn_id, confirm, "ME->C quickack");
+            }
+            wait_for_traffic_budget_or_cancel(
+                traffic_lease,
+                RateDirection::Down,
+                4,
+                cancel,
+                stats,
+                None,
+            )
+            .await?;
+            write_client_ack(client_writer, proto_tag, confirm, cancel).await?;
+            stats.increment_me_d2c_ack_frames_total();
+
+            Ok(MeWriterResponseOutcome::Continue {
+                frames: 1,
+                bytes: 4,
+                flush_immediately: ack_flush_immediate,
+            })
+        }
+        MeResponse::Close => {
+            if batched {
+                debug!(conn_id, "ME sent close (batched)");
+            } else {
+                debug!(conn_id, "ME sent close");
+            }
+            Ok(MeWriterResponseOutcome::Close)
+        }
+    }
+}
+
+/// Computes the intermediate/secure wire length while rejecting lossy casts.
+pub(in crate::proxy::middle_relay) fn compute_intermediate_secure_wire_len(
+    data_len: usize,
+    padding_len: usize,
+    quickack: bool,
+) -> Result<(u32, usize)> {
+    let wire_len = data_len
+        .checked_add(padding_len)
+        .ok_or_else(|| ProxyError::Proxy("Frame length overflow".into()))?;
+    if wire_len > 0x7fff_ffffusize {
+        return Err(ProxyError::Proxy(format!(
+            "Intermediate/Secure frame too large: {wire_len}"
+        )));
+    }
+
+    let total = 4usize
+        .checked_add(wire_len)
+        .ok_or_else(|| ProxyError::Proxy("Frame buffer size overflow".into()))?;
+    let mut len_val = u32::try_from(wire_len)
+        .map_err(|_| ProxyError::Proxy("Frame length conversion overflow".into()))?;
+    if quickack {
+        len_val |= 0x8000_0000;
+    }
+    Ok((len_val, total))
+}
+
+pub(super) async fn write_client_payload<W>(
+    client_writer: &mut CryptoWriter<W>,
+    proto_tag: ProtoTag,
+    flags: u32,
+    data: &[u8],
+    rng: &SecureRandom,
+    frame_buf: &mut Vec<u8>,
+    cancel: &CancellationToken,
+) -> Result<MeD2cWriteMode>
+where
+    W: AsyncWrite + Unpin + Send + 'static,
+{
+    let quickack = (flags & RPC_FLAG_QUICKACK) != 0;
+
+    let write_mode = match proto_tag {
+        ProtoTag::Abridged => {
+            if !data.len().is_multiple_of(4) {
+                return Err(ProxyError::Proxy(format!(
+                    "Abridged payload must be 4-byte aligned, got {}",
+                    data.len()
+                )));
+            }
+
+            let len_words = data.len() / 4;
+            if len_words < 0x7f {
+                let mut first = len_words as u8;
+                if quickack {
+                    first |= 0x80;
+                }
+                let wire_len = 1usize.saturating_add(data.len());
+                if wire_len <= ME_D2C_SINGLE_WRITE_COALESCE_MAX_BYTES {
+                    frame_buf.clear();
+                    frame_buf.reserve(wire_len);
+                    frame_buf.push(first);
+                    frame_buf.extend_from_slice(data);
+                    write_all_client_or_cancel(client_writer, frame_buf.as_slice(), cancel).await?;
+                    MeD2cWriteMode::Coalesced
+                } else {
+                    let header = [first];
+                    write_all_client_or_cancel(client_writer, &header, cancel).await?;
+                    write_all_client_or_cancel(client_writer, data, cancel).await?;
+                    MeD2cWriteMode::Split
+                }
+            } else if len_words < (1 << 24) {
+                let mut first = 0x7fu8;
+                if quickack {
+                    first |= 0x80;
+                }
+                let lw = (len_words as u32).to_le_bytes();
+                let wire_len = 4usize.saturating_add(data.len());
+                if wire_len <= ME_D2C_SINGLE_WRITE_COALESCE_MAX_BYTES {
+                    frame_buf.clear();
+                    frame_buf.reserve(wire_len);
+                    frame_buf.extend_from_slice(&[first, lw[0], lw[1], lw[2]]);
+                    frame_buf.extend_from_slice(data);
+                    write_all_client_or_cancel(client_writer, frame_buf.as_slice(), cancel).await?;
+                    MeD2cWriteMode::Coalesced
+                } else {
+                    let header = [first, lw[0], lw[1], lw[2]];
+                    write_all_client_or_cancel(client_writer, &header, cancel).await?;
+                    write_all_client_or_cancel(client_writer, data, cancel).await?;
+                    MeD2cWriteMode::Split
+                }
+            } else {
+                return Err(ProxyError::Proxy(format!(
+                    "Abridged frame too large: {}",
+                    data.len()
+                )));
+            }
+        }
+        ProtoTag::Intermediate | ProtoTag::Secure => {
+            let padding_len = if proto_tag == ProtoTag::Secure {
+                if !is_valid_secure_payload_len(data.len()) {
+                    return Err(ProxyError::Proxy(format!(
+                        "Secure payload must be 4-byte aligned, got {}",
+                        data.len()
+                    )));
+                }
+                secure_padding_len(data.len(), rng)
+            } else {
+                0
+            };
+
+            let (len_val, total) =
+                compute_intermediate_secure_wire_len(data.len(), padding_len, quickack)?;
+            if total <= ME_D2C_SINGLE_WRITE_COALESCE_MAX_BYTES {
+                frame_buf.clear();
+                frame_buf.reserve(total);
+                frame_buf.extend_from_slice(&len_val.to_le_bytes());
+                frame_buf.extend_from_slice(data);
+                if padding_len > 0 {
+                    let start = frame_buf.len();
+                    frame_buf.resize(start + padding_len, 0);
+                    rng.fill(&mut frame_buf[start..]);
+                }
+                write_all_client_or_cancel(client_writer, frame_buf.as_slice(), cancel).await?;
+                MeD2cWriteMode::Coalesced
+            } else {
+                let header = len_val.to_le_bytes();
+                write_all_client_or_cancel(client_writer, &header, cancel).await?;
+                write_all_client_or_cancel(client_writer, data, cancel).await?;
+                if padding_len > 0 {
+                    frame_buf.clear();
+                    if frame_buf.capacity() < padding_len {
+                        frame_buf.reserve(padding_len);
+                    }
+                    frame_buf.resize(padding_len, 0);
+                    rng.fill(frame_buf.as_mut_slice());
+                    write_all_client_or_cancel(client_writer, frame_buf.as_slice(), cancel).await?;
+                }
+                MeD2cWriteMode::Split
+            }
+        }
+    };
+
+    Ok(write_mode)
+}
+
+pub(super) async fn write_client_ack<W>(
+    client_writer: &mut CryptoWriter<W>,
+    proto_tag: ProtoTag,
+    confirm: u32,
+    cancel: &CancellationToken,
+) -> Result<()>
+where
+    W: AsyncWrite + Unpin + Send + 'static,
+{
+    let bytes = if proto_tag == ProtoTag::Abridged {
+        confirm.to_be_bytes()
+    } else {
+        confirm.to_le_bytes()
+    };
+    write_all_client_or_cancel(client_writer, &bytes, cancel).await
+}
+
+pub(super) async fn write_all_client_or_cancel<W>(
+    client_writer: &mut CryptoWriter<W>,
+    bytes: &[u8],
+    cancel: &CancellationToken,
+) -> Result<()>
+where
+    W: AsyncWrite + Unpin + Send + 'static,
+{
+    tokio::select! {
+        result = client_writer.write_all(bytes) => result.map_err(ProxyError::Io),
+        _ = cancel.cancelled() => Err(ProxyError::MiddleClientWriterCancelled),
+    }
+}
+
+pub(super) async fn flush_client_or_cancel<W>(
+    client_writer: &mut CryptoWriter<W>,
+    cancel: &CancellationToken,
+) -> Result<()>
+where
+    W: AsyncWrite + Unpin + Send + 'static,
+{
+    tokio::select! {
+        result = client_writer.flush() => result.map_err(ProxyError::Io),
+        _ = cancel.cancelled() => Err(ProxyError::MiddleClientWriterCancelled),
+    }
+}

src/proxy/middle_relay/desync.rs

@@ -0,0 +1,406 @@
+use super::*;
+
+#[derive(Default)]
+pub(crate) struct DesyncDedupRotationState {
+    current_started_at: Option<Instant>,
+}
+
+pub(in crate::proxy::middle_relay) struct RelayForensicsState {
+    pub(in crate::proxy::middle_relay) trace_id: u64,
+    pub(in crate::proxy::middle_relay) conn_id: u64,
+    pub(in crate::proxy::middle_relay) user: String,
+    pub(in crate::proxy::middle_relay) peer: SocketAddr,
+    pub(in crate::proxy::middle_relay) peer_hash: u64,
+    pub(in crate::proxy::middle_relay) started_at: Instant,
+    pub(in crate::proxy::middle_relay) bytes_c2me: u64,
+    pub(in crate::proxy::middle_relay) bytes_me2c: Arc<AtomicU64>,
+    pub(in crate::proxy::middle_relay) desync_all_full: bool,
+}
+
+#[cfg(test)]
+pub(crate) fn hash_value<T: Hash>(value: &T) -> u64 {
+    let mut hasher = DefaultHasher::new();
+    value.hash(&mut hasher);
+    hasher.finish()
+}
+
+fn hash_value_in<T: Hash>(shared: &ProxySharedState, value: &T) -> u64 {
+    shared.middle_relay.desync_hasher.hash_one(value)
+}
+
+#[cfg(test)]
+pub(crate) fn hash_ip(ip: IpAddr) -> u64 {
+    hash_value(&ip)
+}
+
+pub(super) fn hash_ip_in(shared: &ProxySharedState, ip: IpAddr) -> u64 {
+    hash_value_in(shared, &ip)
+}
+
+fn should_emit_full_desync_in(
+    shared: &ProxySharedState,
+    key: u64,
+    all_full: bool,
+    now: Instant,
+) -> bool {
+    if all_full {
+        return true;
+    }
+
+    let dedup_current = &shared.middle_relay.desync_dedup;
+    let dedup_previous = &shared.middle_relay.desync_dedup_previous;
+    let rotation_state = &shared.middle_relay.desync_dedup_rotation_state;
+
+    let mut state = match rotation_state.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => {
+            let mut guard = poisoned.into_inner();
+            *guard = DesyncDedupRotationState::default();
+            rotation_state.clear_poison();
+            guard
+        }
+    };
+
+    let rotate_now = match state.current_started_at {
+        Some(current_started_at) => match now.checked_duration_since(current_started_at) {
+            Some(elapsed) => elapsed >= DESYNC_DEDUP_WINDOW,
+            None => true,
+        },
+        None => true,
+    };
+    if rotate_now {
+        dedup_previous.clear();
+        for entry in dedup_current.iter() {
+            dedup_previous.insert(*entry.key(), *entry.value());
+        }
+        dedup_current.clear();
+        state.current_started_at = Some(now);
+    }
+
+    if let Some(seen_at) = dedup_current.get(&key).map(|entry| *entry.value()) {
+        let within_window = match now.checked_duration_since(seen_at) {
+            Some(elapsed) => elapsed < DESYNC_DEDUP_WINDOW,
+            None => true,
+        };
+        if within_window {
+            return false;
+        }
+        dedup_current.insert(key, now);
+        return true;
+    }
+
+    if let Some(seen_at) = dedup_previous.get(&key).map(|entry| *entry.value()) {
+        let within_window = match now.checked_duration_since(seen_at) {
+            Some(elapsed) => elapsed < DESYNC_DEDUP_WINDOW,
+            None => true,
+        };
+        if within_window {
+            dedup_current.insert(key, seen_at);
+            return false;
+        }
+        dedup_previous.remove(&key);
+    }
+
+    if dedup_current.len() >= DESYNC_DEDUP_MAX_ENTRIES {
+        dedup_previous.clear();
+        for entry in dedup_current.iter() {
+            dedup_previous.insert(*entry.key(), *entry.value());
+        }
+        dedup_current.clear();
+        state.current_started_at = Some(now);
+        dedup_current.insert(key, now);
+        should_emit_full_desync_full_cache_in(shared, now)
+    } else {
+        dedup_current.insert(key, now);
+        true
+    }
+}
+
+fn should_emit_full_desync_full_cache_in(shared: &ProxySharedState, now: Instant) -> bool {
+    let gate = &shared.middle_relay.desync_full_cache_last_emit_at;
+    let Ok(mut last_emit_at) = gate.lock() else {
+        return false;
+    };
+
+    match *last_emit_at {
+        None => {
+            *last_emit_at = Some(now);
+            true
+        }
+        Some(last) => {
+            let Some(elapsed) = now.checked_duration_since(last) else {
+                *last_emit_at = Some(now);
+                return true;
+            };
+            if elapsed >= DESYNC_FULL_CACHE_EMIT_MIN_INTERVAL {
+                *last_emit_at = Some(now);
+                true
+            } else {
+                false
+            }
+        }
+    }
+}
+
+pub(crate) fn desync_forensics_len_bytes(len: usize) -> ([u8; 4], bool) {
+    match u32::try_from(len) {
+        Ok(value) => (value.to_le_bytes(), false),
+        Err(_) => (u32::MAX.to_le_bytes(), true),
+    }
+}
+
+pub(super) fn report_desync_frame_too_large_in(
+    shared: &ProxySharedState,
+    state: &RelayForensicsState,
+    proto_tag: ProtoTag,
+    frame_counter: u64,
+    max_frame: usize,
+    len: usize,
+    raw_len_bytes: Option<[u8; 4]>,
+    stats: &Stats,
+) -> ProxyError {
+    let (fallback_len_buf, len_buf_truncated) = desync_forensics_len_bytes(len);
+    let len_buf = raw_len_bytes.unwrap_or(fallback_len_buf);
+    let looks_like_tls = raw_len_bytes
+        .map(|b| b[0] == 0x16 && b[1] == 0x03)
+        .unwrap_or(false);
+    let looks_like_http = raw_len_bytes
+        .map(|b| matches!(b[0], b'G' | b'P' | b'H' | b'C' | b'D'))
+        .unwrap_or(false);
+    let now = Instant::now();
+    let dedup_key = hash_value_in(
+        shared,
+        &(
+            state.user.as_str(),
+            state.peer_hash,
+            proto_tag,
+            DESYNC_ERROR_CLASS,
+        ),
+    );
+    let emit_full = should_emit_full_desync_in(shared, dedup_key, state.desync_all_full, now);
+    let duration_ms = state.started_at.elapsed().as_millis() as u64;
+    let bytes_me2c = state.bytes_me2c.load(Ordering::Relaxed);
+
+    stats.increment_desync_total();
+    stats.increment_relay_protocol_desync_close_total();
+    stats.observe_desync_frames_ok(frame_counter);
+    if emit_full {
+        stats.increment_desync_full_logged();
+        warn!(
+            trace_id = format_args!("0x{:016x}", state.trace_id),
+            conn_id = state.conn_id,
+            user = %state.user,
+            peer_hash = format_args!("0x{:016x}", state.peer_hash),
+            proto = ?proto_tag,
+            mode = "middle_proxy",
+            is_tls = true,
+            duration_ms,
+            bytes_c2me = state.bytes_c2me,
+            bytes_me2c,
+            raw_len = len,
+            raw_len_hex = format_args!("0x{:08x}", len),
+            raw_len_bytes_truncated = len_buf_truncated,
+            raw_bytes = format_args!(
+                "{:02x} {:02x} {:02x} {:02x}",
+                len_buf[0], len_buf[1], len_buf[2], len_buf[3]
+            ),
+            max_frame,
+            tls_like = looks_like_tls,
+            http_like = looks_like_http,
+            frames_ok = frame_counter,
+            dedup_window_secs = DESYNC_DEDUP_WINDOW.as_secs(),
+            desync_all_full = state.desync_all_full,
+            full_reason = if state.desync_all_full { "desync_all_full" } else { "first_in_dedup_window" },
+            error_class = DESYNC_ERROR_CLASS,
+            "Frame too large — crypto desync forensics"
+        );
+        debug!(
+            trace_id = format_args!("0x{:016x}", state.trace_id),
+            conn_id = state.conn_id,
+            user = %state.user,
+            peer = %state.peer,
+            "Frame too large forensic peer detail"
+        );
+    } else {
+        stats.increment_desync_suppressed();
+        debug!(
+            trace_id = format_args!("0x{:016x}", state.trace_id),
+            conn_id = state.conn_id,
+            user = %state.user,
+            peer_hash = format_args!("0x{:016x}", state.peer_hash),
+            proto = ?proto_tag,
+            duration_ms,
+            bytes_c2me = state.bytes_c2me,
+            bytes_me2c,
+            raw_len = len,
+            frames_ok = frame_counter,
+            dedup_window_secs = DESYNC_DEDUP_WINDOW.as_secs(),
+            error_class = DESYNC_ERROR_CLASS,
+            "Frame too large — crypto desync forensic suppressed"
+        );
+    }
+
+    ProxyError::Proxy(format!(
+        "Frame too large: {len} (max {max_frame}), frames_ok={frame_counter}, conn_id={}, trace_id=0x{:016x}",
+        state.conn_id, state.trace_id
+    ))
+}
+
+#[cfg(test)]
+pub(crate) fn report_desync_frame_too_large(
+    state: &RelayForensicsState,
+    proto_tag: ProtoTag,
+    frame_counter: u64,
+    max_frame: usize,
+    len: usize,
+    raw_len_bytes: Option<[u8; 4]>,
+    stats: &Stats,
+) -> ProxyError {
+    let shared = ProxySharedState::new();
+    report_desync_frame_too_large_in(
+        shared.as_ref(),
+        state,
+        proto_tag,
+        frame_counter,
+        max_frame,
+        len,
+        raw_len_bytes,
+        stats,
+    )
+}
+
+#[cfg(test)]
+pub(crate) fn should_emit_full_desync_for_testing(
+    shared: &ProxySharedState,
+    key: u64,
+    all_full: bool,
+    now: Instant,
+) -> bool {
+    if all_full {
+        return true;
+    }
+
+    let dedup_current = &shared.middle_relay.desync_dedup;
+    let dedup_previous = &shared.middle_relay.desync_dedup_previous;
+
+    let Ok(mut state) = shared.middle_relay.desync_dedup_rotation_state.lock() else {
+        return false;
+    };
+
+    let rotate_now = match state.current_started_at {
+        Some(current_started_at) => match now.checked_duration_since(current_started_at) {
+            Some(elapsed) => elapsed >= DESYNC_DEDUP_WINDOW,
+            None => true,
+        },
+        None => true,
+    };
+    if rotate_now {
+        dedup_previous.clear();
+        for entry in dedup_current.iter() {
+            dedup_previous.insert(*entry.key(), *entry.value());
+        }
+        dedup_current.clear();
+        state.current_started_at = Some(now);
+    }
+
+    if let Some(seen_at) = dedup_current.get(&key).map(|entry| *entry.value()) {
+        let within_window = match now.checked_duration_since(seen_at) {
+            Some(elapsed) => elapsed < DESYNC_DEDUP_WINDOW,
+            None => true,
+        };
+        if within_window {
+            return false;
+        }
+        dedup_current.insert(key, now);
+        return true;
+    }
+
+    if let Some(seen_at) = dedup_previous.get(&key).map(|entry| *entry.value()) {
+        let within_window = match now.checked_duration_since(seen_at) {
+            Some(elapsed) => elapsed < DESYNC_DEDUP_WINDOW,
+            None => true,
+        };
+        if within_window {
+            dedup_current.insert(key, seen_at);
+            return false;
+        }
+        dedup_previous.remove(&key);
+    }
+
+    if dedup_current.len() >= DESYNC_DEDUP_MAX_ENTRIES {
+        dedup_previous.clear();
+        for entry in dedup_current.iter() {
+            dedup_previous.insert(*entry.key(), *entry.value());
+        }
+        dedup_current.clear();
+        state.current_started_at = Some(now);
+        dedup_current.insert(key, now);
+        let Ok(mut last_emit_at) = shared.middle_relay.desync_full_cache_last_emit_at.lock() else {
+            return false;
+        };
+        return match *last_emit_at {
+            None => {
+                *last_emit_at = Some(now);
+                true
+            }
+            Some(last) => {
+                let Some(elapsed) = now.checked_duration_since(last) else {
+                    *last_emit_at = Some(now);
+                    return true;
+                };
+                if elapsed >= DESYNC_FULL_CACHE_EMIT_MIN_INTERVAL {
+                    *last_emit_at = Some(now);
+                    true
+                } else {
+                    false
+                }
+            }
+        };
+    }
+
+    dedup_current.insert(key, now);
+    true
+}
+
+#[cfg(test)]
+pub(crate) fn clear_desync_dedup_for_testing_in_shared(shared: &ProxySharedState) {
+    shared.middle_relay.desync_dedup.clear();
+    shared.middle_relay.desync_dedup_previous.clear();
+    if let Ok(mut rotation_state) = shared.middle_relay.desync_dedup_rotation_state.lock() {
+        *rotation_state = DesyncDedupRotationState::default();
+    }
+    if let Ok(mut last_emit_at) = shared.middle_relay.desync_full_cache_last_emit_at.lock() {
+        *last_emit_at = None;
+    }
+}
+
+#[cfg(test)]
+pub(crate) fn desync_dedup_len_for_testing(shared: &ProxySharedState) -> usize {
+    shared.middle_relay.desync_dedup.len()
+}
+
+#[cfg(test)]
+pub(crate) fn desync_dedup_insert_for_testing(shared: &ProxySharedState, key: u64, at: Instant) {
+    shared.middle_relay.desync_dedup.insert(key, at);
+}
+
+#[cfg(test)]
+pub(crate) fn desync_dedup_get_for_testing(shared: &ProxySharedState, key: u64) -> Option<Instant> {
+    shared
+        .middle_relay
+        .desync_dedup
+        .get(&key)
+        .map(|entry| *entry.value())
+}
+
+#[cfg(test)]
+pub(crate) fn desync_dedup_keys_for_testing(
+    shared: &ProxySharedState,
+) -> std::collections::HashSet<u64> {
+    shared
+        .middle_relay
+        .desync_dedup
+        .iter()
+        .map(|entry| *entry.key())
+        .collect()
+}

src/proxy/middle_relay/idle.rs

@@ -0,0 +1,341 @@
+use super::*;
+
+mod read;
+
+pub(crate) use self::read::read_client_payload_with_idle_policy_in;
+#[cfg(test)]
+pub(crate) use self::read::{
+    read_client_payload, read_client_payload_legacy, read_client_payload_with_idle_policy,
+};
+
+#[derive(Default)]
+pub(crate) struct RelayIdleCandidateRegistry {
+    pub(in crate::proxy::middle_relay) by_conn_id: HashMap<u64, RelayIdleCandidateMeta>,
+    pub(in crate::proxy::middle_relay) ordered: BTreeSet<(u64, u64)>,
+    pressure_event_seq: u64,
+    pressure_consumed_seq: u64,
+}
+
+/// Queue metadata used to preserve FIFO ordering for idle relay eviction.
+#[derive(Clone, Copy)]
+pub(in crate::proxy::middle_relay) struct RelayIdleCandidateMeta {
+    pub(in crate::proxy::middle_relay) mark_order_seq: u64,
+    pub(in crate::proxy::middle_relay) mark_pressure_seq: u64,
+}
+
+pub(super) fn relay_idle_candidate_registry_lock_in(
+    shared: &ProxySharedState,
+) -> std::sync::MutexGuard<'_, RelayIdleCandidateRegistry> {
+    let registry = &shared.middle_relay.relay_idle_registry;
+    match registry.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => {
+            let mut guard = poisoned.into_inner();
+            *guard = RelayIdleCandidateRegistry::default();
+            registry.clear_poison();
+            guard
+        }
+    }
+}
+
+pub(super) fn mark_relay_idle_candidate_in(shared: &ProxySharedState, conn_id: u64) -> bool {
+    let mut guard = relay_idle_candidate_registry_lock_in(shared);
+
+    if guard.by_conn_id.contains_key(&conn_id) {
+        return false;
+    }
+
+    let mark_order_seq = shared
+        .middle_relay
+        .relay_idle_mark_seq
+        .fetch_add(1, Ordering::Relaxed)
+        .saturating_add(1);
+    let meta = RelayIdleCandidateMeta {
+        mark_order_seq,
+        mark_pressure_seq: guard.pressure_event_seq,
+    };
+    guard.by_conn_id.insert(conn_id, meta);
+    guard.ordered.insert((meta.mark_order_seq, conn_id));
+    true
+}
+
+pub(super) fn clear_relay_idle_candidate_in(shared: &ProxySharedState, conn_id: u64) {
+    let mut guard = relay_idle_candidate_registry_lock_in(shared);
+
+    if let Some(meta) = guard.by_conn_id.remove(&conn_id) {
+        guard.ordered.remove(&(meta.mark_order_seq, conn_id));
+    }
+}
+
+pub(super) fn note_relay_pressure_event_in(shared: &ProxySharedState) {
+    let mut guard = relay_idle_candidate_registry_lock_in(shared);
+    guard.pressure_event_seq = guard.pressure_event_seq.wrapping_add(1);
+}
+
+pub(crate) fn note_global_relay_pressure(shared: &ProxySharedState) {
+    note_relay_pressure_event_in(shared);
+}
+
+pub(super) fn relay_pressure_event_seq_in(shared: &ProxySharedState) -> u64 {
+    let guard = relay_idle_candidate_registry_lock_in(shared);
+    guard.pressure_event_seq
+}
+
+pub(super) fn maybe_evict_idle_candidate_on_pressure_in(
+    shared: &ProxySharedState,
+    conn_id: u64,
+    seen_pressure_seq: &mut u64,
+    stats: &Stats,
+) -> bool {
+    let mut guard = relay_idle_candidate_registry_lock_in(shared);
+
+    let latest_pressure_seq = guard.pressure_event_seq;
+    if latest_pressure_seq == *seen_pressure_seq {
+        return false;
+    }
+    *seen_pressure_seq = latest_pressure_seq;
+
+    if latest_pressure_seq == guard.pressure_consumed_seq {
+        return false;
+    }
+
+    if guard.ordered.is_empty() {
+        guard.pressure_consumed_seq = latest_pressure_seq;
+        return false;
+    }
+
+    let oldest = guard
+        .ordered
+        .iter()
+        .next()
+        .map(|(_, candidate_conn_id)| *candidate_conn_id);
+    if oldest != Some(conn_id) {
+        return false;
+    }
+
+    let Some(candidate_meta) = guard.by_conn_id.get(&conn_id).copied() else {
+        return false;
+    };
+
+    if latest_pressure_seq == candidate_meta.mark_pressure_seq {
+        return false;
+    }
+
+    if let Some(meta) = guard.by_conn_id.remove(&conn_id) {
+        guard.ordered.remove(&(meta.mark_order_seq, conn_id));
+    }
+    guard.pressure_consumed_seq = latest_pressure_seq;
+    stats.increment_relay_pressure_evict_total();
+    true
+}
+
+#[derive(Clone, Copy)]
+pub(in crate::proxy::middle_relay) struct RelayClientIdlePolicy {
+    pub(in crate::proxy::middle_relay) enabled: bool,
+    pub(in crate::proxy::middle_relay) soft_idle: Duration,
+    pub(in crate::proxy::middle_relay) hard_idle: Duration,
+    pub(in crate::proxy::middle_relay) grace_after_downstream_activity: Duration,
+    pub(in crate::proxy::middle_relay) legacy_frame_read_timeout: Duration,
+}
+
+impl RelayClientIdlePolicy {
+    pub(super) fn from_config(config: &ProxyConfig) -> Self {
+        let frame_read_timeout =
+            Duration::from_secs(config.timeouts.relay_client_idle_hard_secs.max(1));
+        if !config.timeouts.relay_idle_policy_v2_enabled {
+            return Self::disabled(frame_read_timeout);
+        }
+
+        let soft_idle = Duration::from_secs(config.timeouts.relay_client_idle_soft_secs.max(1));
+        let hard_idle = Duration::from_secs(config.timeouts.relay_client_idle_hard_secs.max(1));
+        let grace_after_downstream_activity = Duration::from_secs(
+            config
+                .timeouts
+                .relay_idle_grace_after_downstream_activity_secs,
+        );
+
+        Self {
+            enabled: true,
+            soft_idle,
+            hard_idle,
+            grace_after_downstream_activity,
+            legacy_frame_read_timeout: frame_read_timeout,
+        }
+    }
+
+    pub(in crate::proxy::middle_relay) fn disabled(frame_read_timeout: Duration) -> Self {
+        Self {
+            enabled: false,
+            soft_idle: frame_read_timeout,
+            hard_idle: frame_read_timeout,
+            grace_after_downstream_activity: Duration::ZERO,
+            legacy_frame_read_timeout: frame_read_timeout,
+        }
+    }
+
+    pub(super) fn apply_pressure_caps(&mut self, profile: ConntrackPressureProfile) {
+        let pressure_soft_idle_cap = Duration::from_secs(profile.middle_soft_idle_cap_secs());
+        let pressure_hard_idle_cap = Duration::from_secs(profile.middle_hard_idle_cap_secs());
+
+        self.soft_idle = self.soft_idle.min(pressure_soft_idle_cap);
+        self.hard_idle = self.hard_idle.min(pressure_hard_idle_cap);
+        if self.soft_idle > self.hard_idle {
+            self.soft_idle = self.hard_idle;
+        }
+        self.legacy_frame_read_timeout = self.legacy_frame_read_timeout.min(pressure_hard_idle_cap);
+        if self.grace_after_downstream_activity > self.hard_idle {
+            self.grace_after_downstream_activity = self.hard_idle;
+        }
+    }
+}
+
+#[derive(Clone, Copy)]
+pub(in crate::proxy::middle_relay) struct RelayClientIdleState {
+    pub(in crate::proxy::middle_relay) last_client_frame_at: Instant,
+    pub(in crate::proxy::middle_relay) soft_idle_marked: bool,
+    pub(in crate::proxy::middle_relay) tiny_frame_debt: u32,
+}
+
+impl RelayClientIdleState {
+    pub(super) fn new(now: Instant) -> Self {
+        Self {
+            last_client_frame_at: now,
+            soft_idle_marked: false,
+            tiny_frame_debt: 0,
+        }
+    }
+
+    pub(super) fn on_client_frame(&mut self, now: Instant) {
+        self.last_client_frame_at = now;
+        self.soft_idle_marked = false;
+    }
+
+    pub(super) fn on_client_tiny_frame(&mut self, now: Instant) {
+        self.last_client_frame_at = now;
+    }
+}
+
+#[cfg(test)]
+pub(crate) fn mark_relay_idle_candidate_for_testing(
+    shared: &ProxySharedState,
+    conn_id: u64,
+) -> bool {
+    let registry = &shared.middle_relay.relay_idle_registry;
+    let mut guard = match registry.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => {
+            let mut guard = poisoned.into_inner();
+            *guard = RelayIdleCandidateRegistry::default();
+            registry.clear_poison();
+            guard
+        }
+    };
+
+    if guard.by_conn_id.contains_key(&conn_id) {
+        return false;
+    }
+
+    let mark_order_seq = shared
+        .middle_relay
+        .relay_idle_mark_seq
+        .fetch_add(1, Ordering::Relaxed);
+    let mark_pressure_seq = guard.pressure_event_seq;
+    let meta = RelayIdleCandidateMeta {
+        mark_order_seq,
+        mark_pressure_seq,
+    };
+    guard.by_conn_id.insert(conn_id, meta);
+    guard.ordered.insert((mark_order_seq, conn_id));
+    true
+}
+
+#[cfg(test)]
+pub(crate) fn oldest_relay_idle_candidate_for_testing(shared: &ProxySharedState) -> Option<u64> {
+    let registry = &shared.middle_relay.relay_idle_registry;
+    let guard = match registry.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => {
+            let mut guard = poisoned.into_inner();
+            *guard = RelayIdleCandidateRegistry::default();
+            registry.clear_poison();
+            guard
+        }
+    };
+    guard.ordered.iter().next().map(|(_, conn_id)| *conn_id)
+}
+
+#[cfg(test)]
+pub(crate) fn clear_relay_idle_candidate_for_testing(shared: &ProxySharedState, conn_id: u64) {
+    let registry = &shared.middle_relay.relay_idle_registry;
+    let mut guard = match registry.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => {
+            let mut guard = poisoned.into_inner();
+            *guard = RelayIdleCandidateRegistry::default();
+            registry.clear_poison();
+            guard
+        }
+    };
+    if let Some(meta) = guard.by_conn_id.remove(&conn_id) {
+        guard.ordered.remove(&(meta.mark_order_seq, conn_id));
+    }
+}
+
+#[cfg(test)]
+pub(crate) fn clear_relay_idle_pressure_state_for_testing_in_shared(shared: &ProxySharedState) {
+    if let Ok(mut guard) = shared.middle_relay.relay_idle_registry.lock() {
+        *guard = RelayIdleCandidateRegistry::default();
+    }
+    shared
+        .middle_relay
+        .relay_idle_mark_seq
+        .store(0, Ordering::Relaxed);
+}
+
+#[cfg(test)]
+pub(crate) fn note_relay_pressure_event_for_testing(shared: &ProxySharedState) {
+    note_relay_pressure_event_in(shared);
+}
+
+#[cfg(test)]
+pub(crate) fn relay_pressure_event_seq_for_testing(shared: &ProxySharedState) -> u64 {
+    relay_pressure_event_seq_in(shared)
+}
+
+#[cfg(test)]
+pub(crate) fn relay_idle_mark_seq_for_testing(shared: &ProxySharedState) -> u64 {
+    shared
+        .middle_relay
+        .relay_idle_mark_seq
+        .load(Ordering::Relaxed)
+}
+
+#[cfg(test)]
+pub(crate) fn maybe_evict_idle_candidate_on_pressure_for_testing(
+    shared: &ProxySharedState,
+    conn_id: u64,
+    seen_pressure_seq: &mut u64,
+    stats: &Stats,
+) -> bool {
+    maybe_evict_idle_candidate_on_pressure_in(shared, conn_id, seen_pressure_seq, stats)
+}
+
+#[cfg(test)]
+pub(crate) fn set_relay_pressure_state_for_testing(
+    shared: &ProxySharedState,
+    pressure_event_seq: u64,
+    pressure_consumed_seq: u64,
+) {
+    let registry = &shared.middle_relay.relay_idle_registry;
+    let mut guard = match registry.lock() {
+        Ok(guard) => guard,
+        Err(poisoned) => {
+            let mut guard = poisoned.into_inner();
+            *guard = RelayIdleCandidateRegistry::default();
+            registry.clear_poison();
+            guard
+        }
+    };
+    guard.pressure_event_seq = pressure_event_seq;
+    guard.pressure_consumed_seq = pressure_consumed_seq;
+}

src/proxy/middle_relay/idle/read.rs

@@ -0,0 +1,442 @@
+use super::*;
+
+pub(crate) async fn read_client_payload_with_idle_policy_in<R>(
+    client_reader: &mut CryptoReader<R>,
+    proto_tag: ProtoTag,
+    max_frame: usize,
+    buffer_pool: &Arc<BufferPool>,
+    forensics: &RelayForensicsState,
+    frame_counter: &mut u64,
+    stats: &Stats,
+    shared: &ProxySharedState,
+    idle_policy: &RelayClientIdlePolicy,
+    idle_state: &mut RelayClientIdleState,
+    last_downstream_activity_ms: &AtomicU64,
+    session_started_at: Instant,
+) -> Result<Option<(PooledBuffer, bool)>>
+where
+    R: AsyncRead + Unpin + Send + 'static,
+{
+    const LEGACY_MAX_CONSECUTIVE_ZERO_LEN_FRAMES: u32 = 4;
+
+    async fn read_exact_with_policy<R>(
+        client_reader: &mut CryptoReader<R>,
+        buf: &mut [u8],
+        idle_policy: &RelayClientIdlePolicy,
+        idle_state: &mut RelayClientIdleState,
+        last_downstream_activity_ms: &AtomicU64,
+        session_started_at: Instant,
+        forensics: &RelayForensicsState,
+        stats: &Stats,
+        shared: &ProxySharedState,
+        read_label: &'static str,
+    ) -> Result<()>
+    where
+        R: AsyncRead + Unpin + Send + 'static,
+    {
+        fn hard_deadline(
+            idle_policy: &RelayClientIdlePolicy,
+            idle_state: &RelayClientIdleState,
+            session_started_at: Instant,
+            last_downstream_activity_ms: u64,
+        ) -> Instant {
+            let mut deadline = idle_state.last_client_frame_at + idle_policy.hard_idle;
+            if idle_policy.grace_after_downstream_activity.is_zero() {
+                return deadline;
+            }
+
+            let downstream_at =
+                session_started_at + Duration::from_millis(last_downstream_activity_ms);
+            if downstream_at > idle_state.last_client_frame_at {
+                let grace_deadline = downstream_at + idle_policy.grace_after_downstream_activity;
+                if grace_deadline > deadline {
+                    deadline = grace_deadline;
+                }
+            }
+            deadline
+        }
+
+        let mut filled = 0usize;
+        while filled < buf.len() {
+            let timeout_window = if idle_policy.enabled {
+                let now = Instant::now();
+                let downstream_ms = last_downstream_activity_ms.load(Ordering::Relaxed);
+                let hard_deadline =
+                    hard_deadline(idle_policy, idle_state, session_started_at, downstream_ms);
+                if !idle_state.soft_idle_marked
+                    && now.saturating_duration_since(idle_state.last_client_frame_at)
+                        >= idle_policy.soft_idle
+                {
+                    idle_state.soft_idle_marked = true;
+                    if mark_relay_idle_candidate_in(shared, forensics.conn_id) {
+                        stats.increment_relay_idle_soft_mark_total();
+                    }
+                    info!(
+                        trace_id = format_args!("0x{:016x}", forensics.trace_id),
+                        conn_id = forensics.conn_id,
+                        user = %forensics.user,
+                        read_label,
+                        soft_idle_secs = idle_policy.soft_idle.as_secs(),
+                        hard_idle_secs = idle_policy.hard_idle.as_secs(),
+                        grace_secs = idle_policy.grace_after_downstream_activity.as_secs(),
+                        "Middle-relay soft idle mark"
+                    );
+                }
+
+                let soft_deadline = idle_state.last_client_frame_at + idle_policy.soft_idle;
+                let next_deadline = if idle_state.soft_idle_marked {
+                    hard_deadline
+                } else {
+                    soft_deadline.min(hard_deadline)
+                };
+                let mut remaining = next_deadline.saturating_duration_since(now);
+                if remaining.is_zero() {
+                    remaining = Duration::from_millis(1);
+                }
+                remaining.min(RELAY_IDLE_IO_POLL_MAX)
+            } else {
+                idle_policy.legacy_frame_read_timeout
+            };
+
+            let read_result = timeout(timeout_window, client_reader.read(&mut buf[filled..])).await;
+            match read_result {
+                Ok(Ok(0)) => {
+                    return Err(ProxyError::Io(std::io::Error::from(
+                        std::io::ErrorKind::UnexpectedEof,
+                    )));
+                }
+                Ok(Ok(n)) => {
+                    filled = filled.saturating_add(n);
+                }
+                Ok(Err(e)) => return Err(ProxyError::Io(e)),
+                Err(_) if !idle_policy.enabled => {
+                    return Err(ProxyError::Io(std::io::Error::new(
+                        std::io::ErrorKind::TimedOut,
+                        format!(
+                            "middle-relay client frame read timeout while reading {read_label}"
+                        ),
+                    )));
+                }
+                Err(_) => {
+                    let now = Instant::now();
+                    let downstream_ms = last_downstream_activity_ms.load(Ordering::Relaxed);
+                    let hard_deadline =
+                        hard_deadline(idle_policy, idle_state, session_started_at, downstream_ms);
+                    if now >= hard_deadline {
+                        clear_relay_idle_candidate_in(shared, forensics.conn_id);
+                        stats.increment_relay_idle_hard_close_total();
+                        let client_idle_secs = now
+                            .saturating_duration_since(idle_state.last_client_frame_at)
+                            .as_secs();
+                        let downstream_idle_secs = now
+                            .saturating_duration_since(
+                                session_started_at + Duration::from_millis(downstream_ms),
+                            )
+                            .as_secs();
+                        warn!(
+                            trace_id = format_args!("0x{:016x}", forensics.trace_id),
+                            conn_id = forensics.conn_id,
+                            user = %forensics.user,
+                            read_label,
+                            client_idle_secs,
+                            downstream_idle_secs,
+                            soft_idle_secs = idle_policy.soft_idle.as_secs(),
+                            hard_idle_secs = idle_policy.hard_idle.as_secs(),
+                            grace_secs = idle_policy.grace_after_downstream_activity.as_secs(),
+                            "Middle-relay hard idle close"
+                        );
+                        return Err(ProxyError::Io(std::io::Error::new(
+                            std::io::ErrorKind::TimedOut,
+                            format!(
+                                "middle-relay hard idle timeout while reading {read_label}: client_idle_secs={client_idle_secs}, downstream_idle_secs={downstream_idle_secs}, soft_idle_secs={}, hard_idle_secs={}, grace_secs={}",
+                                idle_policy.soft_idle.as_secs(),
+                                idle_policy.hard_idle.as_secs(),
+                                idle_policy.grace_after_downstream_activity.as_secs(),
+                            ),
+                        )));
+                    }
+                }
+            }
+        }
+
+        Ok(())
+    }
+
+    let mut consecutive_zero_len_frames = 0u32;
+    loop {
+        let (len, quickack, raw_len_bytes) = match proto_tag {
+            ProtoTag::Abridged => {
+                let mut first = [0u8; 1];
+                match read_exact_with_policy(
+                    client_reader,
+                    &mut first,
+                    idle_policy,
+                    idle_state,
+                    last_downstream_activity_ms,
+                    session_started_at,
+                    forensics,
+                    stats,
+                    shared,
+                    "abridged.first_len_byte",
+                )
+                .await
+                {
+                    Ok(()) => {}
+                    Err(ProxyError::Io(e)) if e.kind() == std::io::ErrorKind::UnexpectedEof => {
+                        return Ok(None);
+                    }
+                    Err(e) => return Err(e),
+                }
+
+                let quickack = (first[0] & 0x80) != 0;
+                let len_words = if (first[0] & 0x7f) == 0x7f {
+                    let mut ext = [0u8; 3];
+                    read_exact_with_policy(
+                        client_reader,
+                        &mut ext,
+                        idle_policy,
+                        idle_state,
+                        last_downstream_activity_ms,
+                        session_started_at,
+                        forensics,
+                        stats,
+                        shared,
+                        "abridged.extended_len",
+                    )
+                    .await?;
+                    u32::from_le_bytes([ext[0], ext[1], ext[2], 0]) as usize
+                } else {
+                    (first[0] & 0x7f) as usize
+                };
+
+                let len = len_words
+                    .checked_mul(4)
+                    .ok_or_else(|| ProxyError::Proxy("Abridged frame length overflow".into()))?;
+                (len, quickack, None)
+            }
+            ProtoTag::Intermediate | ProtoTag::Secure => {
+                let mut len_buf = [0u8; 4];
+                match read_exact_with_policy(
+                    client_reader,
+                    &mut len_buf,
+                    idle_policy,
+                    idle_state,
+                    last_downstream_activity_ms,
+                    session_started_at,
+                    forensics,
+                    stats,
+                    shared,
+                    "len_prefix",
+                )
+                .await
+                {
+                    Ok(()) => {}
+                    Err(ProxyError::Io(e)) if e.kind() == std::io::ErrorKind::UnexpectedEof => {
+                        return Ok(None);
+                    }
+                    Err(e) => return Err(e),
+                }
+                let quickack = (len_buf[3] & 0x80) != 0;
+                (
+                    (u32::from_le_bytes(len_buf) & 0x7fff_ffff) as usize,
+                    quickack,
+                    Some(len_buf),
+                )
+            }
+        };
+
+        if len == 0 {
+            idle_state.on_client_tiny_frame(Instant::now());
+            idle_state.tiny_frame_debt = idle_state
+                .tiny_frame_debt
+                .saturating_add(TINY_FRAME_DEBT_PER_TINY);
+            if idle_state.tiny_frame_debt >= TINY_FRAME_DEBT_LIMIT {
+                stats.increment_relay_protocol_desync_close_total();
+                return Err(ProxyError::Proxy(format!(
+                    "Tiny frame overhead limit exceeded: debt={}, conn_id={}",
+                    idle_state.tiny_frame_debt, forensics.conn_id
+                )));
+            }
+
+            if !idle_policy.enabled {
+                consecutive_zero_len_frames = consecutive_zero_len_frames.saturating_add(1);
+                if consecutive_zero_len_frames > LEGACY_MAX_CONSECUTIVE_ZERO_LEN_FRAMES {
+                    stats.increment_relay_protocol_desync_close_total();
+                    return Err(ProxyError::Proxy(
+                        "Excessive zero-length abridged frames".to_string(),
+                    ));
+                }
+            }
+            continue;
+        }
+        if len < 4 && proto_tag != ProtoTag::Abridged {
+            warn!(
+                trace_id = format_args!("0x{:016x}", forensics.trace_id),
+                conn_id = forensics.conn_id,
+                user = %forensics.user,
+                len,
+                proto = ?proto_tag,
+                "Frame too small — corrupt or probe"
+            );
+            stats.increment_relay_protocol_desync_close_total();
+            return Err(ProxyError::Proxy(format!("Frame too small: {len}")));
+        }
+
+        if len > max_frame {
+            return Err(report_desync_frame_too_large_in(
+                shared,
+                forensics,
+                proto_tag,
+                *frame_counter,
+                max_frame,
+                len,
+                raw_len_bytes,
+                stats,
+            ));
+        }
+
+        let secure_payload_len = if proto_tag == ProtoTag::Secure {
+            match secure_payload_len_from_wire_len(len) {
+                Some(payload_len) => payload_len,
+                None => {
+                    stats.increment_secure_padding_invalid();
+                    stats.increment_relay_protocol_desync_close_total();
+                    return Err(ProxyError::Proxy(format!(
+                        "Invalid secure frame length: {len}"
+                    )));
+                }
+            }
+        } else {
+            len
+        };
+
+        let mut payload = buffer_pool.get();
+        payload.clear();
+        let current_cap = payload.capacity();
+        if current_cap < len {
+            payload.reserve(len - current_cap);
+        }
+        payload.resize(len, 0);
+        read_exact_with_policy(
+            client_reader,
+            &mut payload[..len],
+            idle_policy,
+            idle_state,
+            last_downstream_activity_ms,
+            session_started_at,
+            forensics,
+            stats,
+            shared,
+            "payload",
+        )
+        .await?;
+
+        // Secure Intermediate: strip validated trailing padding bytes.
+        if proto_tag == ProtoTag::Secure {
+            payload.truncate(secure_payload_len);
+        }
+        *frame_counter += 1;
+        idle_state.on_client_frame(Instant::now());
+        idle_state.tiny_frame_debt = idle_state.tiny_frame_debt.saturating_sub(1);
+        clear_relay_idle_candidate_in(shared, forensics.conn_id);
+        return Ok(Some((payload, quickack)));
+    }
+}
+
+#[cfg(test)]
+pub(crate) async fn read_client_payload_with_idle_policy<R>(
+    client_reader: &mut CryptoReader<R>,
+    proto_tag: ProtoTag,
+    max_frame: usize,
+    buffer_pool: &Arc<BufferPool>,
+    forensics: &RelayForensicsState,
+    frame_counter: &mut u64,
+    stats: &Stats,
+    idle_policy: &RelayClientIdlePolicy,
+    idle_state: &mut RelayClientIdleState,
+    last_downstream_activity_ms: &AtomicU64,
+    session_started_at: Instant,
+) -> Result<Option<(PooledBuffer, bool)>>
+where
+    R: AsyncRead + Unpin + Send + 'static,
+{
+    let shared = ProxySharedState::new();
+    read_client_payload_with_idle_policy_in(
+        client_reader,
+        proto_tag,
+        max_frame,
+        buffer_pool,
+        forensics,
+        frame_counter,
+        stats,
+        shared.as_ref(),
+        idle_policy,
+        idle_state,
+        last_downstream_activity_ms,
+        session_started_at,
+    )
+    .await
+}
+
+#[cfg(test)]
+pub(crate) async fn read_client_payload_legacy<R>(
+    client_reader: &mut CryptoReader<R>,
+    proto_tag: ProtoTag,
+    max_frame: usize,
+    frame_read_timeout: Duration,
+    buffer_pool: &Arc<BufferPool>,
+    forensics: &RelayForensicsState,
+    frame_counter: &mut u64,
+    stats: &Stats,
+) -> Result<Option<(PooledBuffer, bool)>>
+where
+    R: AsyncRead + Unpin + Send + 'static,
+{
+    let now = Instant::now();
+    let shared = ProxySharedState::new();
+    let mut idle_state = RelayClientIdleState::new(now);
+    let last_downstream_activity_ms = AtomicU64::new(0);
+    let idle_policy = RelayClientIdlePolicy::disabled(frame_read_timeout);
+    read_client_payload_with_idle_policy_in(
+        client_reader,
+        proto_tag,
+        max_frame,
+        buffer_pool,
+        forensics,
+        frame_counter,
+        stats,
+        shared.as_ref(),
+        &idle_policy,
+        &mut idle_state,
+        &last_downstream_activity_ms,
+        now,
+    )
+    .await
+}
+
+#[cfg(test)]
+pub(crate) async fn read_client_payload<R>(
+    client_reader: &mut CryptoReader<R>,
+    proto_tag: ProtoTag,
+    max_frame: usize,
+    frame_read_timeout: Duration,
+    buffer_pool: &Arc<BufferPool>,
+    forensics: &RelayForensicsState,
+    frame_counter: &mut u64,
+    stats: &Stats,
+) -> Result<Option<(PooledBuffer, bool)>>
+where
+    R: AsyncRead + Unpin + Send + 'static,
+{
+    read_client_payload_legacy(
+        client_reader,
+        proto_tag,
+        max_frame,
+        frame_read_timeout,
+        buffer_pool,
+        forensics,
+        frame_counter,
+        stats,
+    )
+    .await
+}

src/proxy/middle_relay/quota.rs

@@ -0,0 +1,151 @@
+use super::*;
+
+pub(super) enum MiddleQuotaReserveError {
+    LimitExceeded,
+    Contended,
+    Cancelled,
+    DeadlineExceeded,
+}
+
+pub(super) fn quota_soft_cap(limit: u64, overshoot: u64) -> u64 {
+    limit.saturating_add(overshoot)
+}
+
+pub(super) async fn reserve_user_quota_with_yield(
+    user_stats: &UserStats,
+    bytes: u64,
+    limit: u64,
+    stats: &Stats,
+    cancel: &CancellationToken,
+    deadline: Option<Instant>,
+) -> std::result::Result<u64, MiddleQuotaReserveError> {
+    let mut backoff_ms = QUOTA_RESERVE_BACKOFF_MIN_MS;
+    let mut backoff_rounds = 0usize;
+    loop {
+        for _ in 0..QUOTA_RESERVE_SPIN_RETRIES {
+            match user_stats.quota_try_reserve(bytes, limit) {
+                Ok(total) => return Ok(total),
+                Err(QuotaReserveError::LimitExceeded) => {
+                    return Err(MiddleQuotaReserveError::LimitExceeded);
+                }
+                Err(QuotaReserveError::Contended) => {
+                    stats.increment_quota_contention_total();
+                    std::hint::spin_loop();
+                }
+            }
+        }
+
+        tokio::task::yield_now().await;
+        if deadline.is_some_and(|deadline| Instant::now() >= deadline) {
+            stats.increment_quota_contention_timeout_total();
+            return Err(MiddleQuotaReserveError::DeadlineExceeded);
+        }
+        tokio::select! {
+            _ = tokio::time::sleep(Duration::from_millis(backoff_ms)) => {}
+            _ = cancel.cancelled() => {
+                stats.increment_quota_acquire_cancelled_total();
+                return Err(MiddleQuotaReserveError::Cancelled);
+            }
+        }
+        backoff_rounds = backoff_rounds.saturating_add(1);
+        if backoff_rounds >= QUOTA_RESERVE_MAX_BACKOFF_ROUNDS {
+            stats.increment_quota_contention_timeout_total();
+            return Err(MiddleQuotaReserveError::Contended);
+        }
+        backoff_ms = backoff_ms
+            .saturating_mul(2)
+            .min(QUOTA_RESERVE_BACKOFF_MAX_MS);
+    }
+}
+
+pub(super) async fn wait_for_traffic_budget(
+    lease: Option<&Arc<TrafficLease>>,
+    direction: RateDirection,
+    bytes: u64,
+    deadline: Option<Instant>,
+) -> Result<()> {
+    if bytes == 0 {
+        return Ok(());
+    }
+    let Some(lease) = lease else {
+        return Ok(());
+    };
+
+    let mut remaining = bytes;
+    while remaining > 0 {
+        let consume = lease.try_consume(direction, remaining);
+        if consume.granted > 0 {
+            remaining = remaining.saturating_sub(consume.granted);
+            continue;
+        }
+
+        let wait_started_at = Instant::now();
+        if deadline.is_some_and(|deadline| wait_started_at >= deadline) {
+            return Err(ProxyError::TrafficBudgetWaitDeadlineExceeded);
+        }
+        tokio::time::sleep(next_refill_delay()).await;
+        let wait_ms = wait_started_at
+            .elapsed()
+            .as_millis()
+            .min(u128::from(u64::MAX)) as u64;
+        lease.observe_wait_ms(
+            direction,
+            consume.blocked_user,
+            consume.blocked_cidr,
+            wait_ms,
+        );
+    }
+
+    Ok(())
+}
+
+pub(super) async fn wait_for_traffic_budget_or_cancel(
+    lease: Option<&Arc<TrafficLease>>,
+    direction: RateDirection,
+    bytes: u64,
+    cancel: &CancellationToken,
+    stats: &Stats,
+    deadline: Option<Instant>,
+) -> Result<()> {
+    if bytes == 0 {
+        return Ok(());
+    }
+    let Some(lease) = lease else {
+        return Ok(());
+    };
+
+    let mut remaining = bytes;
+    while remaining > 0 {
+        let consume = lease.try_consume(direction, remaining);
+        if consume.granted > 0 {
+            remaining = remaining.saturating_sub(consume.granted);
+            continue;
+        }
+
+        let wait_started_at = Instant::now();
+        if deadline.is_some_and(|deadline| wait_started_at >= deadline) {
+            stats.increment_flow_wait_middle_rate_limit_cancelled_total();
+            return Err(ProxyError::TrafficBudgetWaitDeadlineExceeded);
+        }
+        tokio::select! {
+            _ = tokio::time::sleep(next_refill_delay()) => {}
+            _ = cancel.cancelled() => {
+                stats.increment_flow_wait_middle_rate_limit_cancelled_total();
+                return Err(ProxyError::TrafficBudgetWaitCancelled);
+            }
+        }
+        let wait_ms = wait_started_at
+            .elapsed()
+            .as_millis()
+            .min(u128::from(u64::MAX)) as u64;
+        lease.observe_wait_ms(
+            direction,
+            consume.blocked_user,
+            consume.blocked_cidr,
+            wait_ms,
+        );
+        stats.observe_flow_wait_middle_rate_limit_ms(wait_ms);
+    }
+
+    Ok(())
+}

src/proxy/middle_relay/session.rs

@@ -0,0 +1,830 @@
+use super::*;
+
+pub(crate) async fn handle_via_middle_proxy<R, W>(
+    mut crypto_reader: CryptoReader<R>,
+    crypto_writer: CryptoWriter<W>,
+    success: HandshakeSuccess,
+    me_pool: Arc<MePool>,
+    stats: Arc<Stats>,
+    config: Arc<ProxyConfig>,
+    buffer_pool: Arc<BufferPool>,
+    local_addr: SocketAddr,
+    rng: Arc<SecureRandom>,
+    mut route_rx: watch::Receiver<RouteCutoverState>,
+    route_snapshot: RouteCutoverState,
+    session_id: u64,
+    shared: Arc<ProxySharedState>,
+) -> Result<()>
+where
+    R: AsyncRead + Unpin + Send + 'static,
+    W: AsyncWrite + Unpin + Send + 'static,
+{
+    let user = success.user.clone();
+    let quota_limit = config.access.user_data_quota.get(&user).copied();
+    let quota_user_stats = quota_limit.map(|_| stats.get_or_create_user_stats_handle(&user));
+    let peer = success.peer;
+    let traffic_lease = shared.traffic_limiter.acquire_lease(&user, peer.ip());
+    let proto_tag = success.proto_tag;
+    let pool_generation = me_pool.current_generation();
+
+    debug!(
+        user = %user,
+        peer = %peer,
+        dc = success.dc_idx,
+        proto = ?proto_tag,
+        mode = "middle_proxy",
+        pool_generation,
+        "Routing via Middle-End"
+    );
+
+    let (conn_id, me_rx) = me_pool.registry().register().await;
+    let trace_id = session_id;
+    let bytes_me2c = Arc::new(AtomicU64::new(0));
+    let mut forensics = RelayForensicsState {
+        trace_id,
+        conn_id,
+        user: user.clone(),
+        peer,
+        peer_hash: hash_ip_in(shared.as_ref(), peer.ip()),
+        started_at: Instant::now(),
+        bytes_c2me: 0,
+        bytes_me2c: bytes_me2c.clone(),
+        desync_all_full: config.general.desync_all_full,
+    };
+
+    stats.increment_user_connects(&user);
+    let _me_connection_lease = stats.acquire_me_connection_lease();
+
+    if let Some(cutover) =
+        affected_cutover_state(&route_rx, RelayRouteMode::Middle, route_snapshot.generation)
+    {
+        let delay = cutover_stagger_delay(session_id, cutover.generation);
+        warn!(
+            conn_id,
+            target_mode = cutover.mode.as_str(),
+            cutover_generation = cutover.generation,
+            delay_ms = delay.as_millis() as u64,
+            "Cutover affected middle session before relay start, closing client connection"
+        );
+        let _cutover_park_lease = stats.acquire_middle_cutover_park_lease();
+        tokio::time::sleep(delay).await;
+        let _ = me_pool.send_close(conn_id).await;
+        me_pool.registry().unregister(conn_id).await;
+        return Err(ProxyError::RouteSwitched);
+    }
+
+    // Per-user ad_tag from access.user_ad_tags; fallback to general.ad_tag (hot-reloadable)
+    let user_tag: Option<Vec<u8>> = config
+        .access
+        .user_ad_tags
+        .get(&user)
+        .and_then(|s| hex::decode(s).ok())
+        .filter(|v| v.len() == 16);
+    let global_tag: Option<Vec<u8>> = config
+        .general
+        .ad_tag
+        .as_ref()
+        .and_then(|s| hex::decode(s).ok())
+        .filter(|v| v.len() == 16);
+    let effective_tag = user_tag.or(global_tag);
+
+    let proto_flags = proto_flags_for_tag(proto_tag, effective_tag.is_some());
+    let effective_tag_array = effective_tag
+        .as_deref()
+        .and_then(|tag| <[u8; 16]>::try_from(tag).ok());
+    debug!(
+        trace_id = format_args!("0x{:016x}", trace_id),
+        user = %user,
+        conn_id,
+        peer_hash = format_args!("0x{:016x}", forensics.peer_hash),
+        desync_all_full = forensics.desync_all_full,
+        proto_flags = format_args!("0x{:08x}", proto_flags),
+        pool_generation,
+        "ME relay started"
+    );
+
+    let translated_local_addr = me_pool.translate_our_addr(local_addr);
+
+    let frame_limit = config.general.max_client_frame;
+    let mut relay_idle_policy = RelayClientIdlePolicy::from_config(&config);
+    let mut pressure_caps_applied = false;
+    if shared.conntrack_pressure_active() {
+        relay_idle_policy.apply_pressure_caps(config.server.conntrack_control.profile);
+        pressure_caps_applied = true;
+    }
+    let session_started_at = forensics.started_at;
+    let mut relay_idle_state = RelayClientIdleState::new(session_started_at);
+    let last_downstream_activity_ms = Arc::new(AtomicU64::new(0));
+
+    let c2me_channel_capacity = config
+        .general
+        .me_c2me_channel_capacity
+        .max(C2ME_CHANNEL_CAPACITY_FALLBACK);
+    let c2me_send_timeout = match config.general.me_c2me_send_timeout_ms {
+        0 => None,
+        timeout_ms => Some(Duration::from_millis(timeout_ms)),
+    };
+    let c2me_byte_budget = c2me_queued_permit_budget(c2me_channel_capacity, frame_limit);
+    let c2me_byte_semaphore = Arc::new(Semaphore::new(c2me_byte_budget));
+    let (c2me_tx, mut c2me_rx) = mpsc::channel::<C2MeCommand>(c2me_channel_capacity);
+    let me_pool_c2me = me_pool.clone();
+    let mut c2me_sender = tokio::spawn(async move {
+        let mut sent_since_yield = 0usize;
+        while let Some(cmd) = c2me_rx.recv().await {
+            match cmd {
+                C2MeCommand::Data {
+                    payload,
+                    flags,
+                    _permit,
+                } => {
+                    me_pool_c2me
+                        .send_proxy_req_pooled(
+                            conn_id,
+                            success.dc_idx,
+                            peer,
+                            translated_local_addr,
+                            payload,
+                            flags,
+                            effective_tag_array,
+                        )
+                        .await?;
+                    sent_since_yield = sent_since_yield.saturating_add(1);
+                    if should_yield_c2me_sender(sent_since_yield, !c2me_rx.is_empty()) {
+                        sent_since_yield = 0;
+                        tokio::task::yield_now().await;
+                    }
+                }
+                C2MeCommand::Close => {
+                    let _ = me_pool_c2me.send_close(conn_id).await;
+                    return Ok(());
+                }
+            }
+        }
+        Ok(())
+    });
+
+    let (stop_tx, mut stop_rx) = oneshot::channel::<()>();
+    let flow_cancel = CancellationToken::new();
+    let mut me_rx_task = me_rx;
+    let stats_clone = stats.clone();
+    let rng_clone = rng.clone();
+    let user_clone = user.clone();
+    let quota_user_stats_me_writer = quota_user_stats.clone();
+    let traffic_lease_me_writer = traffic_lease.clone();
+    let flow_cancel_me_writer = flow_cancel.clone();
+    let last_downstream_activity_ms_clone = last_downstream_activity_ms.clone();
+    let bytes_me2c_clone = bytes_me2c.clone();
+    let d2c_flush_policy = MeD2cFlushPolicy::from_config(&config);
+    let mut me_writer = tokio::spawn(async move {
+        let mut writer = crypto_writer;
+        let mut frame_buf = Vec::with_capacity(16 * 1024);
+        let shrink_threshold = d2c_flush_policy.frame_buf_shrink_threshold_bytes;
+
+        fn shrink_session_vec(buf: &mut Vec<u8>, threshold: usize) {
+            if buf.capacity() > threshold {
+                buf.clear();
+                buf.shrink_to(threshold);
+            } else {
+                buf.clear();
+            }
+        }
+
+        loop {
+            tokio::select! {
+                msg = me_rx_task.recv() => {
+                    let Some(first) = msg else {
+                        debug!(conn_id, "ME channel closed");
+                        shrink_session_vec(&mut frame_buf, shrink_threshold);
+                        return Err(ProxyError::MiddleConnectionLost);
+                    };
+
+                    let mut batch_frames = 0usize;
+                    let mut batch_bytes = 0usize;
+                    let mut flush_immediately;
+                    let mut max_delay_fired = false;
+
+                    let first_is_downstream_activity =
+                        matches!(&first, MeResponse::Data { .. } | MeResponse::Ack(_));
+                    match process_me_writer_response_with_traffic_lease(
+                        first,
+                        &mut writer,
+                        proto_tag,
+                        rng_clone.as_ref(),
+                        &mut frame_buf,
+                        stats_clone.as_ref(),
+                        &user_clone,
+                        quota_user_stats_me_writer.as_deref(),
+                        quota_limit,
+                        d2c_flush_policy.quota_soft_overshoot_bytes,
+                        traffic_lease_me_writer.as_ref(),
+                        &flow_cancel_me_writer,
+                        bytes_me2c_clone.as_ref(),
+                        conn_id,
+                        d2c_flush_policy.ack_flush_immediate,
+                        false,
+                    ).await? {
+                        MeWriterResponseOutcome::Continue { frames, bytes, flush_immediately: immediate } => {
+                            if first_is_downstream_activity {
+                                last_downstream_activity_ms_clone
+                                    .store(session_started_at.elapsed().as_millis() as u64, Ordering::Relaxed);
+                            }
+                            batch_frames = batch_frames.saturating_add(frames);
+                            batch_bytes = batch_bytes.saturating_add(bytes);
+                            flush_immediately = immediate;
+                        }
+                        MeWriterResponseOutcome::Close => {
+                            let flush_started_at = if stats_clone.telemetry_policy().me_level.allows_debug() {
+                                Some(Instant::now())
+                            } else {
+                                None
+                            };
+                            let _ = flush_client_or_cancel(&mut writer, &flow_cancel_me_writer).await;
+                            let flush_duration_us = flush_started_at.map(|started| {
+                                started
+                                    .elapsed()
+                                    .as_micros()
+                                    .min(u128::from(u64::MAX)) as u64
+                            });
+                            observe_me_d2c_flush_event(
+                                stats_clone.as_ref(),
+                                MeD2cFlushReason::Close,
+                                batch_frames,
+                                batch_bytes,
+                                flush_duration_us,
+                            );
+                            shrink_session_vec(&mut frame_buf, shrink_threshold);
+                            return Ok(());
+                        }
+                    }
+
+                    while !flush_immediately
+                        && batch_frames < d2c_flush_policy.max_frames
+                        && batch_bytes < d2c_flush_policy.max_bytes
+                    {
+                        let Ok(next) = me_rx_task.try_recv() else {
+                            break;
+                        };
+
+                        let next_is_downstream_activity =
+                            matches!(&next, MeResponse::Data { .. } | MeResponse::Ack(_));
+                        match process_me_writer_response_with_traffic_lease(
+                            next,
+                            &mut writer,
+                            proto_tag,
+                            rng_clone.as_ref(),
+                            &mut frame_buf,
+                            stats_clone.as_ref(),
+                            &user_clone,
+                            quota_user_stats_me_writer.as_deref(),
+                            quota_limit,
+                            d2c_flush_policy.quota_soft_overshoot_bytes,
+                            traffic_lease_me_writer.as_ref(),
+                            &flow_cancel_me_writer,
+                            bytes_me2c_clone.as_ref(),
+                            conn_id,
+                            d2c_flush_policy.ack_flush_immediate,
+                            true,
+                        ).await? {
+                            MeWriterResponseOutcome::Continue { frames, bytes, flush_immediately: immediate } => {
+                                if next_is_downstream_activity {
+                                    last_downstream_activity_ms_clone
+                                        .store(session_started_at.elapsed().as_millis() as u64, Ordering::Relaxed);
+                                }
+                                batch_frames = batch_frames.saturating_add(frames);
+                                batch_bytes = batch_bytes.saturating_add(bytes);
+                                flush_immediately |= immediate;
+                            }
+                            MeWriterResponseOutcome::Close => {
+                                let flush_started_at =
+                                    if stats_clone.telemetry_policy().me_level.allows_debug() {
+                                        Some(Instant::now())
+                                    } else {
+                                        None
+                                    };
+                                let _ =
+                                    flush_client_or_cancel(&mut writer, &flow_cancel_me_writer).await;
+                                let flush_duration_us = flush_started_at.map(|started| {
+                                    started
+                                        .elapsed()
+                                        .as_micros()
+                                        .min(u128::from(u64::MAX))
+                                        as u64
+                                });
+                                observe_me_d2c_flush_event(
+                                    stats_clone.as_ref(),
+                                    MeD2cFlushReason::Close,
+                                    batch_frames,
+                                    batch_bytes,
+                                    flush_duration_us,
+                                );
+                                shrink_session_vec(&mut frame_buf, shrink_threshold);
+                                return Ok(());
+                            }
+                        }
+                    }
+
+                    if !flush_immediately
+                        && !d2c_flush_policy.max_delay.is_zero()
+                        && batch_frames < d2c_flush_policy.max_frames
+                        && batch_bytes < d2c_flush_policy.max_bytes
+                    {
+                        stats_clone.increment_me_d2c_batch_timeout_armed_total();
+                        match tokio::time::timeout(d2c_flush_policy.max_delay, me_rx_task.recv()).await {
+                            Ok(Some(next)) => {
+                                let next_is_downstream_activity =
+                                    matches!(&next, MeResponse::Data { .. } | MeResponse::Ack(_));
+                                match process_me_writer_response_with_traffic_lease(
+                                    next,
+                                    &mut writer,
+                                    proto_tag,
+                                    rng_clone.as_ref(),
+                                    &mut frame_buf,
+                                    stats_clone.as_ref(),
+                                    &user_clone,
+                                    quota_user_stats_me_writer.as_deref(),
+                                    quota_limit,
+                                    d2c_flush_policy.quota_soft_overshoot_bytes,
+                                    traffic_lease_me_writer.as_ref(),
+                                    &flow_cancel_me_writer,
+                                    bytes_me2c_clone.as_ref(),
+                                    conn_id,
+                                    d2c_flush_policy.ack_flush_immediate,
+                                    true,
+                                ).await? {
+                                    MeWriterResponseOutcome::Continue { frames, bytes, flush_immediately: immediate } => {
+                                        if next_is_downstream_activity {
+                                            last_downstream_activity_ms_clone
+                                                .store(session_started_at.elapsed().as_millis() as u64, Ordering::Relaxed);
+                                        }
+                                        batch_frames = batch_frames.saturating_add(frames);
+                                        batch_bytes = batch_bytes.saturating_add(bytes);
+                                        flush_immediately |= immediate;
+                                    }
+                                    MeWriterResponseOutcome::Close => {
+                                        let flush_started_at = if stats_clone
+                                            .telemetry_policy()
+                                            .me_level
+                                            .allows_debug()
+                                        {
+                                            Some(Instant::now())
+                                        } else {
+                                            None
+                                        };
+                                        let _ = flush_client_or_cancel(
+                                            &mut writer,
+                                            &flow_cancel_me_writer,
+                                        )
+                                        .await;
+                                        let flush_duration_us = flush_started_at.map(|started| {
+                                            started
+                                                .elapsed()
+                                                .as_micros()
+                                                .min(u128::from(u64::MAX))
+                                                as u64
+                                        });
+                                        observe_me_d2c_flush_event(
+                                            stats_clone.as_ref(),
+                                            MeD2cFlushReason::Close,
+                                            batch_frames,
+                                            batch_bytes,
+                                            flush_duration_us,
+                                        );
+                                        shrink_session_vec(&mut frame_buf, shrink_threshold);
+                                        return Ok(());
+                                    }
+                                }
+
+                                while !flush_immediately
+                                    && batch_frames < d2c_flush_policy.max_frames
+                                    && batch_bytes < d2c_flush_policy.max_bytes
+                                {
+                                    let Ok(extra) = me_rx_task.try_recv() else {
+                                        break;
+                                    };
+
+                                    let extra_is_downstream_activity =
+                                        matches!(&extra, MeResponse::Data { .. } | MeResponse::Ack(_));
+                                    match process_me_writer_response_with_traffic_lease(
+                                        extra,
+                                        &mut writer,
+                                        proto_tag,
+                                        rng_clone.as_ref(),
+                                        &mut frame_buf,
+                                        stats_clone.as_ref(),
+                                        &user_clone,
+                                        quota_user_stats_me_writer.as_deref(),
+                                        quota_limit,
+                                        d2c_flush_policy.quota_soft_overshoot_bytes,
+                                        traffic_lease_me_writer.as_ref(),
+                                        &flow_cancel_me_writer,
+                                        bytes_me2c_clone.as_ref(),
+                                        conn_id,
+                                        d2c_flush_policy.ack_flush_immediate,
+                                        true,
+                                    ).await? {
+                                        MeWriterResponseOutcome::Continue { frames, bytes, flush_immediately: immediate } => {
+                                            if extra_is_downstream_activity {
+                                                last_downstream_activity_ms_clone
+                                                    .store(session_started_at.elapsed().as_millis() as u64, Ordering::Relaxed);
+                                            }
+                                            batch_frames = batch_frames.saturating_add(frames);
+                                            batch_bytes = batch_bytes.saturating_add(bytes);
+                                            flush_immediately |= immediate;
+                                        }
+                                        MeWriterResponseOutcome::Close => {
+                                            let flush_started_at = if stats_clone
+                                                .telemetry_policy()
+                                                .me_level
+                                                .allows_debug()
+                                            {
+                                                Some(Instant::now())
+                                            } else {
+                                                None
+                                            };
+                                            let _ = flush_client_or_cancel(
+                                                &mut writer,
+                                                &flow_cancel_me_writer,
+                                            )
+                                            .await;
+                                            let flush_duration_us = flush_started_at.map(|started| {
+                                                started
+                                                    .elapsed()
+                                                    .as_micros()
+                                                    .min(u128::from(u64::MAX))
+                                                    as u64
+                                            });
+                                            observe_me_d2c_flush_event(
+                                                stats_clone.as_ref(),
+                                                MeD2cFlushReason::Close,
+                                                batch_frames,
+                                                batch_bytes,
+                                                flush_duration_us,
+                                            );
+                                            shrink_session_vec(&mut frame_buf, shrink_threshold);
+                                            return Ok(());
+                                        }
+                                    }
+                                }
+                            }
+                            Ok(None) => {
+                                debug!(conn_id, "ME channel closed");
+                                shrink_session_vec(&mut frame_buf, shrink_threshold);
+                                return Err(ProxyError::MiddleConnectionLost);
+                            }
+                            Err(_) => {
+                                max_delay_fired = true;
+                                stats_clone.increment_me_d2c_batch_timeout_fired_total();
+                            }
+                        }
+                    }
+
+                    let flush_reason = classify_me_d2c_flush_reason(
+                        flush_immediately,
+                        batch_frames,
+                        d2c_flush_policy.max_frames,
+                        batch_bytes,
+                        d2c_flush_policy.max_bytes,
+                        max_delay_fired,
+                    );
+                    let flush_started_at = if stats_clone.telemetry_policy().me_level.allows_debug() {
+                        Some(Instant::now())
+                    } else {
+                        None
+                    };
+                    flush_client_or_cancel(&mut writer, &flow_cancel_me_writer).await?;
+                    let flush_duration_us = flush_started_at.map(|started| {
+                        started
+                            .elapsed()
+                            .as_micros()
+                            .min(u128::from(u64::MAX)) as u64
+                    });
+                    observe_me_d2c_flush_event(
+                        stats_clone.as_ref(),
+                        flush_reason,
+                        batch_frames,
+                        batch_bytes,
+                        flush_duration_us,
+                    );
+                    let shrink_threshold = d2c_flush_policy.frame_buf_shrink_threshold_bytes;
+                    let shrink_trigger = shrink_threshold
+                        .saturating_mul(ME_D2C_FRAME_BUF_SHRINK_HYSTERESIS_FACTOR);
+                    if frame_buf.capacity() > shrink_trigger {
+                        let cap_before = frame_buf.capacity();
+                        frame_buf.shrink_to(shrink_threshold);
+                        let cap_after = frame_buf.capacity();
+                        let bytes_freed = cap_before.saturating_sub(cap_after) as u64;
+                        stats_clone.observe_me_d2c_frame_buf_shrink(bytes_freed);
+                    }
+                }
+                _ = &mut stop_rx => {
+                    debug!(conn_id, "ME writer stop signal");
+                    shrink_session_vec(&mut frame_buf, shrink_threshold);
+                    return Ok(());
+                }
+            }
+        }
+    });
+
+    let mut main_result: Result<()> = Ok(());
+    let mut client_closed = false;
+    let mut frame_counter: u64 = 0;
+    let mut route_watch_open = true;
+    let mut seen_pressure_seq = relay_pressure_event_seq_in(shared.as_ref());
+    loop {
+        if shared.conntrack_pressure_active() && !pressure_caps_applied {
+            relay_idle_policy.apply_pressure_caps(config.server.conntrack_control.profile);
+            pressure_caps_applied = true;
+        }
+
+        if relay_idle_policy.enabled
+            && maybe_evict_idle_candidate_on_pressure_in(
+                shared.as_ref(),
+                conn_id,
+                &mut seen_pressure_seq,
+                stats.as_ref(),
+            )
+        {
+            info!(
+                conn_id,
+                trace_id = format_args!("0x{:016x}", trace_id),
+                user = %user,
+                "Middle-relay pressure eviction for idle-candidate session"
+            );
+            let _ = enqueue_c2me_command_in(
+                shared.as_ref(),
+                &c2me_tx,
+                C2MeCommand::Close,
+                c2me_send_timeout,
+                stats.as_ref(),
+            )
+            .await;
+            main_result = Err(ProxyError::Proxy(
+                "middle-relay session evicted under pressure (idle-candidate)".to_string(),
+            ));
+            break;
+        }
+
+        if let Some(cutover) =
+            affected_cutover_state(&route_rx, RelayRouteMode::Middle, route_snapshot.generation)
+        {
+            let delay = cutover_stagger_delay(session_id, cutover.generation);
+            warn!(
+                conn_id,
+                target_mode = cutover.mode.as_str(),
+                cutover_generation = cutover.generation,
+                delay_ms = delay.as_millis() as u64,
+                "Cutover affected middle session, closing client connection"
+            );
+            let _cutover_park_lease = stats.acquire_middle_cutover_park_lease();
+            tokio::time::sleep(delay).await;
+            let _ = enqueue_c2me_command_in(
+                shared.as_ref(),
+                &c2me_tx,
+                C2MeCommand::Close,
+                c2me_send_timeout,
+                stats.as_ref(),
+            )
+            .await;
+            main_result = Err(ProxyError::RouteSwitched);
+            break;
+        }
+
+        tokio::select! {
+            changed = route_rx.changed(), if route_watch_open => {
+                if changed.is_err() {
+                    route_watch_open = false;
+                }
+            }
+            payload_result = read_client_payload_with_idle_policy_in(
+                &mut crypto_reader,
+                proto_tag,
+                frame_limit,
+                &buffer_pool,
+                &forensics,
+                &mut frame_counter,
+                &stats,
+                shared.as_ref(),
+                &relay_idle_policy,
+                &mut relay_idle_state,
+                last_downstream_activity_ms.as_ref(),
+                session_started_at,
+            ) => {
+                match payload_result {
+                    Ok(Some((payload, quickack))) => {
+                        trace!(conn_id, bytes = payload.len(), "C->ME frame");
+                        wait_for_traffic_budget(
+                            traffic_lease.as_ref(),
+                            RateDirection::Up,
+                            payload.len() as u64,
+                            None,
+                        )
+                        .await?;
+                        forensics.bytes_c2me = forensics
+                            .bytes_c2me
+                            .saturating_add(payload.len() as u64);
+                        if let (Some(limit), Some(user_stats)) =
+                            (quota_limit, quota_user_stats.as_deref())
+                        {
+                            match reserve_user_quota_with_yield(
+                                user_stats,
+                                payload.len() as u64,
+                                limit,
+                                stats.as_ref(),
+                                &flow_cancel,
+                                None,
+                            )
+                            .await
+                            {
+                                Ok(_) => {}
+                                Err(MiddleQuotaReserveError::LimitExceeded) => {
+                                    main_result = Err(ProxyError::DataQuotaExceeded {
+                                        user: user.clone(),
+                                    });
+                                    break;
+                                }
+                                Err(MiddleQuotaReserveError::Contended) => {
+                                    main_result = Err(ProxyError::Proxy(
+                                        "ME C->ME quota reservation contended".into(),
+                                    ));
+                                    break;
+                                }
+                                Err(MiddleQuotaReserveError::Cancelled) => {
+                                    main_result = Err(ProxyError::Proxy(
+                                        "ME C->ME quota reservation cancelled".into(),
+                                    ));
+                                    break;
+                                }
+                                Err(MiddleQuotaReserveError::DeadlineExceeded) => {
+                                    main_result = Err(ProxyError::Proxy(
+                                        "ME C->ME quota reservation deadline exceeded".into(),
+                                    ));
+                                    break;
+                                }
+                            }
+                            stats.add_user_octets_from_handle(user_stats, payload.len() as u64);
+                        } else {
+                            stats.add_user_octets_from(&user, payload.len() as u64);
+                        }
+                        let mut flags = proto_flags;
+                        if quickack {
+                            flags |= RPC_FLAG_QUICKACK;
+                        }
+                        if payload.len() >= 8 && payload[..8].iter().all(|b| *b == 0) {
+                            flags |= RPC_FLAG_NOT_ENCRYPTED;
+                        }
+                        let payload_permit = match acquire_c2me_payload_permit(
+                            &c2me_byte_semaphore,
+                            payload.len(),
+                            c2me_send_timeout,
+                            stats.as_ref(),
+                        )
+                        .await
+                        {
+                            Ok(permit) => permit,
+                            Err(e) => {
+                                main_result = Err(e);
+                                break;
+                            }
+                        };
+                        // Keep client read loop lightweight: route heavy ME send path via a dedicated task.
+                        if enqueue_c2me_command_in(
+                            shared.as_ref(),
+                            &c2me_tx,
+                            C2MeCommand::Data {
+                                payload,
+                                flags,
+                                _permit: payload_permit,
+                            },
+                            c2me_send_timeout,
+                            stats.as_ref(),
+                        )
+                        .await
+                            .is_err()
+                        {
+                            main_result = Err(ProxyError::Proxy("ME sender channel closed".into()));
+                            break;
+                        }
+                    }
+                    Ok(None) => {
+                        debug!(conn_id, "Client EOF");
+                        client_closed = true;
+                        let _ = enqueue_c2me_command_in(
+                            shared.as_ref(),
+                            &c2me_tx,
+                            C2MeCommand::Close,
+                            c2me_send_timeout,
+                            stats.as_ref(),
+                        )
+                        .await;
+                        break;
+                    }
+                    Err(e) => {
+                        main_result = Err(e);
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
+    drop(c2me_tx);
+    let c2me_result = match timeout(ME_CHILD_JOIN_TIMEOUT, &mut c2me_sender).await {
+        Ok(joined) => {
+            joined.unwrap_or_else(|e| Err(ProxyError::Proxy(format!("ME sender join error: {e}"))))
+        }
+        Err(_) => {
+            stats.increment_me_child_join_timeout_total();
+            stats.increment_me_child_abort_total();
+            c2me_sender.abort();
+            Err(ProxyError::Proxy("ME sender join timeout".into()))
+        }
+    };
+
+    flow_cancel.cancel();
+    let _ = stop_tx.send(());
+    let mut writer_result = match timeout(ME_CHILD_JOIN_TIMEOUT, &mut me_writer).await {
+        Ok(joined) => {
+            joined.unwrap_or_else(|e| Err(ProxyError::Proxy(format!("ME writer join error: {e}"))))
+        }
+        Err(_) => {
+            stats.increment_me_child_join_timeout_total();
+            stats.increment_me_child_abort_total();
+            me_writer.abort();
+            Err(ProxyError::Proxy("ME writer join timeout".into()))
+        }
+    };
+
+    // When client closes, but ME channel stopped as unregistered - it isnt error
+    if client_closed && matches!(writer_result, Err(ProxyError::MiddleConnectionLost)) {
+        writer_result = Ok(());
+    }
+
+    let result = match (main_result, c2me_result, writer_result) {
+        (Ok(()), Ok(()), Ok(())) => Ok(()),
+        (Err(e), _, _) => Err(e),
+        (_, Err(e), _) => Err(e),
+        (_, _, Err(e)) => Err(e),
+    };
+
+    debug!(
+        user = %user,
+        conn_id,
+        trace_id = format_args!("0x{:016x}", trace_id),
+        duration_ms = forensics.started_at.elapsed().as_millis() as u64,
+        bytes_c2me = forensics.bytes_c2me,
+        bytes_me2c = forensics.bytes_me2c.load(Ordering::Relaxed),
+        frames_ok = frame_counter,
+        "ME relay cleanup"
+    );
+
+    let close_reason = classify_conntrack_close_reason(&result);
+    let publish_result = shared.publish_conntrack_close_event(ConntrackCloseEvent {
+        src: peer,
+        dst: local_addr,
+        reason: close_reason,
+    });
+    if !matches!(
+        publish_result,
+        ConntrackClosePublishResult::Sent | ConntrackClosePublishResult::Disabled
+    ) {
+        stats.increment_conntrack_close_event_drop_total();
+    }
+
+    clear_relay_idle_candidate_in(shared.as_ref(), conn_id);
+    me_pool.registry().unregister(conn_id).await;
+    buffer_pool.trim_to(buffer_pool.max_buffers().min(64));
+    let pool_snapshot = buffer_pool.stats();
+    stats.set_buffer_pool_gauges(
+        pool_snapshot.pooled,
+        pool_snapshot.allocated,
+        pool_snapshot.allocated.saturating_sub(pool_snapshot.pooled),
+    );
+    result
+}
+
+fn classify_conntrack_close_reason(result: &Result<()>) -> ConntrackCloseReason {
+    match result {
+        Ok(()) => ConntrackCloseReason::NormalEof,
+        Err(ProxyError::Io(error)) if matches!(error.kind(), std::io::ErrorKind::TimedOut) => {
+            ConntrackCloseReason::Timeout
+        }
+        Err(ProxyError::Io(error))
+            if matches!(
+                error.kind(),
+                std::io::ErrorKind::ConnectionReset
+                    | std::io::ErrorKind::ConnectionAborted
+                    | std::io::ErrorKind::BrokenPipe
+                    | std::io::ErrorKind::NotConnected
+                    | std::io::ErrorKind::UnexpectedEof
+            ) =>
+        {
+            ConntrackCloseReason::Reset
+        }
+        Err(ProxyError::Proxy(message))
+            if message.contains("pressure") || message.contains("evicted") =>
+        {
+            ConntrackCloseReason::Pressure
+        }
+        Err(_) => ConntrackCloseReason::Other,
+    }
+}

src/proxy/relay.rs

@@ -52,18 +52,15 @@
 //! - `SharedCounters` (atomics) let the watchdog read stats without locking
 
 use crate::error::{ProxyError, Result};
-use crate::proxy::traffic_limiter::{RateDirection, TrafficLease, next_refill_delay};
-use crate::stats::{Stats, UserStats};
+use crate::proxy::traffic_limiter::TrafficLease;
+use crate::stats::Stats;
 use crate::stream::BufferPool;
-use std::io;
-use std::pin::Pin;
 use std::sync::Arc;
-use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
-use std::task::{Context, Poll};
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
-use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt, ReadBuf, copy_bidirectional_with_sizes};
-use tokio::time::{Instant, Sleep};
-use tracing::{debug, trace, warn};
+use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt, copy_bidirectional_with_sizes};
+use tokio::time::Instant;
+use tracing::{debug, warn};
 
 // ============= Constants =============
 
@@ -85,704 +82,11 @@ fn watchdog_delta(current: u64, previous: u64) -> u64 {
     current.saturating_sub(previous)
 }
 
-// ============= CombinedStream =============
-
-/// Combines separate read and write halves into a single bidirectional stream.
-///
-/// `copy_bidirectional` requires `AsyncRead + AsyncWrite` on each side,
-/// but the handshake layer produces split reader/writer pairs
-/// (e.g. `CryptoReader<FakeTlsReader<OwnedReadHalf>>` + `CryptoWriter<...>`).
-///
-/// This wrapper reunifies them with zero overhead — each trait method
-/// delegates directly to the corresponding half. No buffering, no copies.
-///
-/// Safety: `poll_read` only touches `reader`, `poll_write` only touches `writer`,
-/// so there's no aliasing even though both are called on the same `&mut self`.
-struct CombinedStream<R, W> {
-    reader: R,
-    writer: W,
-}
-
-impl<R, W> CombinedStream<R, W> {
-    fn new(reader: R, writer: W) -> Self {
-        Self { reader, writer }
-    }
-}
-
-impl<R: AsyncRead + Unpin, W: Unpin> AsyncRead for CombinedStream<R, W> {
-    #[inline]
-    fn poll_read(
-        self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-        buf: &mut ReadBuf<'_>,
-    ) -> Poll<io::Result<()>> {
-        Pin::new(&mut self.get_mut().reader).poll_read(cx, buf)
-    }
-}
-
-impl<R: Unpin, W: AsyncWrite + Unpin> AsyncWrite for CombinedStream<R, W> {
-    #[inline]
-    fn poll_write(
-        self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-        buf: &[u8],
-    ) -> Poll<io::Result<usize>> {
-        Pin::new(&mut self.get_mut().writer).poll_write(cx, buf)
-    }
-
-    #[inline]
-    fn poll_flush(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
-        Pin::new(&mut self.get_mut().writer).poll_flush(cx)
-    }
-
-    #[inline]
-    fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
-        Pin::new(&mut self.get_mut().writer).poll_shutdown(cx)
-    }
-}
-
-// ============= SharedCounters =============
-
-/// Atomic counters shared between the relay (via StatsIo) and the watchdog task.
-///
-/// Using `Relaxed` ordering is sufficient because:
-/// - Counters are monotonically increasing (no ABA problem)
-/// - Slight staleness in watchdog reads is harmless (±10s check interval anyway)
-/// - No ordering dependencies between different counters
-struct SharedCounters {
-    /// Bytes read from client (C→S direction)
-    c2s_bytes: AtomicU64,
-    /// Bytes written to client (S→C direction)
-    s2c_bytes: AtomicU64,
-    /// Number of poll_read completions (≈ C→S chunks)
-    c2s_ops: AtomicU64,
-    /// Number of poll_write completions (≈ S→C chunks)
-    s2c_ops: AtomicU64,
-    /// Milliseconds since relay epoch of last I/O activity
-    last_activity_ms: AtomicU64,
-}
-
-impl SharedCounters {
-    fn new() -> Self {
-        Self {
-            c2s_bytes: AtomicU64::new(0),
-            s2c_bytes: AtomicU64::new(0),
-            c2s_ops: AtomicU64::new(0),
-            s2c_ops: AtomicU64::new(0),
-            last_activity_ms: AtomicU64::new(0),
-        }
-    }
-
-    /// Record activity at this instant.
-    #[inline]
-    fn touch(&self, now: Instant, epoch: Instant) {
-        let ms = now.duration_since(epoch).as_millis() as u64;
-        self.last_activity_ms.store(ms, Ordering::Relaxed);
-    }
-
-    /// How long since last recorded activity.
-    fn idle_duration(&self, now: Instant, epoch: Instant) -> Duration {
-        let last_ms = self.last_activity_ms.load(Ordering::Relaxed);
-        let now_ms = now.duration_since(epoch).as_millis() as u64;
-        Duration::from_millis(now_ms.saturating_sub(last_ms))
-    }
-}
-
-// ============= StatsIo =============
-
-/// Transparent I/O wrapper that tracks per-user statistics and activity.
-///
-/// Wraps the **client** side of the relay. Direction mapping:
-///
-/// | poll method  | direction | stats updated                        |
-/// |-------------|-----------|--------------------------------------|
-/// | `poll_read`  | C→S       | `octets_from`, `msgs_from`, counters |
-/// | `poll_write` | S→C       | `octets_to`, `msgs_to`, counters     |
-///
-/// Both update the shared activity timestamp for the watchdog.
-///
-/// Note on message counts: the original code counted one `read()`/`write_all()`
-/// as one "message". Here we count `poll_read`/`poll_write` completions instead.
-/// Byte counts are identical; op counts may differ slightly due to different
-/// internal buffering in `copy_bidirectional`. This is fine for monitoring.
-struct StatsIo<S> {
-    inner: S,
-    counters: Arc<SharedCounters>,
-    stats: Arc<Stats>,
-    user: String,
-    user_stats: Arc<UserStats>,
-    traffic_lease: Option<Arc<TrafficLease>>,
-    c2s_rate_debt_bytes: u64,
-    c2s_wait: RateWaitState,
-    s2c_wait: RateWaitState,
-    quota_wait: RateWaitState,
-    quota_limit: Option<u64>,
-    quota_exceeded: Arc<AtomicBool>,
-    quota_bytes_since_check: u64,
-    epoch: Instant,
-}
-
-#[derive(Default)]
-struct RateWaitState {
-    sleep: Option<Pin<Box<Sleep>>>,
-    started_at: Option<Instant>,
-    blocked_user: bool,
-    blocked_cidr: bool,
-}
-
-impl<S> StatsIo<S> {
-    #[cfg(test)]
-    fn new(
-        inner: S,
-        counters: Arc<SharedCounters>,
-        stats: Arc<Stats>,
-        user: String,
-        quota_limit: Option<u64>,
-        quota_exceeded: Arc<AtomicBool>,
-        epoch: Instant,
-    ) -> Self {
-        Self::new_with_traffic_lease(
-            inner,
-            counters,
-            stats,
-            user,
-            None,
-            quota_limit,
-            quota_exceeded,
-            epoch,
-        )
-    }
-
-    fn new_with_traffic_lease(
-        inner: S,
-        counters: Arc<SharedCounters>,
-        stats: Arc<Stats>,
-        user: String,
-        traffic_lease: Option<Arc<TrafficLease>>,
-        quota_limit: Option<u64>,
-        quota_exceeded: Arc<AtomicBool>,
-        epoch: Instant,
-    ) -> Self {
-        // Mark initial activity so the watchdog doesn't fire before data flows
-        counters.touch(Instant::now(), epoch);
-        let user_stats = stats.get_or_create_user_stats_handle(&user);
-        Self {
-            inner,
-            counters,
-            stats,
-            user,
-            user_stats,
-            traffic_lease,
-            c2s_rate_debt_bytes: 0,
-            c2s_wait: RateWaitState::default(),
-            s2c_wait: RateWaitState::default(),
-            quota_wait: RateWaitState::default(),
-            quota_limit,
-            quota_exceeded,
-            quota_bytes_since_check: 0,
-            epoch,
-        }
-    }
-
-    fn record_wait(
-        wait: &mut RateWaitState,
-        lease: Option<&Arc<TrafficLease>>,
-        direction: RateDirection,
-    ) {
-        let Some(started_at) = wait.started_at.take() else {
-            return;
-        };
-        let wait_ms = started_at.elapsed().as_millis().min(u128::from(u64::MAX)) as u64;
-        if let Some(lease) = lease {
-            lease.observe_wait_ms(direction, wait.blocked_user, wait.blocked_cidr, wait_ms);
-        }
-        wait.blocked_user = false;
-        wait.blocked_cidr = false;
-    }
-
-    fn arm_wait(wait: &mut RateWaitState, blocked_user: bool, blocked_cidr: bool) {
-        if wait.sleep.is_none() {
-            wait.sleep = Some(Box::pin(tokio::time::sleep(next_refill_delay())));
-            wait.started_at = Some(Instant::now());
-        }
-        wait.blocked_user |= blocked_user;
-        wait.blocked_cidr |= blocked_cidr;
-    }
-
-    fn poll_wait(
-        wait: &mut RateWaitState,
-        cx: &mut Context<'_>,
-        lease: Option<&Arc<TrafficLease>>,
-        direction: RateDirection,
-    ) -> Poll<()> {
-        let Some(sleep) = wait.sleep.as_mut() else {
-            return Poll::Ready(());
-        };
-        if sleep.as_mut().poll(cx).is_pending() {
-            return Poll::Pending;
-        }
-        wait.sleep = None;
-        Self::record_wait(wait, lease, direction);
-        Poll::Ready(())
-    }
-
-    fn settle_c2s_rate_debt(&mut self, cx: &mut Context<'_>) -> Poll<()> {
-        let Some(lease) = self.traffic_lease.as_ref() else {
-            self.c2s_rate_debt_bytes = 0;
-            return Poll::Ready(());
-        };
-
-        while self.c2s_rate_debt_bytes > 0 {
-            let consume = lease.try_consume(RateDirection::Up, self.c2s_rate_debt_bytes);
-            if consume.granted > 0 {
-                self.c2s_rate_debt_bytes = self.c2s_rate_debt_bytes.saturating_sub(consume.granted);
-                continue;
-            }
-            Self::arm_wait(
-                &mut self.c2s_wait,
-                consume.blocked_user,
-                consume.blocked_cidr,
-            );
-            if Self::poll_wait(&mut self.c2s_wait, cx, Some(lease), RateDirection::Up).is_pending()
-            {
-                return Poll::Pending;
-            }
-        }
-
-        if Self::poll_wait(&mut self.c2s_wait, cx, Some(lease), RateDirection::Up).is_pending() {
-            return Poll::Pending;
-        }
-
-        Poll::Ready(())
-    }
-
-    fn arm_quota_wait(&mut self, cx: &mut Context<'_>) -> Poll<()> {
-        Self::arm_wait(&mut self.quota_wait, false, false);
-        Self::poll_wait(&mut self.quota_wait, cx, None, RateDirection::Up)
-    }
-}
-
-#[derive(Debug)]
-struct QuotaIoSentinel;
-
-impl std::fmt::Display for QuotaIoSentinel {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        f.write_str("user data quota exceeded")
-    }
-}
-
-impl std::error::Error for QuotaIoSentinel {}
-
-fn quota_io_error() -> io::Error {
-    io::Error::new(io::ErrorKind::PermissionDenied, QuotaIoSentinel)
-}
-
-fn is_quota_io_error(err: &io::Error) -> bool {
-    err.kind() == io::ErrorKind::PermissionDenied
-        && err
-            .get_ref()
-            .and_then(|source| source.downcast_ref::<QuotaIoSentinel>())
-            .is_some()
-}
-
-const QUOTA_NEAR_LIMIT_BYTES: u64 = 64 * 1024;
-const QUOTA_LARGE_CHARGE_BYTES: u64 = 16 * 1024;
-const QUOTA_ADAPTIVE_INTERVAL_MIN_BYTES: u64 = 4 * 1024;
-const QUOTA_ADAPTIVE_INTERVAL_MAX_BYTES: u64 = 64 * 1024;
-const QUOTA_RESERVE_SPIN_RETRIES: usize = 64;
-const QUOTA_RESERVE_MAX_ROUNDS: usize = 8;
-
-#[inline]
-fn quota_adaptive_interval_bytes(remaining_before: u64) -> u64 {
-    remaining_before.saturating_div(2).clamp(
-        QUOTA_ADAPTIVE_INTERVAL_MIN_BYTES,
-        QUOTA_ADAPTIVE_INTERVAL_MAX_BYTES,
-    )
-}
-
-#[inline]
-fn should_immediate_quota_check(remaining_before: u64, charge_bytes: u64) -> bool {
-    remaining_before <= QUOTA_NEAR_LIMIT_BYTES || charge_bytes >= QUOTA_LARGE_CHARGE_BYTES
-}
-
-fn refund_reserved_quota_bytes(user_stats: &UserStats, reserved_bytes: u64) {
-    if reserved_bytes == 0 {
-        return;
-    }
-    let mut current = user_stats.quota_used.load(Ordering::Relaxed);
-    loop {
-        let next = current.saturating_sub(reserved_bytes);
-        match user_stats.quota_used.compare_exchange_weak(
-            current,
-            next,
-            Ordering::Relaxed,
-            Ordering::Relaxed,
-        ) {
-            Ok(_) => return,
-            Err(observed) => current = observed,
-        }
-    }
-}
-
-impl<S: AsyncRead + Unpin> AsyncRead for StatsIo<S> {
-    fn poll_read(
-        self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-        buf: &mut ReadBuf<'_>,
-    ) -> Poll<io::Result<()>> {
-        let this = self.get_mut();
-        if this.quota_exceeded.load(Ordering::Acquire) {
-            return Poll::Ready(Err(quota_io_error()));
-        }
-        if this.settle_c2s_rate_debt(cx).is_pending() {
-            return Poll::Pending;
-        }
-        if buf.remaining() == 0 {
-            return Pin::new(&mut this.inner).poll_read(cx, buf);
-        }
-
-        let mut remaining_before = None;
-        let mut reserved_read_bytes = 0u64;
-        let mut read_limit = buf.remaining();
-        if let Some(limit) = this.quota_limit {
-            let used_before = this.user_stats.quota_used();
-            let remaining = limit.saturating_sub(used_before);
-            if remaining == 0 {
-                this.quota_exceeded.store(true, Ordering::Release);
-                return Poll::Ready(Err(quota_io_error()));
-            }
-            remaining_before = Some(remaining);
-            read_limit = read_limit.min(remaining as usize);
-            if read_limit == 0 {
-                this.quota_exceeded.store(true, Ordering::Release);
-                return Poll::Ready(Err(quota_io_error()));
-            }
-
-            let desired = read_limit as u64;
-            let mut reserve_rounds = 0usize;
-            while reserved_read_bytes == 0 {
-                for _ in 0..QUOTA_RESERVE_SPIN_RETRIES {
-                    match this.user_stats.quota_try_reserve(desired, limit) {
-                        Ok(_) => {
-                            reserved_read_bytes = desired;
-                            break;
-                        }
-                        Err(crate::stats::QuotaReserveError::LimitExceeded) => {
-                            this.quota_exceeded.store(true, Ordering::Release);
-                            return Poll::Ready(Err(quota_io_error()));
-                        }
-                        Err(crate::stats::QuotaReserveError::Contended) => {
-                            this.stats.increment_quota_contention_total();
-                        }
-                    }
-                }
-
-                if reserved_read_bytes == 0 {
-                    reserve_rounds = reserve_rounds.saturating_add(1);
-                    if reserve_rounds >= QUOTA_RESERVE_MAX_ROUNDS {
-                        this.stats.increment_quota_contention_timeout_total();
-                        if this.arm_quota_wait(cx).is_pending() {
-                            return Poll::Pending;
-                        }
-                        reserve_rounds = 0;
-                    }
-                }
-            }
-        }
-
-        let limited_read = read_limit < buf.remaining();
-        let read_result = if limited_read {
-            let mut limited_buf = ReadBuf::new(buf.initialize_unfilled_to(read_limit));
-            match Pin::new(&mut this.inner).poll_read(cx, &mut limited_buf) {
-                Poll::Ready(Ok(())) => {
-                    let n = limited_buf.filled().len();
-                    buf.advance(n);
-                    Poll::Ready(Ok(n))
-                }
-                Poll::Ready(Err(err)) => Poll::Ready(Err(err)),
-                Poll::Pending => Poll::Pending,
-            }
-        } else {
-            let before = buf.filled().len();
-            match Pin::new(&mut this.inner).poll_read(cx, buf) {
-                Poll::Ready(Ok(())) => {
-                    let n = buf.filled().len() - before;
-                    Poll::Ready(Ok(n))
-                }
-                Poll::Ready(Err(err)) => Poll::Ready(Err(err)),
-                Poll::Pending => Poll::Pending,
-            }
-        };
-
-        match read_result {
-            Poll::Ready(Ok(n)) => {
-                if reserved_read_bytes > n as u64 {
-                    let refund_bytes = reserved_read_bytes - n as u64;
-                    refund_reserved_quota_bytes(this.user_stats.as_ref(), refund_bytes);
-                    this.stats.add_quota_refund_bytes_total(refund_bytes);
-                }
-                if n > 0 {
-                    let n_to_charge = n as u64;
-
-                    if let Some(remaining) = remaining_before {
-                        if should_immediate_quota_check(remaining, n_to_charge) {
-                            this.quota_bytes_since_check = 0;
-                        } else {
-                            this.quota_bytes_since_check =
-                                this.quota_bytes_since_check.saturating_add(n_to_charge);
-                            let interval = quota_adaptive_interval_bytes(remaining);
-                            if this.quota_bytes_since_check >= interval {
-                                this.quota_bytes_since_check = 0;
-                            }
-                        }
-                    }
-                    if let Some(limit) = this.quota_limit
-                        && this.user_stats.quota_used() >= limit
-                    {
-                        this.quota_exceeded.store(true, Ordering::Release);
-                    }
-
-                    // C→S: client sent data
-                    this.counters
-                        .c2s_bytes
-                        .fetch_add(n_to_charge, Ordering::Relaxed);
-                    this.counters.c2s_ops.fetch_add(1, Ordering::Relaxed);
-                    this.counters.touch(Instant::now(), this.epoch);
-
-                    this.stats
-                        .add_user_octets_from_handle(this.user_stats.as_ref(), n_to_charge);
-                    this.stats
-                        .increment_user_msgs_from_handle(this.user_stats.as_ref());
-                    if this.traffic_lease.is_some() {
-                        this.c2s_rate_debt_bytes =
-                            this.c2s_rate_debt_bytes.saturating_add(n_to_charge);
-                        let _ = this.settle_c2s_rate_debt(cx);
-                    }
-
-                    trace!(user = %this.user, bytes = n, "C->S");
-                }
-                Poll::Ready(Ok(()))
-            }
-            Poll::Pending => {
-                if reserved_read_bytes > 0 {
-                    refund_reserved_quota_bytes(this.user_stats.as_ref(), reserved_read_bytes);
-                    this.stats.add_quota_refund_bytes_total(reserved_read_bytes);
-                }
-                Poll::Pending
-            }
-            Poll::Ready(Err(err)) => {
-                if reserved_read_bytes > 0 {
-                    refund_reserved_quota_bytes(this.user_stats.as_ref(), reserved_read_bytes);
-                    this.stats.add_quota_refund_bytes_total(reserved_read_bytes);
-                }
-                Poll::Ready(Err(err))
-            }
-        }
-    }
-}
-
-impl<S: AsyncWrite + Unpin> AsyncWrite for StatsIo<S> {
-    fn poll_write(
-        self: Pin<&mut Self>,
-        cx: &mut Context<'_>,
-        buf: &[u8],
-    ) -> Poll<io::Result<usize>> {
-        let this = self.get_mut();
-        if this.quota_exceeded.load(Ordering::Acquire) {
-            return Poll::Ready(Err(quota_io_error()));
-        }
-
-        let mut shaper_reserved_bytes = 0u64;
-        let mut write_buf = buf;
-        if let Some(lease) = this.traffic_lease.as_ref() {
-            if !buf.is_empty() {
-                loop {
-                    let consume = lease.try_consume(RateDirection::Down, buf.len() as u64);
-                    if consume.granted > 0 {
-                        shaper_reserved_bytes = consume.granted;
-                        if consume.granted < buf.len() as u64 {
-                            write_buf = &buf[..consume.granted as usize];
-                        }
-                        let _ = Self::poll_wait(
-                            &mut this.s2c_wait,
-                            cx,
-                            Some(lease),
-                            RateDirection::Down,
-                        );
-                        break;
-                    }
-
-                    Self::arm_wait(
-                        &mut this.s2c_wait,
-                        consume.blocked_user,
-                        consume.blocked_cidr,
-                    );
-                    if Self::poll_wait(&mut this.s2c_wait, cx, Some(lease), RateDirection::Down)
-                        .is_pending()
-                    {
-                        return Poll::Pending;
-                    }
-                }
-            } else {
-                let _ = Self::poll_wait(&mut this.s2c_wait, cx, Some(lease), RateDirection::Down);
-            }
-        }
-
-        let mut remaining_before = None;
-        let mut reserved_bytes = 0u64;
-        if let Some(limit) = this.quota_limit {
-            if !write_buf.is_empty() {
-                let mut reserve_rounds = 0usize;
-                while reserved_bytes == 0 {
-                    let used_before = this.user_stats.quota_used();
-                    let remaining = limit.saturating_sub(used_before);
-                    if remaining == 0 {
-                        if let Some(lease) = this.traffic_lease.as_ref() {
-                            lease.refund(RateDirection::Down, shaper_reserved_bytes);
-                        }
-                        this.quota_exceeded.store(true, Ordering::Release);
-                        return Poll::Ready(Err(quota_io_error()));
-                    }
-                    remaining_before = Some(remaining);
-
-                    let desired = remaining.min(write_buf.len() as u64);
-                    let mut saw_contention = false;
-                    for _ in 0..QUOTA_RESERVE_SPIN_RETRIES {
-                        match this.user_stats.quota_try_reserve(desired, limit) {
-                            Ok(_) => {
-                                reserved_bytes = desired;
-                                write_buf = &write_buf[..desired as usize];
-                                break;
-                            }
-                            Err(crate::stats::QuotaReserveError::LimitExceeded) => {
-                                break;
-                            }
-                            Err(crate::stats::QuotaReserveError::Contended) => {
-                                this.stats.increment_quota_contention_total();
-                                saw_contention = true;
-                            }
-                        }
-                    }
-
-                    if reserved_bytes == 0 {
-                        reserve_rounds = reserve_rounds.saturating_add(1);
-                        if reserve_rounds >= QUOTA_RESERVE_MAX_ROUNDS {
-                            this.stats.increment_quota_contention_timeout_total();
-                            if let Some(lease) = this.traffic_lease.as_ref() {
-                                lease.refund(RateDirection::Down, shaper_reserved_bytes);
-                            }
-                            let _ = this.arm_quota_wait(cx);
-                            return Poll::Pending;
-                        } else if saw_contention {
-                            std::hint::spin_loop();
-                        }
-                    }
-                }
-            } else {
-                let used_before = this.user_stats.quota_used();
-                let remaining = limit.saturating_sub(used_before);
-                if remaining == 0 {
-                    if let Some(lease) = this.traffic_lease.as_ref() {
-                        lease.refund(RateDirection::Down, shaper_reserved_bytes);
-                    }
-                    this.quota_exceeded.store(true, Ordering::Release);
-                    return Poll::Ready(Err(quota_io_error()));
-                }
-                remaining_before = Some(remaining);
-            }
-        }
-
-        match Pin::new(&mut this.inner).poll_write(cx, write_buf) {
-            Poll::Ready(Ok(n)) => {
-                if reserved_bytes > n as u64 {
-                    let refund_bytes = reserved_bytes - n as u64;
-                    refund_reserved_quota_bytes(this.user_stats.as_ref(), refund_bytes);
-                    this.stats.add_quota_refund_bytes_total(refund_bytes);
-                }
-                if shaper_reserved_bytes > n as u64
-                    && let Some(lease) = this.traffic_lease.as_ref()
-                {
-                    lease.refund(RateDirection::Down, shaper_reserved_bytes - n as u64);
-                }
-                if n > 0 {
-                    if let Some(lease) = this.traffic_lease.as_ref() {
-                        Self::record_wait(&mut this.s2c_wait, Some(lease), RateDirection::Down);
-                    }
-                    let n_to_charge = n as u64;
-
-                    // S→C: data written to client
-                    this.counters
-                        .s2c_bytes
-                        .fetch_add(n_to_charge, Ordering::Relaxed);
-                    this.counters.s2c_ops.fetch_add(1, Ordering::Relaxed);
-                    this.counters.touch(Instant::now(), this.epoch);
-
-                    this.stats
-                        .add_user_octets_to_handle(this.user_stats.as_ref(), n_to_charge);
-                    this.stats
-                        .increment_user_msgs_to_handle(this.user_stats.as_ref());
-
-                    if let (Some(limit), Some(remaining)) = (this.quota_limit, remaining_before) {
-                        if should_immediate_quota_check(remaining, n_to_charge) {
-                            this.quota_bytes_since_check = 0;
-                            if this.user_stats.quota_used() >= limit {
-                                this.quota_exceeded.store(true, Ordering::Release);
-                            }
-                        } else {
-                            this.quota_bytes_since_check =
-                                this.quota_bytes_since_check.saturating_add(n_to_charge);
-                            let interval = quota_adaptive_interval_bytes(remaining);
-                            if this.quota_bytes_since_check >= interval {
-                                this.quota_bytes_since_check = 0;
-                                if this.user_stats.quota_used() >= limit {
-                                    this.quota_exceeded.store(true, Ordering::Release);
-                                }
-                            }
-                        }
-                    }
-
-                    trace!(user = %this.user, bytes = n, "S->C");
-                }
-                Poll::Ready(Ok(n))
-            }
-            Poll::Ready(Err(err)) => {
-                if reserved_bytes > 0 {
-                    refund_reserved_quota_bytes(this.user_stats.as_ref(), reserved_bytes);
-                    this.stats.add_quota_refund_bytes_total(reserved_bytes);
-                }
-                if shaper_reserved_bytes > 0
-                    && let Some(lease) = this.traffic_lease.as_ref()
-                {
-                    lease.refund(RateDirection::Down, shaper_reserved_bytes);
-                }
-                Poll::Ready(Err(err))
-            }
-            Poll::Pending => {
-                if reserved_bytes > 0 {
-                    refund_reserved_quota_bytes(this.user_stats.as_ref(), reserved_bytes);
-                    this.stats.add_quota_refund_bytes_total(reserved_bytes);
-                }
-                if shaper_reserved_bytes > 0
-                    && let Some(lease) = this.traffic_lease.as_ref()
-                {
-                    lease.refund(RateDirection::Down, shaper_reserved_bytes);
-                }
-                Poll::Pending
-            }
-        }
-    }
-
-    #[inline]
-    fn poll_flush(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
-        Pin::new(&mut self.get_mut().inner).poll_flush(cx)
-    }
-
-    #[inline]
-    fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
-        Pin::new(&mut self.get_mut().inner).poll_shutdown(cx)
-    }
-}
+mod io;
 
+use self::io::{CombinedStream, SharedCounters, StatsIo, is_quota_io_error};
+#[cfg(test)]
+use self::io::{quota_adaptive_interval_bytes, should_immediate_quota_check};
 // ============= Relay =============
 
 /// Relay data bidirectionally between client and server.

src/proxy/relay/io.rs

@@ -0,0 +1,551 @@
+use crate::proxy::traffic_limiter::{RateDirection, TrafficLease, next_refill_delay};
+use crate::stats::{Stats, UserStats};
+use std::io;
+use std::pin::Pin;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::task::{Context, Poll};
+use tokio::io::{AsyncRead, AsyncWrite, ReadBuf};
+use tokio::time::{Instant, Sleep};
+use tracing::trace;
+
+mod combined;
+mod counters;
+mod quota;
+
+pub(super) use self::combined::CombinedStream;
+pub(super) use self::counters::SharedCounters;
+pub(super) use self::quota::is_quota_io_error;
+use self::quota::{
+    QUOTA_RESERVE_MAX_ROUNDS, QUOTA_RESERVE_SPIN_RETRIES, quota_io_error,
+    refund_reserved_quota_bytes,
+};
+pub(super) use self::quota::{quota_adaptive_interval_bytes, should_immediate_quota_check};
+
+/// Transparent I/O wrapper that tracks per-user statistics and activity.
+///
+/// Wraps the **client** side of the relay. Direction mapping:
+///
+/// | poll method  | direction | stats updated                        |
+/// |-------------|-----------|--------------------------------------|
+/// | `poll_read`  | C→S       | `octets_from`, `msgs_from`, counters |
+/// | `poll_write` | S→C       | `octets_to`, `msgs_to`, counters     |
+///
+/// Both update the shared activity timestamp for the watchdog.
+///
+/// Note on message counts: the original code counted one `read()`/`write_all()`
+/// as one "message". Here we count `poll_read`/`poll_write` completions instead.
+/// Byte counts are identical; op counts may differ slightly due to different
+/// internal buffering in `copy_bidirectional`. This is fine for monitoring.
+pub(super) struct StatsIo<S> {
+    inner: S,
+    counters: Arc<SharedCounters>,
+    stats: Arc<Stats>,
+    user: String,
+    user_stats: Arc<UserStats>,
+    traffic_lease: Option<Arc<TrafficLease>>,
+    c2s_rate_debt_bytes: u64,
+    c2s_wait: RateWaitState,
+    s2c_wait: RateWaitState,
+    quota_wait: RateWaitState,
+    quota_limit: Option<u64>,
+    quota_exceeded: Arc<AtomicBool>,
+    pub(super) quota_bytes_since_check: u64,
+    epoch: Instant,
+}
+
+#[derive(Default)]
+struct RateWaitState {
+    sleep: Option<Pin<Box<Sleep>>>,
+    started_at: Option<Instant>,
+    blocked_user: bool,
+    blocked_cidr: bool,
+}
+
+impl<S> StatsIo<S> {
+    /// Creates a StatsIo wrapper without a traffic lease for relay unit tests.
+    #[cfg(test)]
+    pub(super) fn new(
+        inner: S,
+        counters: Arc<SharedCounters>,
+        stats: Arc<Stats>,
+        user: String,
+        quota_limit: Option<u64>,
+        quota_exceeded: Arc<AtomicBool>,
+        epoch: Instant,
+    ) -> Self {
+        Self::new_with_traffic_lease(
+            inner,
+            counters,
+            stats,
+            user,
+            None,
+            quota_limit,
+            quota_exceeded,
+            epoch,
+        )
+    }
+
+    pub(super) fn new_with_traffic_lease(
+        inner: S,
+        counters: Arc<SharedCounters>,
+        stats: Arc<Stats>,
+        user: String,
+        traffic_lease: Option<Arc<TrafficLease>>,
+        quota_limit: Option<u64>,
+        quota_exceeded: Arc<AtomicBool>,
+        epoch: Instant,
+    ) -> Self {
+        // Mark initial activity so the watchdog doesn't fire before data flows
+        counters.touch(Instant::now(), epoch);
+        let user_stats = stats.get_or_create_user_stats_handle(&user);
+        Self {
+            inner,
+            counters,
+            stats,
+            user,
+            user_stats,
+            traffic_lease,
+            c2s_rate_debt_bytes: 0,
+            c2s_wait: RateWaitState::default(),
+            s2c_wait: RateWaitState::default(),
+            quota_wait: RateWaitState::default(),
+            quota_limit,
+            quota_exceeded,
+            quota_bytes_since_check: 0,
+            epoch,
+        }
+    }
+
+    fn record_wait(
+        wait: &mut RateWaitState,
+        lease: Option<&Arc<TrafficLease>>,
+        direction: RateDirection,
+    ) {
+        let Some(started_at) = wait.started_at.take() else {
+            return;
+        };
+        let wait_ms = started_at.elapsed().as_millis().min(u128::from(u64::MAX)) as u64;
+        if let Some(lease) = lease {
+            lease.observe_wait_ms(direction, wait.blocked_user, wait.blocked_cidr, wait_ms);
+        }
+        wait.blocked_user = false;
+        wait.blocked_cidr = false;
+    }
+
+    fn arm_wait(wait: &mut RateWaitState, blocked_user: bool, blocked_cidr: bool) {
+        if wait.sleep.is_none() {
+            wait.sleep = Some(Box::pin(tokio::time::sleep(next_refill_delay())));
+            wait.started_at = Some(Instant::now());
+        }
+        wait.blocked_user |= blocked_user;
+        wait.blocked_cidr |= blocked_cidr;
+    }
+
+    fn poll_wait(
+        wait: &mut RateWaitState,
+        cx: &mut Context<'_>,
+        lease: Option<&Arc<TrafficLease>>,
+        direction: RateDirection,
+    ) -> Poll<()> {
+        let Some(sleep) = wait.sleep.as_mut() else {
+            return Poll::Ready(());
+        };
+        if sleep.as_mut().poll(cx).is_pending() {
+            return Poll::Pending;
+        }
+        wait.sleep = None;
+        Self::record_wait(wait, lease, direction);
+        Poll::Ready(())
+    }
+
+    fn settle_c2s_rate_debt(&mut self, cx: &mut Context<'_>) -> Poll<()> {
+        let Some(lease) = self.traffic_lease.as_ref() else {
+            self.c2s_rate_debt_bytes = 0;
+            return Poll::Ready(());
+        };
+
+        while self.c2s_rate_debt_bytes > 0 {
+            let consume = lease.try_consume(RateDirection::Up, self.c2s_rate_debt_bytes);
+            if consume.granted > 0 {
+                self.c2s_rate_debt_bytes = self.c2s_rate_debt_bytes.saturating_sub(consume.granted);
+                continue;
+            }
+            Self::arm_wait(
+                &mut self.c2s_wait,
+                consume.blocked_user,
+                consume.blocked_cidr,
+            );
+            if Self::poll_wait(&mut self.c2s_wait, cx, Some(lease), RateDirection::Up).is_pending()
+            {
+                return Poll::Pending;
+            }
+        }
+
+        if Self::poll_wait(&mut self.c2s_wait, cx, Some(lease), RateDirection::Up).is_pending() {
+            return Poll::Pending;
+        }
+
+        Poll::Ready(())
+    }
+
+    fn arm_quota_wait(&mut self, cx: &mut Context<'_>) -> Poll<()> {
+        Self::arm_wait(&mut self.quota_wait, false, false);
+        Self::poll_wait(&mut self.quota_wait, cx, None, RateDirection::Up)
+    }
+}
+
+impl<S: AsyncRead + Unpin> AsyncRead for StatsIo<S> {
+    fn poll_read(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &mut ReadBuf<'_>,
+    ) -> Poll<io::Result<()>> {
+        let this = self.get_mut();
+        if this.quota_exceeded.load(Ordering::Acquire) {
+            return Poll::Ready(Err(quota_io_error()));
+        }
+        if this.settle_c2s_rate_debt(cx).is_pending() {
+            return Poll::Pending;
+        }
+        if buf.remaining() == 0 {
+            return Pin::new(&mut this.inner).poll_read(cx, buf);
+        }
+
+        let mut remaining_before = None;
+        let mut reserved_read_bytes = 0u64;
+        let mut read_limit = buf.remaining();
+        if let Some(limit) = this.quota_limit {
+            let used_before = this.user_stats.quota_used();
+            let remaining = limit.saturating_sub(used_before);
+            if remaining == 0 {
+                this.quota_exceeded.store(true, Ordering::Release);
+                return Poll::Ready(Err(quota_io_error()));
+            }
+            remaining_before = Some(remaining);
+            read_limit = read_limit.min(remaining as usize);
+            if read_limit == 0 {
+                this.quota_exceeded.store(true, Ordering::Release);
+                return Poll::Ready(Err(quota_io_error()));
+            }
+
+            let desired = read_limit as u64;
+            let mut reserve_rounds = 0usize;
+            while reserved_read_bytes == 0 {
+                for _ in 0..QUOTA_RESERVE_SPIN_RETRIES {
+                    match this.user_stats.quota_try_reserve(desired, limit) {
+                        Ok(_) => {
+                            reserved_read_bytes = desired;
+                            break;
+                        }
+                        Err(crate::stats::QuotaReserveError::LimitExceeded) => {
+                            this.quota_exceeded.store(true, Ordering::Release);
+                            return Poll::Ready(Err(quota_io_error()));
+                        }
+                        Err(crate::stats::QuotaReserveError::Contended) => {
+                            this.stats.increment_quota_contention_total();
+                        }
+                    }
+                }
+
+                if reserved_read_bytes == 0 {
+                    reserve_rounds = reserve_rounds.saturating_add(1);
+                    if reserve_rounds >= QUOTA_RESERVE_MAX_ROUNDS {
+                        this.stats.increment_quota_contention_timeout_total();
+                        if this.arm_quota_wait(cx).is_pending() {
+                            return Poll::Pending;
+                        }
+                        reserve_rounds = 0;
+                    }
+                }
+            }
+        }
+
+        let limited_read = read_limit < buf.remaining();
+        let read_result = if limited_read {
+            let mut limited_buf = ReadBuf::new(buf.initialize_unfilled_to(read_limit));
+            match Pin::new(&mut this.inner).poll_read(cx, &mut limited_buf) {
+                Poll::Ready(Ok(())) => {
+                    let n = limited_buf.filled().len();
+                    buf.advance(n);
+                    Poll::Ready(Ok(n))
+                }
+                Poll::Ready(Err(err)) => Poll::Ready(Err(err)),
+                Poll::Pending => Poll::Pending,
+            }
+        } else {
+            let before = buf.filled().len();
+            match Pin::new(&mut this.inner).poll_read(cx, buf) {
+                Poll::Ready(Ok(())) => {
+                    let n = buf.filled().len() - before;
+                    Poll::Ready(Ok(n))
+                }
+                Poll::Ready(Err(err)) => Poll::Ready(Err(err)),
+                Poll::Pending => Poll::Pending,
+            }
+        };
+
+        match read_result {
+            Poll::Ready(Ok(n)) => {
+                if reserved_read_bytes > n as u64 {
+                    let refund_bytes = reserved_read_bytes - n as u64;
+                    refund_reserved_quota_bytes(this.user_stats.as_ref(), refund_bytes);
+                    this.stats.add_quota_refund_bytes_total(refund_bytes);
+                }
+                if n > 0 {
+                    let n_to_charge = n as u64;
+
+                    if let Some(remaining) = remaining_before {
+                        if should_immediate_quota_check(remaining, n_to_charge) {
+                            this.quota_bytes_since_check = 0;
+                        } else {
+                            this.quota_bytes_since_check =
+                                this.quota_bytes_since_check.saturating_add(n_to_charge);
+                            let interval = quota_adaptive_interval_bytes(remaining);
+                            if this.quota_bytes_since_check >= interval {
+                                this.quota_bytes_since_check = 0;
+                            }
+                        }
+                    }
+                    if let Some(limit) = this.quota_limit
+                        && this.user_stats.quota_used() >= limit
+                    {
+                        this.quota_exceeded.store(true, Ordering::Release);
+                    }
+
+                    // C→S: client sent data
+                    this.counters
+                        .c2s_bytes
+                        .fetch_add(n_to_charge, Ordering::Relaxed);
+                    this.counters.c2s_ops.fetch_add(1, Ordering::Relaxed);
+                    this.counters.touch(Instant::now(), this.epoch);
+
+                    this.stats
+                        .add_user_traffic_from_handle(this.user_stats.as_ref(), n_to_charge);
+                    if this.traffic_lease.is_some() {
+                        this.c2s_rate_debt_bytes =
+                            this.c2s_rate_debt_bytes.saturating_add(n_to_charge);
+                        let _ = this.settle_c2s_rate_debt(cx);
+                    }
+
+                    trace!(user = %this.user, bytes = n, "C->S");
+                }
+                Poll::Ready(Ok(()))
+            }
+            Poll::Pending => {
+                if reserved_read_bytes > 0 {
+                    refund_reserved_quota_bytes(this.user_stats.as_ref(), reserved_read_bytes);
+                    this.stats.add_quota_refund_bytes_total(reserved_read_bytes);
+                }
+                Poll::Pending
+            }
+            Poll::Ready(Err(err)) => {
+                if reserved_read_bytes > 0 {
+                    refund_reserved_quota_bytes(this.user_stats.as_ref(), reserved_read_bytes);
+                    this.stats.add_quota_refund_bytes_total(reserved_read_bytes);
+                }
+                Poll::Ready(Err(err))
+            }
+        }
+    }
+}
+
+impl<S: AsyncWrite + Unpin> AsyncWrite for StatsIo<S> {
+    fn poll_write(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &[u8],
+    ) -> Poll<io::Result<usize>> {
+        let this = self.get_mut();
+        if this.quota_exceeded.load(Ordering::Acquire) {
+            return Poll::Ready(Err(quota_io_error()));
+        }
+
+        let mut shaper_reserved_bytes = 0u64;
+        let mut write_buf = buf;
+        if let Some(lease) = this.traffic_lease.as_ref() {
+            if !buf.is_empty() {
+                loop {
+                    let consume = lease.try_consume(RateDirection::Down, buf.len() as u64);
+                    if consume.granted > 0 {
+                        shaper_reserved_bytes = consume.granted;
+                        if consume.granted < buf.len() as u64 {
+                            write_buf = &buf[..consume.granted as usize];
+                        }
+                        let _ = Self::poll_wait(
+                            &mut this.s2c_wait,
+                            cx,
+                            Some(lease),
+                            RateDirection::Down,
+                        );
+                        break;
+                    }
+
+                    Self::arm_wait(
+                        &mut this.s2c_wait,
+                        consume.blocked_user,
+                        consume.blocked_cidr,
+                    );
+                    if Self::poll_wait(&mut this.s2c_wait, cx, Some(lease), RateDirection::Down)
+                        .is_pending()
+                    {
+                        return Poll::Pending;
+                    }
+                }
+            } else {
+                let _ = Self::poll_wait(&mut this.s2c_wait, cx, Some(lease), RateDirection::Down);
+            }
+        }
+
+        let mut remaining_before = None;
+        let mut reserved_bytes = 0u64;
+        if let Some(limit) = this.quota_limit {
+            if !write_buf.is_empty() {
+                let mut reserve_rounds = 0usize;
+                while reserved_bytes == 0 {
+                    let used_before = this.user_stats.quota_used();
+                    let remaining = limit.saturating_sub(used_before);
+                    if remaining == 0 {
+                        if let Some(lease) = this.traffic_lease.as_ref() {
+                            lease.refund(RateDirection::Down, shaper_reserved_bytes);
+                        }
+                        this.quota_exceeded.store(true, Ordering::Release);
+                        return Poll::Ready(Err(quota_io_error()));
+                    }
+                    remaining_before = Some(remaining);
+
+                    let desired = remaining.min(write_buf.len() as u64);
+                    let mut saw_contention = false;
+                    for _ in 0..QUOTA_RESERVE_SPIN_RETRIES {
+                        match this.user_stats.quota_try_reserve(desired, limit) {
+                            Ok(_) => {
+                                reserved_bytes = desired;
+                                write_buf = &write_buf[..desired as usize];
+                                break;
+                            }
+                            Err(crate::stats::QuotaReserveError::LimitExceeded) => {
+                                break;
+                            }
+                            Err(crate::stats::QuotaReserveError::Contended) => {
+                                this.stats.increment_quota_contention_total();
+                                saw_contention = true;
+                            }
+                        }
+                    }
+
+                    if reserved_bytes == 0 {
+                        reserve_rounds = reserve_rounds.saturating_add(1);
+                        if reserve_rounds >= QUOTA_RESERVE_MAX_ROUNDS {
+                            this.stats.increment_quota_contention_timeout_total();
+                            if let Some(lease) = this.traffic_lease.as_ref() {
+                                lease.refund(RateDirection::Down, shaper_reserved_bytes);
+                            }
+                            let _ = this.arm_quota_wait(cx);
+                            return Poll::Pending;
+                        } else if saw_contention {
+                            std::hint::spin_loop();
+                        }
+                    }
+                }
+            } else {
+                let used_before = this.user_stats.quota_used();
+                let remaining = limit.saturating_sub(used_before);
+                if remaining == 0 {
+                    if let Some(lease) = this.traffic_lease.as_ref() {
+                        lease.refund(RateDirection::Down, shaper_reserved_bytes);
+                    }
+                    this.quota_exceeded.store(true, Ordering::Release);
+                    return Poll::Ready(Err(quota_io_error()));
+                }
+                remaining_before = Some(remaining);
+            }
+        }
+
+        match Pin::new(&mut this.inner).poll_write(cx, write_buf) {
+            Poll::Ready(Ok(n)) => {
+                if reserved_bytes > n as u64 {
+                    let refund_bytes = reserved_bytes - n as u64;
+                    refund_reserved_quota_bytes(this.user_stats.as_ref(), refund_bytes);
+                    this.stats.add_quota_refund_bytes_total(refund_bytes);
+                }
+                if shaper_reserved_bytes > n as u64
+                    && let Some(lease) = this.traffic_lease.as_ref()
+                {
+                    lease.refund(RateDirection::Down, shaper_reserved_bytes - n as u64);
+                }
+                if n > 0 {
+                    if let Some(lease) = this.traffic_lease.as_ref() {
+                        Self::record_wait(&mut this.s2c_wait, Some(lease), RateDirection::Down);
+                    }
+                    let n_to_charge = n as u64;
+
+                    // S→C: data written to client
+                    this.counters
+                        .s2c_bytes
+                        .fetch_add(n_to_charge, Ordering::Relaxed);
+                    this.counters.s2c_ops.fetch_add(1, Ordering::Relaxed);
+                    this.counters.touch(Instant::now(), this.epoch);
+
+                    this.stats
+                        .add_user_traffic_to_handle(this.user_stats.as_ref(), n_to_charge);
+
+                    if let (Some(limit), Some(remaining)) = (this.quota_limit, remaining_before) {
+                        if should_immediate_quota_check(remaining, n_to_charge) {
+                            this.quota_bytes_since_check = 0;
+                            if this.user_stats.quota_used() >= limit {
+                                this.quota_exceeded.store(true, Ordering::Release);
+                            }
+                        } else {
+                            this.quota_bytes_since_check =
+                                this.quota_bytes_since_check.saturating_add(n_to_charge);
+                            let interval = quota_adaptive_interval_bytes(remaining);
+                            if this.quota_bytes_since_check >= interval {
+                                this.quota_bytes_since_check = 0;
+                                if this.user_stats.quota_used() >= limit {
+                                    this.quota_exceeded.store(true, Ordering::Release);
+                                }
+                            }
+                        }
+                    }
+
+                    trace!(user = %this.user, bytes = n, "S->C");
+                }
+                Poll::Ready(Ok(n))
+            }
+            Poll::Ready(Err(err)) => {
+                if reserved_bytes > 0 {
+                    refund_reserved_quota_bytes(this.user_stats.as_ref(), reserved_bytes);
+                    this.stats.add_quota_refund_bytes_total(reserved_bytes);
+                }
+                if shaper_reserved_bytes > 0
+                    && let Some(lease) = this.traffic_lease.as_ref()
+                {
+                    lease.refund(RateDirection::Down, shaper_reserved_bytes);
+                }
+                Poll::Ready(Err(err))
+            }
+            Poll::Pending => {
+                if reserved_bytes > 0 {
+                    refund_reserved_quota_bytes(this.user_stats.as_ref(), reserved_bytes);
+                    this.stats.add_quota_refund_bytes_total(reserved_bytes);
+                }
+                if shaper_reserved_bytes > 0
+                    && let Some(lease) = this.traffic_lease.as_ref()
+                {
+                    lease.refund(RateDirection::Down, shaper_reserved_bytes);
+                }
+                Poll::Pending
+            }
+        }
+    }
+
+    #[inline]
+    fn poll_flush(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        Pin::new(&mut self.get_mut().inner).poll_flush(cx)
+    }
+
+    #[inline]
+    fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        Pin::new(&mut self.get_mut().inner).poll_shutdown(cx)
+    }
+}

src/proxy/relay/io/combined.rs

@@ -0,0 +1,61 @@
+use std::io;
+use std::pin::Pin;
+use std::task::{Context, Poll};
+
+use tokio::io::{AsyncRead, AsyncWrite, ReadBuf};
+
+// ============= CombinedStream =============
+
+/// Combines separate read and write halves into a single bidirectional stream.
+///
+/// `copy_bidirectional` requires `AsyncRead + AsyncWrite` on each side,
+/// but the handshake layer produces split reader/writer pairs
+/// (e.g. `CryptoReader<FakeTlsReader<OwnedReadHalf>>` + `CryptoWriter<...>`).
+///
+/// This wrapper reunifies them with zero overhead — each trait method
+/// delegates directly to the corresponding half. No buffering, no copies.
+///
+/// Safety: `poll_read` only touches `reader`, `poll_write` only touches `writer`,
+/// so there's no aliasing even though both are called on the same `&mut self`.
+pub(in crate::proxy::relay) struct CombinedStream<R, W> {
+    reader: R,
+    writer: W,
+}
+
+impl<R, W> CombinedStream<R, W> {
+    pub(in crate::proxy::relay) fn new(reader: R, writer: W) -> Self {
+        Self { reader, writer }
+    }
+}
+
+impl<R: AsyncRead + Unpin, W: Unpin> AsyncRead for CombinedStream<R, W> {
+    #[inline]
+    fn poll_read(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &mut ReadBuf<'_>,
+    ) -> Poll<io::Result<()>> {
+        Pin::new(&mut self.get_mut().reader).poll_read(cx, buf)
+    }
+}
+
+impl<R: Unpin, W: AsyncWrite + Unpin> AsyncWrite for CombinedStream<R, W> {
+    #[inline]
+    fn poll_write(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &[u8],
+    ) -> Poll<io::Result<usize>> {
+        Pin::new(&mut self.get_mut().writer).poll_write(cx, buf)
+    }
+
+    #[inline]
+    fn poll_flush(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        Pin::new(&mut self.get_mut().writer).poll_flush(cx)
+    }
+
+    #[inline]
+    fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        Pin::new(&mut self.get_mut().writer).poll_shutdown(cx)
+    }
+}

src/proxy/relay/io/counters.rs

@@ -0,0 +1,51 @@
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::Duration;
+
+use tokio::time::Instant;
+
+// ============= SharedCounters =============
+
+/// Atomic counters shared between the relay (via StatsIo) and the watchdog task.
+///
+/// Using `Relaxed` ordering is sufficient because:
+/// - Counters are monotonically increasing (no ABA problem)
+/// - Slight staleness in watchdog reads is harmless (±10s check interval anyway)
+/// - No ordering dependencies between different counters
+pub(in crate::proxy::relay) struct SharedCounters {
+    /// Bytes read from client (C→S direction)
+    pub(in crate::proxy::relay) c2s_bytes: AtomicU64,
+    /// Bytes written to client (S→C direction)
+    pub(in crate::proxy::relay) s2c_bytes: AtomicU64,
+    /// Number of poll_read completions (≈ C→S chunks)
+    pub(in crate::proxy::relay) c2s_ops: AtomicU64,
+    /// Number of poll_write completions (≈ S→C chunks)
+    pub(in crate::proxy::relay) s2c_ops: AtomicU64,
+    /// Milliseconds since relay epoch of last I/O activity
+    last_activity_ms: AtomicU64,
+}
+
+impl SharedCounters {
+    pub(in crate::proxy::relay) fn new() -> Self {
+        Self {
+            c2s_bytes: AtomicU64::new(0),
+            s2c_bytes: AtomicU64::new(0),
+            c2s_ops: AtomicU64::new(0),
+            s2c_ops: AtomicU64::new(0),
+            last_activity_ms: AtomicU64::new(0),
+        }
+    }
+
+    /// Record activity at this instant.
+    #[inline]
+    pub(in crate::proxy::relay) fn touch(&self, now: Instant, epoch: Instant) {
+        let ms = now.duration_since(epoch).as_millis() as u64;
+        self.last_activity_ms.store(ms, Ordering::Relaxed);
+    }
+
+    /// How long since last recorded activity.
+    pub(in crate::proxy::relay) fn idle_duration(&self, now: Instant, epoch: Instant) -> Duration {
+        let last_ms = self.last_activity_ms.load(Ordering::Relaxed);
+        let now_ms = now.duration_since(epoch).as_millis() as u64;
+        Duration::from_millis(now_ms.saturating_sub(last_ms))
+    }
+}

src/proxy/relay/io/quota.rs

@@ -0,0 +1,68 @@
+use crate::stats::UserStats;
+use std::io;
+use std::sync::atomic::Ordering;
+
+#[derive(Debug)]
+struct QuotaIoSentinel;
+
+impl std::fmt::Display for QuotaIoSentinel {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("user data quota exceeded")
+    }
+}
+
+impl std::error::Error for QuotaIoSentinel {}
+
+pub(super) fn quota_io_error() -> io::Error {
+    io::Error::new(io::ErrorKind::PermissionDenied, QuotaIoSentinel)
+}
+
+pub(in crate::proxy::relay) fn is_quota_io_error(err: &io::Error) -> bool {
+    err.kind() == io::ErrorKind::PermissionDenied
+        && err
+            .get_ref()
+            .and_then(|source| source.downcast_ref::<QuotaIoSentinel>())
+            .is_some()
+}
+
+const QUOTA_NEAR_LIMIT_BYTES: u64 = 64 * 1024;
+const QUOTA_LARGE_CHARGE_BYTES: u64 = 16 * 1024;
+const QUOTA_ADAPTIVE_INTERVAL_MIN_BYTES: u64 = 4 * 1024;
+const QUOTA_ADAPTIVE_INTERVAL_MAX_BYTES: u64 = 64 * 1024;
+pub(super) const QUOTA_RESERVE_SPIN_RETRIES: usize = 64;
+pub(super) const QUOTA_RESERVE_MAX_ROUNDS: usize = 8;
+
+#[inline]
+pub(in crate::proxy::relay) fn quota_adaptive_interval_bytes(remaining_before: u64) -> u64 {
+    remaining_before.saturating_div(2).clamp(
+        QUOTA_ADAPTIVE_INTERVAL_MIN_BYTES,
+        QUOTA_ADAPTIVE_INTERVAL_MAX_BYTES,
+    )
+}
+
+#[inline]
+pub(in crate::proxy::relay) fn should_immediate_quota_check(
+    remaining_before: u64,
+    charge_bytes: u64,
+) -> bool {
+    remaining_before <= QUOTA_NEAR_LIMIT_BYTES || charge_bytes >= QUOTA_LARGE_CHARGE_BYTES
+}
+
+pub(super) fn refund_reserved_quota_bytes(user_stats: &UserStats, reserved_bytes: u64) {
+    if reserved_bytes == 0 {
+        return;
+    }
+    let mut current = user_stats.quota_used.load(Ordering::Relaxed);
+    loop {
+        let next = current.saturating_sub(reserved_bytes);
+        match user_stats.quota_used.compare_exchange_weak(
+            current,
+            next,
+            Ordering::Relaxed,
+            Ordering::Relaxed,
+        ) {
+            Ok(_) => return,
+            Err(observed) => current = observed,
+        }
+    }
+}

src/proxy/tests/relay_baseline_invariant_tests.rs

@@ -3,7 +3,9 @@ use crate::error::ProxyError;
 use crate::stats::Stats;
 use crate::stream::BufferPool;
 use std::io;
+use std::pin::Pin;
 use std::sync::Arc;
+use std::task::{Context, Poll};
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt, ReadBuf, duplex};
 use tokio::time::{Duration, timeout};
 

src/stats/core_counters.rs

@@ -0,0 +1,266 @@
+use super::*;
+
+impl Stats {
+    pub fn apply_telemetry_policy(&self, policy: TelemetryPolicy) {
+        self.telemetry_core_enabled
+            .store(policy.core_enabled, Ordering::Relaxed);
+        self.telemetry_user_enabled
+            .store(policy.user_enabled, Ordering::Relaxed);
+        self.telemetry_me_level
+            .store(policy.me_level.as_u8(), Ordering::Relaxed);
+    }
+
+    pub fn telemetry_policy(&self) -> TelemetryPolicy {
+        TelemetryPolicy {
+            core_enabled: self.telemetry_core_enabled(),
+            user_enabled: self.telemetry_user_enabled(),
+            me_level: self.telemetry_me_level(),
+        }
+    }
+
+    pub fn increment_connects_all(&self) {
+        if self.telemetry_core_enabled() {
+            self.connects_all.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_connects_bad_with_class(&self, class: &'static str) {
+        if !self.telemetry_core_enabled() {
+            return;
+        }
+        self.connects_bad.fetch_add(1, Ordering::Relaxed);
+        let entry = self
+            .connects_bad_classes
+            .entry(class)
+            .or_insert_with(|| AtomicU64::new(0));
+        entry.fetch_add(1, Ordering::Relaxed);
+    }
+
+    pub fn increment_connects_bad(&self) {
+        self.increment_connects_bad_with_class("other");
+    }
+
+    pub fn increment_handshake_failure_class(&self, class: &'static str) {
+        if !self.telemetry_core_enabled() {
+            return;
+        }
+        let entry = self
+            .handshake_failure_classes
+            .entry(class)
+            .or_insert_with(|| AtomicU64::new(0));
+        entry.fetch_add(1, Ordering::Relaxed);
+    }
+    pub fn increment_current_connections_direct(&self) {
+        self.current_connections_direct
+            .fetch_add(1, Ordering::Relaxed);
+    }
+    pub fn decrement_current_connections_direct(&self) {
+        Self::decrement_atomic_saturating(&self.current_connections_direct);
+    }
+    pub fn increment_current_connections_me(&self) {
+        self.current_connections_me.fetch_add(1, Ordering::Relaxed);
+    }
+    pub fn decrement_current_connections_me(&self) {
+        Self::decrement_atomic_saturating(&self.current_connections_me);
+    }
+
+    pub fn acquire_direct_connection_lease(self: &Arc<Self>) -> RouteConnectionLease {
+        self.increment_current_connections_direct();
+        RouteConnectionLease::new(self.clone(), RouteConnectionGauge::Direct)
+    }
+
+    pub fn acquire_me_connection_lease(self: &Arc<Self>) -> RouteConnectionLease {
+        self.increment_current_connections_me();
+        RouteConnectionLease::new(self.clone(), RouteConnectionGauge::Middle)
+    }
+
+    pub(super) fn decrement_route_cutover_parked_direct(&self) {
+        Self::decrement_atomic_saturating(&self.route_cutover_parked_direct_current);
+    }
+
+    pub(super) fn decrement_route_cutover_parked_middle(&self) {
+        Self::decrement_atomic_saturating(&self.route_cutover_parked_middle_current);
+    }
+
+    pub fn acquire_direct_cutover_park_lease(self: &Arc<Self>) -> RouteCutoverParkLease {
+        self.route_cutover_parked_direct_current
+            .fetch_add(1, Ordering::Relaxed);
+        self.route_cutover_parked_direct_total
+            .fetch_add(1, Ordering::Relaxed);
+        RouteCutoverParkLease::new(self.clone(), RouteCutoverParkGauge::Direct)
+    }
+
+    pub fn acquire_middle_cutover_park_lease(self: &Arc<Self>) -> RouteCutoverParkLease {
+        self.route_cutover_parked_middle_current
+            .fetch_add(1, Ordering::Relaxed);
+        self.route_cutover_parked_middle_total
+            .fetch_add(1, Ordering::Relaxed);
+        RouteCutoverParkLease::new(self.clone(), RouteCutoverParkGauge::Middle)
+    }
+    pub fn increment_handshake_timeouts(&self) {
+        if self.telemetry_core_enabled() {
+            self.handshake_timeouts.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_accept_permit_timeout_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.accept_permit_timeout_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn set_conntrack_control_enabled(&self, enabled: bool) {
+        self.conntrack_control_enabled_gauge
+            .store(enabled, Ordering::Relaxed);
+    }
+
+    pub fn set_conntrack_control_available(&self, available: bool) {
+        self.conntrack_control_available_gauge
+            .store(available, Ordering::Relaxed);
+    }
+
+    pub fn set_conntrack_pressure_active(&self, active: bool) {
+        self.conntrack_pressure_active_gauge
+            .store(active, Ordering::Relaxed);
+    }
+
+    pub fn set_conntrack_event_queue_depth(&self, depth: u64) {
+        self.conntrack_event_queue_depth_gauge
+            .store(depth, Ordering::Relaxed);
+    }
+
+    pub fn set_conntrack_rule_apply_ok(&self, ok: bool) {
+        self.conntrack_rule_apply_ok_gauge
+            .store(ok, Ordering::Relaxed);
+    }
+
+    pub fn increment_conntrack_delete_attempt_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.conntrack_delete_attempt_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_conntrack_delete_success_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.conntrack_delete_success_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_conntrack_delete_not_found_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.conntrack_delete_not_found_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_conntrack_delete_error_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.conntrack_delete_error_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_conntrack_close_event_drop_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.conntrack_close_event_drop_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_upstream_connect_attempt_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.upstream_connect_attempt_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_upstream_connect_success_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.upstream_connect_success_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_upstream_connect_fail_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.upstream_connect_fail_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_upstream_connect_failfast_hard_error_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.upstream_connect_failfast_hard_error_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn observe_upstream_connect_attempts_per_request(&self, attempts: u32) {
+        if !self.telemetry_core_enabled() {
+            return;
+        }
+        match attempts {
+            0 => {}
+            1 => {
+                self.upstream_connect_attempts_bucket_1
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            2 => {
+                self.upstream_connect_attempts_bucket_2
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            3..=4 => {
+                self.upstream_connect_attempts_bucket_3_4
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            _ => {
+                self.upstream_connect_attempts_bucket_gt_4
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn observe_upstream_connect_duration_ms(&self, duration_ms: u64, success: bool) {
+        if !self.telemetry_core_enabled() {
+            return;
+        }
+        let bucket = match duration_ms {
+            0..=100 => 0u8,
+            101..=500 => 1u8,
+            501..=1000 => 2u8,
+            _ => 3u8,
+        };
+        match (success, bucket) {
+            (true, 0) => {
+                self.upstream_connect_duration_success_bucket_le_100ms
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            (true, 1) => {
+                self.upstream_connect_duration_success_bucket_101_500ms
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            (true, 2) => {
+                self.upstream_connect_duration_success_bucket_501_1000ms
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            (true, _) => {
+                self.upstream_connect_duration_success_bucket_gt_1000ms
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            (false, 0) => {
+                self.upstream_connect_duration_fail_bucket_le_100ms
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            (false, 1) => {
+                self.upstream_connect_duration_fail_bucket_101_500ms
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            (false, 2) => {
+                self.upstream_connect_duration_fail_bucket_501_1000ms
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            (false, _) => {
+                self.upstream_connect_duration_fail_bucket_gt_1000ms
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+}

src/stats/core_getters.rs

@@ -0,0 +1,283 @@
+use super::*;
+
+impl Stats {
+    pub fn get_connects_all(&self) -> u64 {
+        self.connects_all.load(Ordering::Relaxed)
+    }
+    pub fn get_connects_bad(&self) -> u64 {
+        self.connects_bad.load(Ordering::Relaxed)
+    }
+
+    pub fn get_connects_bad_class_counts(&self) -> Vec<(String, u64)> {
+        let mut out: Vec<(String, u64)> = self
+            .connects_bad_classes
+            .iter()
+            .map(|entry| {
+                (
+                    entry.key().to_string(),
+                    entry.value().load(Ordering::Relaxed),
+                )
+            })
+            .collect();
+        out.sort_by(|a, b| a.0.cmp(&b.0));
+        out
+    }
+
+    pub fn get_handshake_failure_class_counts(&self) -> Vec<(String, u64)> {
+        let mut out: Vec<(String, u64)> = self
+            .handshake_failure_classes
+            .iter()
+            .map(|entry| {
+                (
+                    entry.key().to_string(),
+                    entry.value().load(Ordering::Relaxed),
+                )
+            })
+            .collect();
+        out.sort_by(|a, b| a.0.cmp(&b.0));
+        out
+    }
+
+    pub fn get_accept_permit_timeout_total(&self) -> u64 {
+        self.accept_permit_timeout_total.load(Ordering::Relaxed)
+    }
+    pub fn get_current_connections_direct(&self) -> u64 {
+        self.current_connections_direct.load(Ordering::Relaxed)
+    }
+    pub fn get_current_connections_me(&self) -> u64 {
+        self.current_connections_me.load(Ordering::Relaxed)
+    }
+    pub fn get_route_cutover_parked_direct_current(&self) -> u64 {
+        self.route_cutover_parked_direct_current
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_route_cutover_parked_middle_current(&self) -> u64 {
+        self.route_cutover_parked_middle_current
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_route_cutover_parked_direct_total(&self) -> u64 {
+        self.route_cutover_parked_direct_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_route_cutover_parked_middle_total(&self) -> u64 {
+        self.route_cutover_parked_middle_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_current_connections_total(&self) -> u64 {
+        self.get_current_connections_direct()
+            .saturating_add(self.get_current_connections_me())
+    }
+    pub fn get_conntrack_control_enabled(&self) -> bool {
+        self.conntrack_control_enabled_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_control_available(&self) -> bool {
+        self.conntrack_control_available_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_pressure_active(&self) -> bool {
+        self.conntrack_pressure_active_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_event_queue_depth(&self) -> u64 {
+        self.conntrack_event_queue_depth_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_rule_apply_ok(&self) -> bool {
+        self.conntrack_rule_apply_ok_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_delete_attempt_total(&self) -> u64 {
+        self.conntrack_delete_attempt_total.load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_delete_success_total(&self) -> u64 {
+        self.conntrack_delete_success_total.load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_delete_not_found_total(&self) -> u64 {
+        self.conntrack_delete_not_found_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_delete_error_total(&self) -> u64 {
+        self.conntrack_delete_error_total.load(Ordering::Relaxed)
+    }
+    pub fn get_conntrack_close_event_drop_total(&self) -> u64 {
+        self.conntrack_close_event_drop_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_keepalive_sent(&self) -> u64 {
+        self.me_keepalive_sent.load(Ordering::Relaxed)
+    }
+    pub fn get_me_keepalive_failed(&self) -> u64 {
+        self.me_keepalive_failed.load(Ordering::Relaxed)
+    }
+    pub fn get_me_keepalive_pong(&self) -> u64 {
+        self.me_keepalive_pong.load(Ordering::Relaxed)
+    }
+    pub fn get_me_keepalive_timeout(&self) -> u64 {
+        self.me_keepalive_timeout.load(Ordering::Relaxed)
+    }
+    pub fn get_me_rpc_proxy_req_signal_sent_total(&self) -> u64 {
+        self.me_rpc_proxy_req_signal_sent_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_rpc_proxy_req_signal_failed_total(&self) -> u64 {
+        self.me_rpc_proxy_req_signal_failed_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_rpc_proxy_req_signal_skipped_no_meta_total(&self) -> u64 {
+        self.me_rpc_proxy_req_signal_skipped_no_meta_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_rpc_proxy_req_signal_response_total(&self) -> u64 {
+        self.me_rpc_proxy_req_signal_response_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_rpc_proxy_req_signal_close_sent_total(&self) -> u64 {
+        self.me_rpc_proxy_req_signal_close_sent_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_reconnect_attempts(&self) -> u64 {
+        self.me_reconnect_attempts.load(Ordering::Relaxed)
+    }
+    pub fn get_me_reconnect_success(&self) -> u64 {
+        self.me_reconnect_success.load(Ordering::Relaxed)
+    }
+    pub fn get_me_handshake_reject_total(&self) -> u64 {
+        self.me_handshake_reject_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_reader_eof_total(&self) -> u64 {
+        self.me_reader_eof_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_idle_close_by_peer_total(&self) -> u64 {
+        self.me_idle_close_by_peer_total.load(Ordering::Relaxed)
+    }
+    pub fn get_relay_idle_soft_mark_total(&self) -> u64 {
+        self.relay_idle_soft_mark_total.load(Ordering::Relaxed)
+    }
+    pub fn get_relay_idle_hard_close_total(&self) -> u64 {
+        self.relay_idle_hard_close_total.load(Ordering::Relaxed)
+    }
+    pub fn get_relay_pressure_evict_total(&self) -> u64 {
+        self.relay_pressure_evict_total.load(Ordering::Relaxed)
+    }
+    pub fn get_relay_protocol_desync_close_total(&self) -> u64 {
+        self.relay_protocol_desync_close_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_crc_mismatch(&self) -> u64 {
+        self.me_crc_mismatch.load(Ordering::Relaxed)
+    }
+    pub fn get_me_seq_mismatch(&self) -> u64 {
+        self.me_seq_mismatch.load(Ordering::Relaxed)
+    }
+    pub fn get_me_endpoint_quarantine_total(&self) -> u64 {
+        self.me_endpoint_quarantine_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_endpoint_quarantine_unexpected_total(&self) -> u64 {
+        self.me_endpoint_quarantine_unexpected_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_endpoint_quarantine_draining_suppressed_total(&self) -> u64 {
+        self.me_endpoint_quarantine_draining_suppressed_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_kdf_drift_total(&self) -> u64 {
+        self.me_kdf_drift_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_kdf_port_only_drift_total(&self) -> u64 {
+        self.me_kdf_port_only_drift_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_hardswap_pending_reuse_total(&self) -> u64 {
+        self.me_hardswap_pending_reuse_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_hardswap_pending_ttl_expired_total(&self) -> u64 {
+        self.me_hardswap_pending_ttl_expired_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_single_endpoint_outage_enter_total(&self) -> u64 {
+        self.me_single_endpoint_outage_enter_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_single_endpoint_outage_exit_total(&self) -> u64 {
+        self.me_single_endpoint_outage_exit_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_single_endpoint_outage_reconnect_attempt_total(&self) -> u64 {
+        self.me_single_endpoint_outage_reconnect_attempt_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_single_endpoint_outage_reconnect_success_total(&self) -> u64 {
+        self.me_single_endpoint_outage_reconnect_success_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_single_endpoint_quarantine_bypass_total(&self) -> u64 {
+        self.me_single_endpoint_quarantine_bypass_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_single_endpoint_shadow_rotate_total(&self) -> u64 {
+        self.me_single_endpoint_shadow_rotate_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_single_endpoint_shadow_rotate_skipped_quarantine_total(&self) -> u64 {
+        self.me_single_endpoint_shadow_rotate_skipped_quarantine_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_mode_switch_total(&self) -> u64 {
+        self.me_floor_mode_switch_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_mode_switch_static_to_adaptive_total(&self) -> u64 {
+        self.me_floor_mode_switch_static_to_adaptive_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_mode_switch_adaptive_to_static_total(&self) -> u64 {
+        self.me_floor_mode_switch_adaptive_to_static_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_cpu_cores_detected_gauge(&self) -> u64 {
+        self.me_floor_cpu_cores_detected_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_cpu_cores_effective_gauge(&self) -> u64 {
+        self.me_floor_cpu_cores_effective_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_global_cap_raw_gauge(&self) -> u64 {
+        self.me_floor_global_cap_raw_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_global_cap_effective_gauge(&self) -> u64 {
+        self.me_floor_global_cap_effective_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_target_writers_total_gauge(&self) -> u64 {
+        self.me_floor_target_writers_total_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_active_cap_configured_gauge(&self) -> u64 {
+        self.me_floor_active_cap_configured_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_active_cap_effective_gauge(&self) -> u64 {
+        self.me_floor_active_cap_effective_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_warm_cap_configured_gauge(&self) -> u64 {
+        self.me_floor_warm_cap_configured_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_warm_cap_effective_gauge(&self) -> u64 {
+        self.me_floor_warm_cap_effective_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writers_active_current_gauge(&self) -> u64 {
+        self.me_writers_active_current_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_me_writers_warm_current_gauge(&self) -> u64 {
+        self.me_writers_warm_current_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_cap_block_total(&self) -> u64 {
+        self.me_floor_cap_block_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_swap_idle_total(&self) -> u64 {
+        self.me_floor_swap_idle_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_floor_swap_idle_failed_total(&self) -> u64 {
+        self.me_floor_swap_idle_failed_total.load(Ordering::Relaxed)
+    }
+}

src/stats/helpers.rs

@@ -0,0 +1,208 @@
+use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
+
+use crate::config::MeTelemetryLevel;
+
+use super::*;
+
+impl Stats {
+    pub(super) fn telemetry_me_level(&self) -> MeTelemetryLevel {
+        MeTelemetryLevel::from_u8(self.telemetry_me_level.load(Ordering::Relaxed))
+    }
+
+    pub(super) fn telemetry_core_enabled(&self) -> bool {
+        self.telemetry_core_enabled.load(Ordering::Relaxed)
+    }
+
+    pub(super) fn telemetry_user_enabled(&self) -> bool {
+        self.telemetry_user_enabled.load(Ordering::Relaxed)
+    }
+
+    pub(super) fn telemetry_me_allows_normal(&self) -> bool {
+        self.telemetry_me_level().allows_normal()
+    }
+
+    pub(super) fn telemetry_me_allows_debug(&self) -> bool {
+        self.telemetry_me_level().allows_debug()
+    }
+
+    pub(super) fn decrement_atomic_saturating(counter: &AtomicU64) {
+        let mut current = counter.load(Ordering::Relaxed);
+        loop {
+            if current == 0 {
+                break;
+            }
+            match counter.compare_exchange_weak(
+                current,
+                current - 1,
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => break,
+                Err(actual) => current = actual,
+            }
+        }
+    }
+
+    pub(super) fn now_epoch_secs() -> u64 {
+        SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .unwrap_or_default()
+            .as_secs()
+    }
+
+    pub(super) fn refresh_cached_epoch_secs(&self) -> u64 {
+        let now_epoch_secs = Self::now_epoch_secs();
+        self.cached_epoch_secs
+            .store(now_epoch_secs, Ordering::Relaxed);
+        now_epoch_secs
+    }
+
+    pub(super) fn cached_epoch_secs(&self) -> u64 {
+        let cached = self.cached_epoch_secs.load(Ordering::Relaxed);
+        if cached != 0 {
+            return cached;
+        }
+        self.refresh_cached_epoch_secs()
+    }
+
+    pub(super) fn touch_user_stats(&self, stats: &UserStats) {
+        stats
+            .last_seen_epoch_secs
+            .store(self.cached_epoch_secs(), Ordering::Relaxed);
+    }
+
+    pub(crate) fn get_or_create_user_stats_handle(&self, user: &str) -> Arc<UserStats> {
+        if let Some(existing) = self.user_stats.get(user) {
+            let handle = Arc::clone(existing.value());
+            self.touch_user_stats(handle.as_ref());
+            return handle;
+        }
+
+        let entry = self.user_stats.entry(user.to_string()).or_default();
+        if entry.last_seen_epoch_secs.load(Ordering::Relaxed) == 0 {
+            self.touch_user_stats(entry.value().as_ref());
+        }
+        Arc::clone(entry.value())
+    }
+
+    pub(crate) async fn run_periodic_user_stats_maintenance(self: Arc<Self>) {
+        let mut interval = tokio::time::interval(Duration::from_secs(60));
+        loop {
+            interval.tick().await;
+            self.maybe_cleanup_user_stats();
+        }
+    }
+
+    #[inline]
+    pub(crate) fn add_user_octets_from_handle(&self, user_stats: &UserStats, bytes: u64) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        self.touch_user_stats(user_stats);
+        user_stats
+            .octets_from_client
+            .fetch_add(bytes, Ordering::Relaxed);
+    }
+
+    #[inline]
+    pub(crate) fn add_user_octets_to_handle(&self, user_stats: &UserStats, bytes: u64) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        self.touch_user_stats(user_stats);
+        user_stats
+            .octets_to_client
+            .fetch_add(bytes, Ordering::Relaxed);
+    }
+
+    #[inline]
+    pub(crate) fn add_user_traffic_from_handle(&self, user_stats: &UserStats, bytes: u64) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        self.touch_user_stats(user_stats);
+        user_stats
+            .octets_from_client
+            .fetch_add(bytes, Ordering::Relaxed);
+        user_stats.msgs_from_client.fetch_add(1, Ordering::Relaxed);
+    }
+
+    #[inline]
+    pub(crate) fn add_user_traffic_to_handle(&self, user_stats: &UserStats, bytes: u64) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        self.touch_user_stats(user_stats);
+        user_stats
+            .octets_to_client
+            .fetch_add(bytes, Ordering::Relaxed);
+        user_stats.msgs_to_client.fetch_add(1, Ordering::Relaxed);
+    }
+
+    #[inline]
+    pub(crate) fn increment_user_msgs_from_handle(&self, user_stats: &UserStats) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        self.touch_user_stats(user_stats);
+        user_stats.msgs_from_client.fetch_add(1, Ordering::Relaxed);
+    }
+
+    #[inline]
+    pub(crate) fn increment_user_msgs_to_handle(&self, user_stats: &UserStats) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        self.touch_user_stats(user_stats);
+        user_stats.msgs_to_client.fetch_add(1, Ordering::Relaxed);
+    }
+
+    /// Charges already committed bytes in a post-I/O path.
+    ///
+    /// This helper is intentionally separate from `quota_try_reserve` to avoid
+    /// mixing reserve and post-charge on a single I/O event.
+    #[inline]
+    pub(crate) fn quota_charge_post_write(&self, user_stats: &UserStats, bytes: u64) -> u64 {
+        self.touch_user_stats(user_stats);
+        user_stats
+            .quota_used
+            .fetch_add(bytes, Ordering::Relaxed)
+            .saturating_add(bytes)
+    }
+
+    pub(super) fn maybe_cleanup_user_stats(&self) {
+        const USER_STATS_CLEANUP_INTERVAL_SECS: u64 = 60;
+        const USER_STATS_IDLE_TTL_SECS: u64 = 24 * 60 * 60;
+
+        let now_epoch_secs = self.refresh_cached_epoch_secs();
+        let last_cleanup_epoch_secs = self
+            .user_stats_last_cleanup_epoch_secs
+            .load(Ordering::Relaxed);
+        if now_epoch_secs.saturating_sub(last_cleanup_epoch_secs) < USER_STATS_CLEANUP_INTERVAL_SECS
+        {
+            return;
+        }
+        if self
+            .user_stats_last_cleanup_epoch_secs
+            .compare_exchange(
+                last_cleanup_epoch_secs,
+                now_epoch_secs,
+                Ordering::AcqRel,
+                Ordering::Relaxed,
+            )
+            .is_err()
+        {
+            return;
+        }
+
+        self.user_stats.retain(|_, stats| {
+            if stats.curr_connects.load(Ordering::Relaxed) > 0 {
+                return true;
+            }
+            let last_seen_epoch_secs = stats.last_seen_epoch_secs.load(Ordering::Relaxed);
+            now_epoch_secs.saturating_sub(last_seen_epoch_secs) <= USER_STATS_IDLE_TTL_SECS
+        });
+    }
+}

src/stats/me_counters.rs

@@ -0,0 +1,442 @@
+use super::*;
+
+impl Stats {
+    pub fn increment_me_keepalive_sent(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_keepalive_sent.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_keepalive_failed(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_keepalive_failed.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_keepalive_pong(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_keepalive_pong.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_keepalive_timeout(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_keepalive_timeout.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_keepalive_timeout_by(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_keepalive_timeout
+                .fetch_add(value, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_rpc_proxy_req_signal_sent_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_rpc_proxy_req_signal_sent_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_rpc_proxy_req_signal_failed_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_rpc_proxy_req_signal_failed_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_rpc_proxy_req_signal_skipped_no_meta_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_rpc_proxy_req_signal_skipped_no_meta_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_rpc_proxy_req_signal_response_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_rpc_proxy_req_signal_response_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_rpc_proxy_req_signal_close_sent_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_rpc_proxy_req_signal_close_sent_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_reconnect_attempt(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_reconnect_attempts.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_reconnect_success(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_reconnect_success.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_handshake_reject_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_handshake_reject_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_handshake_error_code(&self, code: i32) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        let entry = self
+            .me_handshake_error_codes
+            .entry(code)
+            .or_insert_with(|| AtomicU64::new(0));
+        entry.fetch_add(1, Ordering::Relaxed);
+    }
+    pub fn increment_me_reader_eof_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_reader_eof_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_idle_close_by_peer_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_idle_close_by_peer_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_relay_idle_soft_mark_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.relay_idle_soft_mark_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_relay_idle_hard_close_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.relay_idle_hard_close_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_relay_pressure_evict_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.relay_pressure_evict_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_relay_protocol_desync_close_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.relay_protocol_desync_close_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_crc_mismatch(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_crc_mismatch.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_seq_mismatch(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_seq_mismatch.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_route_drop_no_conn(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_route_drop_no_conn.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_route_drop_channel_closed(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_route_drop_channel_closed
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_route_drop_queue_full(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_route_drop_queue_full
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_route_drop_queue_full_base(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_route_drop_queue_full_base
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_route_drop_queue_full_high(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_route_drop_queue_full_high
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_fair_pressure_state_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_fair_pressure_state_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_fair_active_flows_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_fair_active_flows_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_fair_queued_bytes_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_fair_queued_bytes_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_fair_standing_flows_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_fair_standing_flows_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_fair_backpressured_flows_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_fair_backpressured_flows_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_fair_scheduler_rounds_total(&self, value: u64) {
+        if self.telemetry_me_allows_normal() && value > 0 {
+            self.me_fair_scheduler_rounds_total
+                .fetch_add(value, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_fair_deficit_grants_total(&self, value: u64) {
+        if self.telemetry_me_allows_normal() && value > 0 {
+            self.me_fair_deficit_grants_total
+                .fetch_add(value, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_fair_deficit_skips_total(&self, value: u64) {
+        if self.telemetry_me_allows_normal() && value > 0 {
+            self.me_fair_deficit_skips_total
+                .fetch_add(value, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_fair_enqueue_rejects_total(&self, value: u64) {
+        if self.telemetry_me_allows_normal() && value > 0 {
+            self.me_fair_enqueue_rejects_total
+                .fetch_add(value, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_fair_shed_drops_total(&self, value: u64) {
+        if self.telemetry_me_allows_normal() && value > 0 {
+            self.me_fair_shed_drops_total
+                .fetch_add(value, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_fair_penalties_total(&self, value: u64) {
+        if self.telemetry_me_allows_normal() && value > 0 {
+            self.me_fair_penalties_total
+                .fetch_add(value, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_fair_downstream_stalls_total(&self, value: u64) {
+        if self.telemetry_me_allows_normal() && value > 0 {
+            self.me_fair_downstream_stalls_total
+                .fetch_add(value, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_d2c_batches_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_d2c_batches_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_d2c_batch_frames_total(&self, frames: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_d2c_batch_frames_total
+                .fetch_add(frames, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_d2c_batch_bytes_total(&self, bytes: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_d2c_batch_bytes_total
+                .fetch_add(bytes, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_d2c_flush_reason(&self, reason: MeD2cFlushReason) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match reason {
+            MeD2cFlushReason::QueueDrain => {
+                self.me_d2c_flush_reason_queue_drain_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeD2cFlushReason::BatchFrames => {
+                self.me_d2c_flush_reason_batch_frames_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeD2cFlushReason::BatchBytes => {
+                self.me_d2c_flush_reason_batch_bytes_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeD2cFlushReason::MaxDelay => {
+                self.me_d2c_flush_reason_max_delay_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeD2cFlushReason::AckImmediate => {
+                self.me_d2c_flush_reason_ack_immediate_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeD2cFlushReason::Close => {
+                self.me_d2c_flush_reason_close_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_me_d2c_data_frames_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_d2c_data_frames_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_d2c_ack_frames_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_d2c_ack_frames_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn add_me_d2c_payload_bytes_total(&self, bytes: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_d2c_payload_bytes_total
+                .fetch_add(bytes, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_d2c_write_mode(&self, mode: MeD2cWriteMode) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match mode {
+            MeD2cWriteMode::Coalesced => {
+                self.me_d2c_write_mode_coalesced_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeD2cWriteMode::Split => {
+                self.me_d2c_write_mode_split_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_me_d2c_quota_reject_total(&self, stage: MeD2cQuotaRejectStage) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match stage {
+            MeD2cQuotaRejectStage::PreWrite => {
+                self.me_d2c_quota_reject_pre_write_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeD2cQuotaRejectStage::PostWrite => {
+                self.me_d2c_quota_reject_post_write_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn observe_me_d2c_frame_buf_shrink(&self, bytes_freed: u64) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        self.me_d2c_frame_buf_shrink_total
+            .fetch_add(1, Ordering::Relaxed);
+        self.me_d2c_frame_buf_shrink_bytes_total
+            .fetch_add(bytes_freed, Ordering::Relaxed);
+    }
+    pub fn observe_me_d2c_batch_frames(&self, frames: u64) {
+        if !self.telemetry_me_allows_debug() {
+            return;
+        }
+        match frames {
+            0 => {}
+            1 => {
+                self.me_d2c_batch_frames_bucket_1
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            2..=4 => {
+                self.me_d2c_batch_frames_bucket_2_4
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            5..=8 => {
+                self.me_d2c_batch_frames_bucket_5_8
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            9..=16 => {
+                self.me_d2c_batch_frames_bucket_9_16
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            17..=32 => {
+                self.me_d2c_batch_frames_bucket_17_32
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            _ => {
+                self.me_d2c_batch_frames_bucket_gt_32
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn observe_me_d2c_batch_bytes(&self, bytes: u64) {
+        if !self.telemetry_me_allows_debug() {
+            return;
+        }
+        match bytes {
+            0..=1024 => {
+                self.me_d2c_batch_bytes_bucket_0_1k
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            1025..=4096 => {
+                self.me_d2c_batch_bytes_bucket_1k_4k
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            4097..=16_384 => {
+                self.me_d2c_batch_bytes_bucket_4k_16k
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            16_385..=65_536 => {
+                self.me_d2c_batch_bytes_bucket_16k_64k
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            65_537..=131_072 => {
+                self.me_d2c_batch_bytes_bucket_64k_128k
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            _ => {
+                self.me_d2c_batch_bytes_bucket_gt_128k
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn observe_me_d2c_flush_duration_us(&self, duration_us: u64) {
+        if !self.telemetry_me_allows_debug() {
+            return;
+        }
+        match duration_us {
+            0..=50 => {
+                self.me_d2c_flush_duration_us_bucket_0_50
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            51..=200 => {
+                self.me_d2c_flush_duration_us_bucket_51_200
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            201..=1000 => {
+                self.me_d2c_flush_duration_us_bucket_201_1000
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            1001..=5000 => {
+                self.me_d2c_flush_duration_us_bucket_1001_5000
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            5001..=20_000 => {
+                self.me_d2c_flush_duration_us_bucket_5001_20000
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            _ => {
+                self.me_d2c_flush_duration_us_bucket_gt_20000
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_me_d2c_batch_timeout_armed_total(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_d2c_batch_timeout_armed_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_d2c_batch_timeout_fired_total(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_d2c_batch_timeout_fired_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+}

src/stats/me_getters.rs

@@ -0,0 +1,398 @@
+use super::*;
+
+impl Stats {
+    pub fn get_me_handshake_error_code_counts(&self) -> Vec<(i32, u64)> {
+        let mut out: Vec<(i32, u64)> = self
+            .me_handshake_error_codes
+            .iter()
+            .map(|entry| (*entry.key(), entry.value().load(Ordering::Relaxed)))
+            .collect();
+        out.sort_by_key(|(code, _)| *code);
+        out
+    }
+    pub fn get_me_route_drop_no_conn(&self) -> u64 {
+        self.me_route_drop_no_conn.load(Ordering::Relaxed)
+    }
+    pub fn get_me_route_drop_channel_closed(&self) -> u64 {
+        self.me_route_drop_channel_closed.load(Ordering::Relaxed)
+    }
+    pub fn get_me_route_drop_queue_full(&self) -> u64 {
+        self.me_route_drop_queue_full.load(Ordering::Relaxed)
+    }
+    pub fn get_me_route_drop_queue_full_base(&self) -> u64 {
+        self.me_route_drop_queue_full_base.load(Ordering::Relaxed)
+    }
+    pub fn get_me_route_drop_queue_full_high(&self) -> u64 {
+        self.me_route_drop_queue_full_high.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_pressure_state_gauge(&self) -> u64 {
+        self.me_fair_pressure_state_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_active_flows_gauge(&self) -> u64 {
+        self.me_fair_active_flows_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_queued_bytes_gauge(&self) -> u64 {
+        self.me_fair_queued_bytes_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_standing_flows_gauge(&self) -> u64 {
+        self.me_fair_standing_flows_gauge.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_backpressured_flows_gauge(&self) -> u64 {
+        self.me_fair_backpressured_flows_gauge
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_scheduler_rounds_total(&self) -> u64 {
+        self.me_fair_scheduler_rounds_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_deficit_grants_total(&self) -> u64 {
+        self.me_fair_deficit_grants_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_deficit_skips_total(&self) -> u64 {
+        self.me_fair_deficit_skips_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_enqueue_rejects_total(&self) -> u64 {
+        self.me_fair_enqueue_rejects_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_shed_drops_total(&self) -> u64 {
+        self.me_fair_shed_drops_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_penalties_total(&self) -> u64 {
+        self.me_fair_penalties_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_fair_downstream_stalls_total(&self) -> u64 {
+        self.me_fair_downstream_stalls_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batches_total(&self) -> u64 {
+        self.me_d2c_batches_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_frames_total(&self) -> u64 {
+        self.me_d2c_batch_frames_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_bytes_total(&self) -> u64 {
+        self.me_d2c_batch_bytes_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_reason_queue_drain_total(&self) -> u64 {
+        self.me_d2c_flush_reason_queue_drain_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_reason_batch_frames_total(&self) -> u64 {
+        self.me_d2c_flush_reason_batch_frames_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_reason_batch_bytes_total(&self) -> u64 {
+        self.me_d2c_flush_reason_batch_bytes_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_reason_max_delay_total(&self) -> u64 {
+        self.me_d2c_flush_reason_max_delay_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_reason_ack_immediate_total(&self) -> u64 {
+        self.me_d2c_flush_reason_ack_immediate_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_reason_close_total(&self) -> u64 {
+        self.me_d2c_flush_reason_close_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_data_frames_total(&self) -> u64 {
+        self.me_d2c_data_frames_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_ack_frames_total(&self) -> u64 {
+        self.me_d2c_ack_frames_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_payload_bytes_total(&self) -> u64 {
+        self.me_d2c_payload_bytes_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_write_mode_coalesced_total(&self) -> u64 {
+        self.me_d2c_write_mode_coalesced_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_write_mode_split_total(&self) -> u64 {
+        self.me_d2c_write_mode_split_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_quota_reject_pre_write_total(&self) -> u64 {
+        self.me_d2c_quota_reject_pre_write_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_quota_reject_post_write_total(&self) -> u64 {
+        self.me_d2c_quota_reject_post_write_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_frame_buf_shrink_total(&self) -> u64 {
+        self.me_d2c_frame_buf_shrink_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_frame_buf_shrink_bytes_total(&self) -> u64 {
+        self.me_d2c_frame_buf_shrink_bytes_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_frames_bucket_1(&self) -> u64 {
+        self.me_d2c_batch_frames_bucket_1.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_frames_bucket_2_4(&self) -> u64 {
+        self.me_d2c_batch_frames_bucket_2_4.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_frames_bucket_5_8(&self) -> u64 {
+        self.me_d2c_batch_frames_bucket_5_8.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_frames_bucket_9_16(&self) -> u64 {
+        self.me_d2c_batch_frames_bucket_9_16.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_frames_bucket_17_32(&self) -> u64 {
+        self.me_d2c_batch_frames_bucket_17_32
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_frames_bucket_gt_32(&self) -> u64 {
+        self.me_d2c_batch_frames_bucket_gt_32
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_bytes_bucket_0_1k(&self) -> u64 {
+        self.me_d2c_batch_bytes_bucket_0_1k.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_bytes_bucket_1k_4k(&self) -> u64 {
+        self.me_d2c_batch_bytes_bucket_1k_4k.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_bytes_bucket_4k_16k(&self) -> u64 {
+        self.me_d2c_batch_bytes_bucket_4k_16k
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_bytes_bucket_16k_64k(&self) -> u64 {
+        self.me_d2c_batch_bytes_bucket_16k_64k
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_bytes_bucket_64k_128k(&self) -> u64 {
+        self.me_d2c_batch_bytes_bucket_64k_128k
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_bytes_bucket_gt_128k(&self) -> u64 {
+        self.me_d2c_batch_bytes_bucket_gt_128k
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_duration_us_bucket_0_50(&self) -> u64 {
+        self.me_d2c_flush_duration_us_bucket_0_50
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_duration_us_bucket_51_200(&self) -> u64 {
+        self.me_d2c_flush_duration_us_bucket_51_200
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_duration_us_bucket_201_1000(&self) -> u64 {
+        self.me_d2c_flush_duration_us_bucket_201_1000
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_duration_us_bucket_1001_5000(&self) -> u64 {
+        self.me_d2c_flush_duration_us_bucket_1001_5000
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_duration_us_bucket_5001_20000(&self) -> u64 {
+        self.me_d2c_flush_duration_us_bucket_5001_20000
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_flush_duration_us_bucket_gt_20000(&self) -> u64 {
+        self.me_d2c_flush_duration_us_bucket_gt_20000
+            .load(Ordering::Relaxed)
+    }
+
+    pub fn get_buffer_pool_pooled_gauge(&self) -> u64 {
+        self.buffer_pool_pooled_gauge.load(Ordering::Relaxed)
+    }
+
+    pub fn get_buffer_pool_allocated_gauge(&self) -> u64 {
+        self.buffer_pool_allocated_gauge.load(Ordering::Relaxed)
+    }
+
+    pub fn get_buffer_pool_in_use_gauge(&self) -> u64 {
+        self.buffer_pool_in_use_gauge.load(Ordering::Relaxed)
+    }
+
+    pub fn get_me_c2me_send_full_total(&self) -> u64 {
+        self.me_c2me_send_full_total.load(Ordering::Relaxed)
+    }
+
+    pub fn get_me_c2me_send_high_water_total(&self) -> u64 {
+        self.me_c2me_send_high_water_total.load(Ordering::Relaxed)
+    }
+
+    pub fn get_me_c2me_send_timeout_total(&self) -> u64 {
+        self.me_c2me_send_timeout_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_timeout_armed_total(&self) -> u64 {
+        self.me_d2c_batch_timeout_armed_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_d2c_batch_timeout_fired_total(&self) -> u64 {
+        self.me_d2c_batch_timeout_fired_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_sorted_rr_success_try_total(&self) -> u64 {
+        self.me_writer_pick_sorted_rr_success_try_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_sorted_rr_success_fallback_total(&self) -> u64 {
+        self.me_writer_pick_sorted_rr_success_fallback_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_sorted_rr_full_total(&self) -> u64 {
+        self.me_writer_pick_sorted_rr_full_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_sorted_rr_closed_total(&self) -> u64 {
+        self.me_writer_pick_sorted_rr_closed_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_sorted_rr_no_candidate_total(&self) -> u64 {
+        self.me_writer_pick_sorted_rr_no_candidate_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_p2c_success_try_total(&self) -> u64 {
+        self.me_writer_pick_p2c_success_try_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_p2c_success_fallback_total(&self) -> u64 {
+        self.me_writer_pick_p2c_success_fallback_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_p2c_full_total(&self) -> u64 {
+        self.me_writer_pick_p2c_full_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_p2c_closed_total(&self) -> u64 {
+        self.me_writer_pick_p2c_closed_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_p2c_no_candidate_total(&self) -> u64 {
+        self.me_writer_pick_p2c_no_candidate_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_blocking_fallback_total(&self) -> u64 {
+        self.me_writer_pick_blocking_fallback_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_pick_mode_switch_total(&self) -> u64 {
+        self.me_writer_pick_mode_switch_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_socks_kdf_strict_reject(&self) -> u64 {
+        self.me_socks_kdf_strict_reject.load(Ordering::Relaxed)
+    }
+    pub fn get_me_socks_kdf_compat_fallback(&self) -> u64 {
+        self.me_socks_kdf_compat_fallback.load(Ordering::Relaxed)
+    }
+    pub fn get_secure_padding_invalid(&self) -> u64 {
+        self.secure_padding_invalid.load(Ordering::Relaxed)
+    }
+    pub fn get_desync_total(&self) -> u64 {
+        self.desync_total.load(Ordering::Relaxed)
+    }
+    pub fn get_desync_full_logged(&self) -> u64 {
+        self.desync_full_logged.load(Ordering::Relaxed)
+    }
+    pub fn get_desync_suppressed(&self) -> u64 {
+        self.desync_suppressed.load(Ordering::Relaxed)
+    }
+    pub fn get_desync_frames_bucket_0(&self) -> u64 {
+        self.desync_frames_bucket_0.load(Ordering::Relaxed)
+    }
+    pub fn get_desync_frames_bucket_1_2(&self) -> u64 {
+        self.desync_frames_bucket_1_2.load(Ordering::Relaxed)
+    }
+    pub fn get_desync_frames_bucket_3_10(&self) -> u64 {
+        self.desync_frames_bucket_3_10.load(Ordering::Relaxed)
+    }
+    pub fn get_desync_frames_bucket_gt_10(&self) -> u64 {
+        self.desync_frames_bucket_gt_10.load(Ordering::Relaxed)
+    }
+    pub fn get_pool_swap_total(&self) -> u64 {
+        self.pool_swap_total.load(Ordering::Relaxed)
+    }
+    pub fn get_pool_drain_active(&self) -> u64 {
+        self.pool_drain_active.load(Ordering::Relaxed)
+    }
+    pub fn get_pool_force_close_total(&self) -> u64 {
+        self.pool_force_close_total.load(Ordering::Relaxed)
+    }
+    pub fn get_pool_stale_pick_total(&self) -> u64 {
+        self.pool_stale_pick_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_removed_total(&self) -> u64 {
+        self.me_writer_removed_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_removed_unexpected_total(&self) -> u64 {
+        self.me_writer_removed_unexpected_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_refill_triggered_total(&self) -> u64 {
+        self.me_refill_triggered_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_refill_skipped_inflight_total(&self) -> u64 {
+        self.me_refill_skipped_inflight_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_refill_failed_total(&self) -> u64 {
+        self.me_refill_failed_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_restored_same_endpoint_total(&self) -> u64 {
+        self.me_writer_restored_same_endpoint_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_writer_restored_fallback_total(&self) -> u64 {
+        self.me_writer_restored_fallback_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_me_no_writer_failfast_total(&self) -> u64 {
+        self.me_no_writer_failfast_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_hybrid_timeout_total(&self) -> u64 {
+        self.me_hybrid_timeout_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_async_recovery_trigger_total(&self) -> u64 {
+        self.me_async_recovery_trigger_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_inline_recovery_total(&self) -> u64 {
+        self.me_inline_recovery_total.load(Ordering::Relaxed)
+    }
+    pub fn get_ip_reservation_rollback_tcp_limit_total(&self) -> u64 {
+        self.ip_reservation_rollback_tcp_limit_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_ip_reservation_rollback_quota_limit_total(&self) -> u64 {
+        self.ip_reservation_rollback_quota_limit_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_quota_refund_bytes_total(&self) -> u64 {
+        self.quota_refund_bytes_total.load(Ordering::Relaxed)
+    }
+    pub fn get_quota_contention_total(&self) -> u64 {
+        self.quota_contention_total.load(Ordering::Relaxed)
+    }
+    pub fn get_quota_contention_timeout_total(&self) -> u64 {
+        self.quota_contention_timeout_total.load(Ordering::Relaxed)
+    }
+    pub fn get_quota_acquire_cancelled_total(&self) -> u64 {
+        self.quota_acquire_cancelled_total.load(Ordering::Relaxed)
+    }
+    pub fn get_quota_write_fail_bytes_total(&self) -> u64 {
+        self.quota_write_fail_bytes_total.load(Ordering::Relaxed)
+    }
+    pub fn get_quota_write_fail_events_total(&self) -> u64 {
+        self.quota_write_fail_events_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_child_join_timeout_total(&self) -> u64 {
+        self.me_child_join_timeout_total.load(Ordering::Relaxed)
+    }
+    pub fn get_me_child_abort_total(&self) -> u64 {
+        self.me_child_abort_total.load(Ordering::Relaxed)
+    }
+    pub fn get_flow_wait_middle_rate_limit_total(&self) -> u64 {
+        self.flow_wait_middle_rate_limit_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_flow_wait_middle_rate_limit_cancelled_total(&self) -> u64 {
+        self.flow_wait_middle_rate_limit_cancelled_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_flow_wait_middle_rate_limit_ms_total(&self) -> u64 {
+        self.flow_wait_middle_rate_limit_ms_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_session_drop_fallback_total(&self) -> u64 {
+        self.session_drop_fallback_total.load(Ordering::Relaxed)
+    }
+}

src/stats/replay.rs

@@ -0,0 +1,356 @@
+use std::borrow::Borrow;
+use std::collections::VecDeque;
+use std::collections::hash_map::DefaultHasher;
+use std::hash::{Hash, Hasher};
+use std::num::NonZeroUsize;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::{Duration, Instant};
+
+use lru::LruCache;
+use parking_lot::Mutex;
+use tracing::debug;
+
+const REPLAY_INLINE_KEY_CAP: usize = 48;
+
+#[derive(Clone)]
+enum ReplayKey {
+    Inline {
+        len: u8,
+        bytes: [u8; REPLAY_INLINE_KEY_CAP],
+    },
+    Heap(Arc<[u8]>),
+}
+
+impl ReplayKey {
+    fn from_slice(key: &[u8]) -> Self {
+        if key.len() <= REPLAY_INLINE_KEY_CAP {
+            let mut bytes = [0u8; REPLAY_INLINE_KEY_CAP];
+            bytes[..key.len()].copy_from_slice(key);
+            return Self::Inline {
+                len: key.len() as u8,
+                bytes,
+            };
+        }
+
+        Self::Heap(Arc::from(key))
+    }
+
+    fn as_slice(&self) -> &[u8] {
+        match self {
+            Self::Inline { len, bytes } => &bytes[..*len as usize],
+            Self::Heap(bytes) => bytes.as_ref(),
+        }
+    }
+}
+
+impl Borrow<[u8]> for ReplayKey {
+    fn borrow(&self) -> &[u8] {
+        self.as_slice()
+    }
+}
+
+impl PartialEq for ReplayKey {
+    fn eq(&self, other: &Self) -> bool {
+        self.as_slice() == other.as_slice()
+    }
+}
+
+impl Eq for ReplayKey {}
+
+impl Hash for ReplayKey {
+    fn hash<H: Hasher>(&self, state: &mut H) {
+        self.as_slice().hash(state);
+    }
+}
+
+pub struct ReplayChecker {
+    handshake_shards: Vec<Mutex<ReplayShard>>,
+    tls_shards: Vec<Mutex<ReplayShard>>,
+    shard_mask: usize,
+    window: Duration,
+    tls_window: Duration,
+    checks: AtomicU64,
+    hits: AtomicU64,
+    additions: AtomicU64,
+    cleanups: AtomicU64,
+}
+
+struct ReplayEntry {
+    seq: u64,
+}
+
+struct ReplayShard {
+    cache: LruCache<ReplayKey, ReplayEntry>,
+    queue: VecDeque<(Instant, ReplayKey, u64)>,
+    seq_counter: u64,
+    capacity: usize,
+}
+
+impl ReplayShard {
+    fn new(cap: NonZeroUsize) -> Self {
+        Self {
+            cache: LruCache::new(cap),
+            queue: VecDeque::with_capacity(cap.get()),
+            seq_counter: 0,
+            capacity: cap.get(),
+        }
+    }
+
+    fn next_seq(&mut self) -> u64 {
+        self.seq_counter += 1;
+        self.seq_counter
+    }
+
+    fn cleanup(&mut self, now: Instant, window: Duration) {
+        if window.is_zero() {
+            self.cache.clear();
+            self.queue.clear();
+            return;
+        }
+        let cutoff = now.checked_sub(window).unwrap_or(now);
+
+        while let Some((ts, _, _)) = self.queue.front() {
+            if *ts >= cutoff {
+                break;
+            }
+            self.evict_queue_front();
+        }
+    }
+
+    fn evict_queue_front(&mut self) {
+        let Some((_, key, queue_seq)) = self.queue.pop_front() else {
+            return;
+        };
+
+        if let Some(entry) = self.cache.peek(key.as_slice())
+            && entry.seq == queue_seq
+        {
+            self.cache.pop(key.as_slice());
+        }
+    }
+
+    fn check(&mut self, key: &[u8], now: Instant, window: Duration) -> bool {
+        if window.is_zero() {
+            return false;
+        }
+        self.cleanup(now, window);
+        self.cache.get(key).is_some()
+    }
+
+    fn add_owned(&mut self, key: ReplayKey, now: Instant, window: Duration) {
+        if window.is_zero() {
+            return;
+        }
+        self.cleanup(now, window);
+        if self.cache.peek(key.as_slice()).is_some() {
+            return;
+        }
+        while self.queue.len() >= self.capacity {
+            self.evict_queue_front();
+        }
+
+        let seq = self.next_seq();
+        self.cache.put(key.clone(), ReplayEntry { seq });
+        self.queue.push_back((now, key, seq));
+    }
+
+    fn len(&self) -> usize {
+        self.cache.len()
+    }
+}
+
+impl ReplayChecker {
+    pub fn new(total_capacity: usize, window: Duration) -> Self {
+        const MIN_TLS_REPLAY_WINDOW: Duration = Duration::from_secs(120);
+        let num_shards = 64;
+        let shard_capacity = (total_capacity / num_shards).max(1);
+        let cap = NonZeroUsize::new(shard_capacity).unwrap();
+
+        let mut handshake_shards = Vec::with_capacity(num_shards);
+        let mut tls_shards = Vec::with_capacity(num_shards);
+        for _ in 0..num_shards {
+            handshake_shards.push(Mutex::new(ReplayShard::new(cap)));
+            tls_shards.push(Mutex::new(ReplayShard::new(cap)));
+        }
+
+        Self {
+            handshake_shards,
+            tls_shards,
+            shard_mask: num_shards - 1,
+            window,
+            tls_window: window.max(MIN_TLS_REPLAY_WINDOW),
+            checks: AtomicU64::new(0),
+            hits: AtomicU64::new(0),
+            additions: AtomicU64::new(0),
+            cleanups: AtomicU64::new(0),
+        }
+    }
+
+    fn get_shard_idx(&self, key: &[u8]) -> usize {
+        let mut hasher = DefaultHasher::new();
+        key.hash(&mut hasher);
+        (hasher.finish() as usize) & self.shard_mask
+    }
+
+    fn check_and_add_internal(
+        &self,
+        data: &[u8],
+        shards: &[Mutex<ReplayShard>],
+        window: Duration,
+    ) -> bool {
+        self.checks.fetch_add(1, Ordering::Relaxed);
+        let idx = self.get_shard_idx(data);
+        let owned_key = ReplayKey::from_slice(data);
+        let mut shard = shards[idx].lock();
+        let now = Instant::now();
+        let found = shard.check(data, now, window);
+        if found {
+            self.hits.fetch_add(1, Ordering::Relaxed);
+        } else {
+            shard.add_owned(owned_key, now, window);
+            self.additions.fetch_add(1, Ordering::Relaxed);
+        }
+        found
+    }
+
+    fn check_only_internal(
+        &self,
+        data: &[u8],
+        shards: &[Mutex<ReplayShard>],
+        window: Duration,
+    ) -> bool {
+        self.checks.fetch_add(1, Ordering::Relaxed);
+        let idx = self.get_shard_idx(data);
+        let mut shard = shards[idx].lock();
+        let found = shard.check(data, Instant::now(), window);
+        if found {
+            self.hits.fetch_add(1, Ordering::Relaxed);
+        }
+        found
+    }
+
+    fn add_only(&self, data: &[u8], shards: &[Mutex<ReplayShard>], window: Duration) {
+        self.additions.fetch_add(1, Ordering::Relaxed);
+        let idx = self.get_shard_idx(data);
+        let owned_key = ReplayKey::from_slice(data);
+        let mut shard = shards[idx].lock();
+        shard.add_owned(owned_key, Instant::now(), window);
+    }
+
+    pub fn check_and_add_handshake(&self, data: &[u8]) -> bool {
+        self.check_and_add_internal(data, &self.handshake_shards, self.window)
+    }
+
+    pub fn check_and_add_tls_digest(&self, data: &[u8]) -> bool {
+        self.check_and_add_internal(data, &self.tls_shards, self.tls_window)
+    }
+
+    pub fn check_handshake(&self, data: &[u8]) -> bool {
+        self.check_and_add_handshake(data)
+    }
+
+    pub fn add_handshake(&self, data: &[u8]) {
+        self.add_only(data, &self.handshake_shards, self.window)
+    }
+
+    pub fn check_tls_digest(&self, data: &[u8]) -> bool {
+        self.check_only_internal(data, &self.tls_shards, self.tls_window)
+    }
+
+    pub fn add_tls_digest(&self, data: &[u8]) {
+        self.add_only(data, &self.tls_shards, self.tls_window)
+    }
+
+    pub fn stats(&self) -> ReplayStats {
+        let mut total_entries = 0;
+        let mut total_queue_len = 0;
+        for shard in &self.handshake_shards {
+            let s = shard.lock();
+            total_entries += s.cache.len();
+            total_queue_len += s.queue.len();
+        }
+        for shard in &self.tls_shards {
+            let s = shard.lock();
+            total_entries += s.cache.len();
+            total_queue_len += s.queue.len();
+        }
+
+        ReplayStats {
+            total_entries,
+            total_queue_len,
+            total_checks: self.checks.load(Ordering::Relaxed),
+            total_hits: self.hits.load(Ordering::Relaxed),
+            total_additions: self.additions.load(Ordering::Relaxed),
+            total_cleanups: self.cleanups.load(Ordering::Relaxed),
+            num_shards: self.handshake_shards.len() + self.tls_shards.len(),
+            window_secs: self.window.as_secs(),
+        }
+    }
+
+    pub async fn run_periodic_cleanup(&self) {
+        let interval = if self.window.as_secs() > 60 {
+            Duration::from_secs(30)
+        } else {
+            Duration::from_secs((self.window.as_secs().max(1) / 2).max(1))
+        };
+
+        loop {
+            tokio::time::sleep(interval).await;
+
+            let now = Instant::now();
+            let mut cleaned = 0usize;
+
+            for shard_mutex in &self.handshake_shards {
+                let mut shard = shard_mutex.lock();
+                let before = shard.len();
+                shard.cleanup(now, self.window);
+                let after = shard.len();
+                cleaned += before.saturating_sub(after);
+            }
+            for shard_mutex in &self.tls_shards {
+                let mut shard = shard_mutex.lock();
+                let before = shard.len();
+                shard.cleanup(now, self.tls_window);
+                let after = shard.len();
+                cleaned += before.saturating_sub(after);
+            }
+
+            self.cleanups.fetch_add(1, Ordering::Relaxed);
+
+            if cleaned > 0 {
+                debug!(cleaned = cleaned, "Replay checker: periodic cleanup");
+            }
+        }
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct ReplayStats {
+    pub total_entries: usize,
+    pub total_queue_len: usize,
+    pub total_checks: u64,
+    pub total_hits: u64,
+    pub total_additions: u64,
+    pub total_cleanups: u64,
+    pub num_shards: usize,
+    pub window_secs: u64,
+}
+
+impl ReplayStats {
+    pub fn hit_rate(&self) -> f64 {
+        if self.total_checks == 0 {
+            0.0
+        } else {
+            (self.total_hits as f64 / self.total_checks as f64) * 100.0
+        }
+    }
+
+    pub fn ghost_ratio(&self) -> f64 {
+        if self.total_entries == 0 {
+            0.0
+        } else {
+            self.total_queue_len as f64 / self.total_entries as f64
+        }
+    }
+}

src/stats/tests.rs

@@ -0,0 +1,317 @@
+use super::*;
+use crate::config::MeTelemetryLevel;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::time::Duration;
+
+#[test]
+fn test_stats_shared_counters() {
+    let stats = Arc::new(Stats::new());
+    stats.increment_connects_all();
+    stats.increment_connects_all();
+    stats.increment_connects_all();
+    assert_eq!(stats.get_connects_all(), 3);
+}
+
+#[test]
+fn test_telemetry_policy_disables_core_and_user_counters() {
+    let stats = Stats::new();
+    stats.apply_telemetry_policy(TelemetryPolicy {
+        core_enabled: false,
+        user_enabled: false,
+        me_level: MeTelemetryLevel::Normal,
+    });
+
+    stats.increment_connects_all();
+    stats.increment_user_connects("alice");
+    stats.add_user_octets_from("alice", 1024);
+    assert_eq!(stats.get_connects_all(), 0);
+    assert_eq!(stats.get_user_curr_connects("alice"), 0);
+    assert_eq!(stats.get_user_total_octets("alice"), 0);
+}
+
+#[test]
+fn test_telemetry_policy_me_silent_blocks_me_counters() {
+    let stats = Stats::new();
+    stats.apply_telemetry_policy(TelemetryPolicy {
+        core_enabled: true,
+        user_enabled: true,
+        me_level: MeTelemetryLevel::Silent,
+    });
+
+    stats.increment_me_crc_mismatch();
+    stats.increment_me_keepalive_sent();
+    stats.increment_me_route_drop_queue_full();
+    stats.increment_me_d2c_batches_total();
+    stats.add_me_d2c_batch_frames_total(4);
+    stats.add_me_d2c_batch_bytes_total(4096);
+    stats.increment_me_d2c_flush_reason(MeD2cFlushReason::BatchBytes);
+    stats.increment_me_d2c_write_mode(MeD2cWriteMode::Coalesced);
+    stats.increment_me_d2c_quota_reject_total(MeD2cQuotaRejectStage::PreWrite);
+    stats.observe_me_d2c_frame_buf_shrink(1024);
+    stats.observe_me_d2c_batch_frames(4);
+    stats.observe_me_d2c_batch_bytes(4096);
+    stats.observe_me_d2c_flush_duration_us(120);
+    stats.increment_me_d2c_batch_timeout_armed_total();
+    stats.increment_me_d2c_batch_timeout_fired_total();
+    assert_eq!(stats.get_me_crc_mismatch(), 0);
+    assert_eq!(stats.get_me_keepalive_sent(), 0);
+    assert_eq!(stats.get_me_route_drop_queue_full(), 0);
+    assert_eq!(stats.get_me_d2c_batches_total(), 0);
+    assert_eq!(stats.get_me_d2c_flush_reason_batch_bytes_total(), 0);
+    assert_eq!(stats.get_me_d2c_write_mode_coalesced_total(), 0);
+    assert_eq!(stats.get_me_d2c_quota_reject_pre_write_total(), 0);
+    assert_eq!(stats.get_me_d2c_frame_buf_shrink_total(), 0);
+    assert_eq!(stats.get_me_d2c_batch_frames_bucket_2_4(), 0);
+    assert_eq!(stats.get_me_d2c_batch_bytes_bucket_1k_4k(), 0);
+    assert_eq!(stats.get_me_d2c_flush_duration_us_bucket_51_200(), 0);
+    assert_eq!(stats.get_me_d2c_batch_timeout_armed_total(), 0);
+    assert_eq!(stats.get_me_d2c_batch_timeout_fired_total(), 0);
+}
+
+#[test]
+fn test_telemetry_policy_me_normal_blocks_d2c_debug_metrics() {
+    let stats = Stats::new();
+    stats.apply_telemetry_policy(TelemetryPolicy {
+        core_enabled: true,
+        user_enabled: true,
+        me_level: MeTelemetryLevel::Normal,
+    });
+
+    stats.increment_me_d2c_batches_total();
+    stats.add_me_d2c_batch_frames_total(2);
+    stats.add_me_d2c_batch_bytes_total(2048);
+    stats.increment_me_d2c_flush_reason(MeD2cFlushReason::QueueDrain);
+    stats.observe_me_d2c_batch_frames(2);
+    stats.observe_me_d2c_batch_bytes(2048);
+    stats.observe_me_d2c_flush_duration_us(100);
+    stats.increment_me_d2c_batch_timeout_armed_total();
+    stats.increment_me_d2c_batch_timeout_fired_total();
+
+    assert_eq!(stats.get_me_d2c_batches_total(), 1);
+    assert_eq!(stats.get_me_d2c_batch_frames_total(), 2);
+    assert_eq!(stats.get_me_d2c_batch_bytes_total(), 2048);
+    assert_eq!(stats.get_me_d2c_flush_reason_queue_drain_total(), 1);
+    assert_eq!(stats.get_me_d2c_batch_frames_bucket_2_4(), 0);
+    assert_eq!(stats.get_me_d2c_batch_bytes_bucket_1k_4k(), 0);
+    assert_eq!(stats.get_me_d2c_flush_duration_us_bucket_51_200(), 0);
+    assert_eq!(stats.get_me_d2c_batch_timeout_armed_total(), 0);
+    assert_eq!(stats.get_me_d2c_batch_timeout_fired_total(), 0);
+}
+
+#[test]
+fn test_telemetry_policy_me_debug_enables_d2c_debug_metrics() {
+    let stats = Stats::new();
+    stats.apply_telemetry_policy(TelemetryPolicy {
+        core_enabled: true,
+        user_enabled: true,
+        me_level: MeTelemetryLevel::Debug,
+    });
+
+    stats.observe_me_d2c_batch_frames(7);
+    stats.observe_me_d2c_batch_bytes(70_000);
+    stats.observe_me_d2c_flush_duration_us(1400);
+    stats.increment_me_d2c_batch_timeout_armed_total();
+    stats.increment_me_d2c_batch_timeout_fired_total();
+
+    assert_eq!(stats.get_me_d2c_batch_frames_bucket_5_8(), 1);
+    assert_eq!(stats.get_me_d2c_batch_bytes_bucket_64k_128k(), 1);
+    assert_eq!(stats.get_me_d2c_flush_duration_us_bucket_1001_5000(), 1);
+    assert_eq!(stats.get_me_d2c_batch_timeout_armed_total(), 1);
+    assert_eq!(stats.get_me_d2c_batch_timeout_fired_total(), 1);
+}
+
+#[test]
+fn test_replay_checker_basic() {
+    let checker = ReplayChecker::new(100, Duration::from_secs(60));
+    assert!(!checker.check_handshake(b"test1")); // first time, inserts
+    assert!(checker.check_handshake(b"test1")); // duplicate
+    assert!(!checker.check_handshake(b"test2")); // new key inserts
+}
+
+#[test]
+fn test_replay_checker_duplicate_add() {
+    let checker = ReplayChecker::new(100, Duration::from_secs(60));
+    checker.add_handshake(b"dup");
+    checker.add_handshake(b"dup");
+    assert!(checker.check_handshake(b"dup"));
+}
+
+#[test]
+fn test_replay_checker_expiration() {
+    let checker = ReplayChecker::new(100, Duration::from_millis(50));
+    assert!(!checker.check_handshake(b"expire"));
+    assert!(checker.check_handshake(b"expire"));
+    std::thread::sleep(Duration::from_millis(100));
+    assert!(!checker.check_handshake(b"expire"));
+}
+
+#[test]
+fn test_replay_checker_zero_window_does_not_retain_entries() {
+    let checker = ReplayChecker::new(100, Duration::ZERO);
+
+    for _ in 0..1_000 {
+        assert!(!checker.check_handshake(b"no-retain"));
+        checker.add_handshake(b"no-retain");
+    }
+
+    let stats = checker.stats();
+    assert_eq!(stats.total_entries, 0);
+    assert_eq!(stats.total_queue_len, 0);
+}
+
+#[test]
+fn test_replay_checker_stats() {
+    let checker = ReplayChecker::new(100, Duration::from_secs(60));
+    assert!(!checker.check_handshake(b"k1"));
+    assert!(!checker.check_handshake(b"k2"));
+    assert!(checker.check_handshake(b"k1"));
+    assert!(!checker.check_handshake(b"k3"));
+    let stats = checker.stats();
+    assert_eq!(stats.total_additions, 3);
+    assert_eq!(stats.total_checks, 4);
+    assert_eq!(stats.total_hits, 1);
+}
+
+#[test]
+fn test_replay_checker_many_keys() {
+    let checker = ReplayChecker::new(10_000, Duration::from_secs(60));
+    for i in 0..500u32 {
+        checker.add_handshake(&i.to_le_bytes());
+    }
+    for i in 0..500u32 {
+        assert!(checker.check_handshake(&i.to_le_bytes()));
+    }
+    assert_eq!(checker.stats().total_entries, 500);
+}
+
+#[test]
+fn test_quota_reserve_under_contention_hits_limit_exactly() {
+    let user_stats = Arc::new(UserStats::default());
+    let successes = Arc::new(AtomicU64::new(0));
+    let limit = 8_192u64;
+    let mut workers = Vec::new();
+
+    for _ in 0..8 {
+        let user_stats = user_stats.clone();
+        let successes = successes.clone();
+        workers.push(std::thread::spawn(move || {
+            loop {
+                match user_stats.quota_try_reserve(1, limit) {
+                    Ok(_) => {
+                        successes.fetch_add(1, Ordering::Relaxed);
+                    }
+                    Err(QuotaReserveError::Contended) => {
+                        std::hint::spin_loop();
+                    }
+                    Err(QuotaReserveError::LimitExceeded) => {
+                        break;
+                    }
+                }
+            }
+        }));
+    }
+
+    for worker in workers {
+        worker.join().expect("worker thread must finish");
+    }
+
+    assert_eq!(
+        successes.load(Ordering::Relaxed),
+        limit,
+        "successful reservations must stop exactly at limit"
+    );
+    assert_eq!(user_stats.quota_used(), limit);
+}
+
+#[test]
+fn test_quota_reserve_200x_1k_reaches_100k_without_overshoot() {
+    let user_stats = Arc::new(UserStats::default());
+    let successes = Arc::new(AtomicU64::new(0));
+    let failures = Arc::new(AtomicU64::new(0));
+    let attempts = 200usize;
+    let reserve_bytes = 1_024u64;
+    let limit = 100 * 1_024u64;
+    let mut workers = Vec::with_capacity(attempts);
+
+    for _ in 0..attempts {
+        let user_stats = user_stats.clone();
+        let successes = successes.clone();
+        let failures = failures.clone();
+        workers.push(std::thread::spawn(move || {
+            loop {
+                match user_stats.quota_try_reserve(reserve_bytes, limit) {
+                    Ok(_) => {
+                        successes.fetch_add(1, Ordering::Relaxed);
+                        return;
+                    }
+                    Err(QuotaReserveError::LimitExceeded) => {
+                        failures.fetch_add(1, Ordering::Relaxed);
+                        return;
+                    }
+                    Err(QuotaReserveError::Contended) => {
+                        std::hint::spin_loop();
+                    }
+                }
+            }
+        }));
+    }
+
+    for worker in workers {
+        worker.join().expect("reservation worker must finish");
+    }
+
+    assert_eq!(
+        successes.load(Ordering::Relaxed),
+        100,
+        "exactly 100 reservations of 1 KiB must fit into a 100 KiB quota"
+    );
+    assert_eq!(
+        failures.load(Ordering::Relaxed),
+        100,
+        "remaining workers must fail once quota is fully reserved"
+    );
+    assert_eq!(user_stats.quota_used(), limit);
+}
+
+#[test]
+fn test_quota_used_is_authoritative_and_independent_from_octets_telemetry() {
+    let stats = Stats::new();
+    let user = "quota-authoritative-user";
+    let user_stats = stats.get_or_create_user_stats_handle(user);
+
+    stats.add_user_octets_to_handle(&user_stats, 5);
+    assert_eq!(stats.get_user_total_octets(user), 5);
+    assert_eq!(stats.get_user_quota_used(user), 0);
+
+    stats.quota_charge_post_write(&user_stats, 7);
+    assert_eq!(stats.get_user_total_octets(user), 5);
+    assert_eq!(stats.get_user_quota_used(user), 7);
+}
+
+#[test]
+fn test_cached_handle_survives_map_cleanup_until_last_drop() {
+    let stats = Stats::new();
+    let user = "quota-handle-lifetime-user";
+    let user_stats = stats.get_or_create_user_stats_handle(user);
+    let weak = Arc::downgrade(&user_stats);
+
+    stats.user_stats.remove(user);
+    assert!(
+        stats.user_stats.get(user).is_none(),
+        "map cleanup should remove idle entry"
+    );
+    assert!(
+        weak.upgrade().is_some(),
+        "cached handle must keep user stats object alive after map removal"
+    );
+
+    stats.quota_charge_post_write(user_stats.as_ref(), 3);
+    assert_eq!(user_stats.quota_used(), 3);
+
+    drop(user_stats);
+    assert!(
+        weak.upgrade().is_none(),
+        "user stats object must be dropped after the last cached handle is released"
+    );
+}

src/stats/users.rs

@@ -0,0 +1,249 @@
+use super::*;
+
+impl Stats {
+    pub fn increment_user_connects(&self, user: &str) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        let stats = self.get_or_create_user_stats_handle(user);
+        self.touch_user_stats(stats.as_ref());
+        stats.connects.fetch_add(1, Ordering::Relaxed);
+    }
+
+    pub fn increment_user_curr_connects(&self, user: &str) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        let stats = self.get_or_create_user_stats_handle(user);
+        self.touch_user_stats(stats.as_ref());
+        stats.curr_connects.fetch_add(1, Ordering::Relaxed);
+    }
+
+    pub fn try_acquire_user_curr_connects(&self, user: &str, limit: Option<u64>) -> bool {
+        if !self.telemetry_user_enabled() {
+            return true;
+        }
+
+        let stats = self.get_or_create_user_stats_handle(user);
+        self.touch_user_stats(stats.as_ref());
+
+        let counter = &stats.curr_connects;
+        let mut current = counter.load(Ordering::Relaxed);
+        loop {
+            if let Some(max) = limit
+                && current >= max
+            {
+                return false;
+            }
+            match counter.compare_exchange_weak(
+                current,
+                current.saturating_add(1),
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => return true,
+                Err(actual) => current = actual,
+            }
+        }
+    }
+
+    pub fn decrement_user_curr_connects(&self, user: &str) {
+        if let Some(stats) = self.user_stats.get(user) {
+            self.touch_user_stats(stats.value().as_ref());
+            let counter = &stats.curr_connects;
+            let mut current = counter.load(Ordering::Relaxed);
+            loop {
+                if current == 0 {
+                    break;
+                }
+                match counter.compare_exchange_weak(
+                    current,
+                    current - 1,
+                    Ordering::Relaxed,
+                    Ordering::Relaxed,
+                ) {
+                    Ok(_) => break,
+                    Err(actual) => current = actual,
+                }
+            }
+        }
+    }
+
+    pub fn get_user_curr_connects(&self, user: &str) -> u64 {
+        self.user_stats
+            .get(user)
+            .map(|s| s.curr_connects.load(Ordering::Relaxed))
+            .unwrap_or(0)
+    }
+
+    pub fn add_user_octets_from(&self, user: &str, bytes: u64) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        let stats = self.get_or_create_user_stats_handle(user);
+        self.add_user_octets_from_handle(stats.as_ref(), bytes);
+    }
+
+    pub fn add_user_octets_to(&self, user: &str, bytes: u64) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        let stats = self.get_or_create_user_stats_handle(user);
+        self.add_user_octets_to_handle(stats.as_ref(), bytes);
+    }
+
+    pub fn increment_user_msgs_from(&self, user: &str) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        let stats = self.get_or_create_user_stats_handle(user);
+        self.increment_user_msgs_from_handle(stats.as_ref());
+    }
+
+    pub fn increment_user_msgs_to(&self, user: &str) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
+        let stats = self.get_or_create_user_stats_handle(user);
+        self.increment_user_msgs_to_handle(stats.as_ref());
+    }
+
+    pub fn get_user_total_octets(&self, user: &str) -> u64 {
+        self.user_stats
+            .get(user)
+            .map(|s| {
+                s.octets_from_client.load(Ordering::Relaxed)
+                    + s.octets_to_client.load(Ordering::Relaxed)
+            })
+            .unwrap_or(0)
+    }
+
+    pub fn get_user_quota_used(&self, user: &str) -> u64 {
+        self.user_stats
+            .get(user)
+            .map(|s| s.quota_used.load(Ordering::Relaxed))
+            .unwrap_or(0)
+    }
+
+    pub fn load_user_quota_state(&self, user: &str, used_bytes: u64, last_reset_epoch_secs: u64) {
+        let stats = self.get_or_create_user_stats_handle(user);
+        stats.quota_used.store(used_bytes, Ordering::Relaxed);
+        stats
+            .quota_last_reset_epoch_secs
+            .store(last_reset_epoch_secs, Ordering::Relaxed);
+    }
+
+    pub fn reset_user_quota(&self, user: &str) -> UserQuotaSnapshot {
+        let stats = self.get_or_create_user_stats_handle(user);
+        let last_reset_epoch_secs = Self::now_epoch_secs();
+        stats.quota_used.store(0, Ordering::Relaxed);
+        stats
+            .quota_last_reset_epoch_secs
+            .store(last_reset_epoch_secs, Ordering::Relaxed);
+        UserQuotaSnapshot {
+            used_bytes: 0,
+            last_reset_epoch_secs,
+        }
+    }
+
+    pub fn user_quota_snapshot(&self) -> HashMap<String, UserQuotaSnapshot> {
+        let mut out = HashMap::new();
+        for entry in self.user_stats.iter() {
+            let stats = entry.value();
+            let used_bytes = stats.quota_used.load(Ordering::Relaxed);
+            let last_reset_epoch_secs = stats.quota_last_reset_epoch_secs.load(Ordering::Relaxed);
+            if used_bytes == 0 && last_reset_epoch_secs == 0 {
+                continue;
+            }
+            out.insert(
+                entry.key().clone(),
+                UserQuotaSnapshot {
+                    used_bytes,
+                    last_reset_epoch_secs,
+                },
+            );
+        }
+        out
+    }
+
+    pub fn get_handshake_timeouts(&self) -> u64 {
+        self.handshake_timeouts.load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_attempt_total(&self) -> u64 {
+        self.upstream_connect_attempt_total.load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_success_total(&self) -> u64 {
+        self.upstream_connect_success_total.load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_fail_total(&self) -> u64 {
+        self.upstream_connect_fail_total.load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_failfast_hard_error_total(&self) -> u64 {
+        self.upstream_connect_failfast_hard_error_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_attempts_bucket_1(&self) -> u64 {
+        self.upstream_connect_attempts_bucket_1
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_attempts_bucket_2(&self) -> u64 {
+        self.upstream_connect_attempts_bucket_2
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_attempts_bucket_3_4(&self) -> u64 {
+        self.upstream_connect_attempts_bucket_3_4
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_attempts_bucket_gt_4(&self) -> u64 {
+        self.upstream_connect_attempts_bucket_gt_4
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_duration_success_bucket_le_100ms(&self) -> u64 {
+        self.upstream_connect_duration_success_bucket_le_100ms
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_duration_success_bucket_101_500ms(&self) -> u64 {
+        self.upstream_connect_duration_success_bucket_101_500ms
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_duration_success_bucket_501_1000ms(&self) -> u64 {
+        self.upstream_connect_duration_success_bucket_501_1000ms
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_duration_success_bucket_gt_1000ms(&self) -> u64 {
+        self.upstream_connect_duration_success_bucket_gt_1000ms
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_duration_fail_bucket_le_100ms(&self) -> u64 {
+        self.upstream_connect_duration_fail_bucket_le_100ms
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_duration_fail_bucket_101_500ms(&self) -> u64 {
+        self.upstream_connect_duration_fail_bucket_101_500ms
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_duration_fail_bucket_501_1000ms(&self) -> u64 {
+        self.upstream_connect_duration_fail_bucket_501_1000ms
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_upstream_connect_duration_fail_bucket_gt_1000ms(&self) -> u64 {
+        self.upstream_connect_duration_fail_bucket_gt_1000ms
+            .load(Ordering::Relaxed)
+    }
+
+    pub fn iter_user_stats(&self) -> dashmap::iter::Iter<'_, String, Arc<UserStats>> {
+        self.user_stats.iter()
+    }
+
+    /// Current number of retained per-user stats entries.
+    pub fn user_stats_len(&self) -> usize {
+        self.user_stats.len()
+    }
+
+    pub fn uptime_secs(&self) -> f64 {
+        self.start_time
+            .read()
+            .map(|t| t.elapsed().as_secs_f64())
+            .unwrap_or(0.0)
+    }
+}

src/stats/writer_counters.rs

@@ -0,0 +1,542 @@
+use super::*;
+
+impl Stats {
+    pub fn increment_me_writer_pick_success_try_total(&self, mode: MeWriterPickMode) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match mode {
+            MeWriterPickMode::SortedRr => {
+                self.me_writer_pick_sorted_rr_success_try_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeWriterPickMode::P2c => {
+                self.me_writer_pick_p2c_success_try_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_me_writer_pick_success_fallback_total(&self, mode: MeWriterPickMode) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match mode {
+            MeWriterPickMode::SortedRr => {
+                self.me_writer_pick_sorted_rr_success_fallback_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeWriterPickMode::P2c => {
+                self.me_writer_pick_p2c_success_fallback_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_me_writer_pick_full_total(&self, mode: MeWriterPickMode) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match mode {
+            MeWriterPickMode::SortedRr => {
+                self.me_writer_pick_sorted_rr_full_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeWriterPickMode::P2c => {
+                self.me_writer_pick_p2c_full_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_me_writer_pick_closed_total(&self, mode: MeWriterPickMode) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match mode {
+            MeWriterPickMode::SortedRr => {
+                self.me_writer_pick_sorted_rr_closed_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeWriterPickMode::P2c => {
+                self.me_writer_pick_p2c_closed_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_me_writer_pick_no_candidate_total(&self, mode: MeWriterPickMode) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match mode {
+            MeWriterPickMode::SortedRr => {
+                self.me_writer_pick_sorted_rr_no_candidate_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            MeWriterPickMode::P2c => {
+                self.me_writer_pick_p2c_no_candidate_total
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_me_writer_pick_blocking_fallback_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_writer_pick_blocking_fallback_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_writer_pick_mode_switch_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_writer_pick_mode_switch_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_socks_kdf_strict_reject(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_socks_kdf_strict_reject
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_socks_kdf_compat_fallback(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_socks_kdf_compat_fallback
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_secure_padding_invalid(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.secure_padding_invalid.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_desync_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.desync_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_desync_full_logged(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.desync_full_logged.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_desync_suppressed(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.desync_suppressed.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn observe_desync_frames_ok(&self, frames_ok: u64) {
+        if !self.telemetry_me_allows_normal() {
+            return;
+        }
+        match frames_ok {
+            0 => {
+                self.desync_frames_bucket_0.fetch_add(1, Ordering::Relaxed);
+            }
+            1..=2 => {
+                self.desync_frames_bucket_1_2
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            3..=10 => {
+                self.desync_frames_bucket_3_10
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+            _ => {
+                self.desync_frames_bucket_gt_10
+                    .fetch_add(1, Ordering::Relaxed);
+            }
+        }
+    }
+    pub fn increment_pool_swap_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.pool_swap_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_pool_drain_active(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.pool_drain_active.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn decrement_pool_drain_active(&self) {
+        if !self.telemetry_me_allows_debug() {
+            return;
+        }
+        let mut current = self.pool_drain_active.load(Ordering::Relaxed);
+        loop {
+            if current == 0 {
+                break;
+            }
+            match self.pool_drain_active.compare_exchange_weak(
+                current,
+                current - 1,
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => break,
+                Err(actual) => current = actual,
+            }
+        }
+    }
+    pub fn increment_pool_force_close_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.pool_force_close_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_pool_stale_pick_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.pool_stale_pick_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_writer_removed_total(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_writer_removed_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_writer_removed_unexpected_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_writer_removed_unexpected_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_refill_triggered_total(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_refill_triggered_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_refill_skipped_inflight_total(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_refill_skipped_inflight_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_refill_failed_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_refill_failed_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_writer_restored_same_endpoint_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_writer_restored_same_endpoint_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_writer_restored_fallback_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_writer_restored_fallback_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_no_writer_failfast_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_no_writer_failfast_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_hybrid_timeout_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_hybrid_timeout_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_async_recovery_trigger_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_async_recovery_trigger_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_inline_recovery_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_inline_recovery_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_ip_reservation_rollback_tcp_limit_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.ip_reservation_rollback_tcp_limit_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_ip_reservation_rollback_quota_limit_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.ip_reservation_rollback_quota_limit_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn add_quota_refund_bytes_total(&self, bytes: u64) {
+        if self.telemetry_core_enabled() {
+            self.quota_refund_bytes_total
+                .fetch_add(bytes, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_quota_contention_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.quota_contention_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_quota_contention_timeout_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.quota_contention_timeout_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_quota_acquire_cancelled_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.quota_acquire_cancelled_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn add_quota_write_fail_bytes_total(&self, bytes: u64) {
+        if self.telemetry_core_enabled() {
+            self.quota_write_fail_bytes_total
+                .fetch_add(bytes, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_quota_write_fail_events_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.quota_write_fail_events_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_child_join_timeout_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.me_child_join_timeout_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_child_abort_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.me_child_abort_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn observe_flow_wait_middle_rate_limit_ms(&self, wait_ms: u64) {
+        if self.telemetry_core_enabled() {
+            self.flow_wait_middle_rate_limit_total
+                .fetch_add(1, Ordering::Relaxed);
+            self.flow_wait_middle_rate_limit_ms_total
+                .fetch_add(wait_ms, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_flow_wait_middle_rate_limit_cancelled_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.flow_wait_middle_rate_limit_cancelled_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_session_drop_fallback_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.session_drop_fallback_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_endpoint_quarantine_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_endpoint_quarantine_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_endpoint_quarantine_unexpected_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_endpoint_quarantine_unexpected_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_endpoint_quarantine_draining_suppressed_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_endpoint_quarantine_draining_suppressed_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_kdf_drift_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_kdf_drift_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_kdf_port_only_drift_total(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_kdf_port_only_drift_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_hardswap_pending_reuse_total(&self) {
+        if self.telemetry_me_allows_debug() {
+            self.me_hardswap_pending_reuse_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_hardswap_pending_ttl_expired_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_hardswap_pending_ttl_expired_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_single_endpoint_outage_enter_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_single_endpoint_outage_enter_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_single_endpoint_outage_exit_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_single_endpoint_outage_exit_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_single_endpoint_outage_reconnect_attempt_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_single_endpoint_outage_reconnect_attempt_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_single_endpoint_outage_reconnect_success_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_single_endpoint_outage_reconnect_success_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_single_endpoint_quarantine_bypass_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_single_endpoint_quarantine_bypass_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_single_endpoint_shadow_rotate_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_single_endpoint_shadow_rotate_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_single_endpoint_shadow_rotate_skipped_quarantine_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_single_endpoint_shadow_rotate_skipped_quarantine_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_floor_mode_switch_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_mode_switch_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_floor_mode_switch_static_to_adaptive_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_mode_switch_static_to_adaptive_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_floor_mode_switch_adaptive_to_static_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_mode_switch_adaptive_to_static_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_cpu_cores_detected_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_cpu_cores_detected_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_cpu_cores_effective_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_cpu_cores_effective_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_global_cap_raw_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_global_cap_raw_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_global_cap_effective_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_global_cap_effective_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_target_writers_total_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_target_writers_total_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_active_cap_configured_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_active_cap_configured_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_active_cap_effective_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_active_cap_effective_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_warm_cap_configured_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_warm_cap_configured_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_floor_warm_cap_effective_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_warm_cap_effective_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_writers_active_current_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_writers_active_current_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+    pub fn set_me_writers_warm_current_gauge(&self, value: u64) {
+        if self.telemetry_me_allows_normal() {
+            self.me_writers_warm_current_gauge
+                .store(value, Ordering::Relaxed);
+        }
+    }
+
+    pub fn set_buffer_pool_gauges(&self, pooled: usize, allocated: usize, in_use: usize) {
+        if self.telemetry_me_allows_normal() {
+            self.buffer_pool_pooled_gauge
+                .store(pooled as u64, Ordering::Relaxed);
+            self.buffer_pool_allocated_gauge
+                .store(allocated as u64, Ordering::Relaxed);
+            self.buffer_pool_in_use_gauge
+                .store(in_use as u64, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_me_c2me_send_full_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_c2me_send_full_total.fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_me_c2me_send_high_water_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_c2me_send_high_water_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+
+    pub fn increment_me_c2me_send_timeout_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_c2me_send_timeout_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_floor_cap_block_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_cap_block_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_floor_swap_idle_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_swap_idle_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_me_floor_swap_idle_failed_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.me_floor_swap_idle_failed_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+}

src/transport/middle_proxy/codec.rs

@@ -4,14 +4,32 @@ use tokio::io::{AsyncReadExt, AsyncWriteExt};
 use crate::crypto::{AesCbc, crc32, crc32c};
 use crate::error::{ProxyError, Result};
 use crate::protocol::constants::*;
+use crate::stream::PooledBuffer;
+
+use super::wire::{append_proxy_req_payload_into, proxy_req_payload_len};
+
+const RPC_WRITER_FRAME_BUF_SHRINK_THRESHOLD: usize = 256 * 1024;
+const RPC_WRITER_FRAME_BUF_RETAIN: usize = 64 * 1024;
 
 /// Commands sent to dedicated writer tasks to avoid mutex contention on TCP writes.
 pub(crate) enum WriterCommand {
     Data(Bytes),
     DataAndFlush(Bytes),
+    ProxyReq(ProxyReqCommand),
+    ControlAndFlush([u8; 12]),
     Close,
 }
 
+/// Structured proxy request command that lets the writer encode directly into its frame buffer.
+pub(crate) struct ProxyReqCommand {
+    pub(crate) conn_id: u64,
+    pub(crate) client_addr: std::net::SocketAddr,
+    pub(crate) our_addr: std::net::SocketAddr,
+    pub(crate) proto_flags: u32,
+    pub(crate) proxy_tag: Option<[u8; 16]>,
+    pub(crate) payload: PooledBuffer,
+}
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub(crate) enum RpcChecksumMode {
     Crc32,
@@ -42,15 +60,35 @@ pub(crate) fn rpc_crc(mode: RpcChecksumMode, data: &[u8]) -> u32 {
     }
 }
 
+/// Builds a fixed-size control payload without heap allocation.
+pub(crate) fn build_control_payload(tag: u32, value: u64) -> [u8; 12] {
+    let mut payload = [0u8; 12];
+    payload[..4].copy_from_slice(&tag.to_le_bytes());
+    payload[4..].copy_from_slice(&value.to_le_bytes());
+    payload
+}
+
 pub(crate) fn build_rpc_frame(seq_no: i32, payload: &[u8], crc_mode: RpcChecksumMode) -> Vec<u8> {
-    let total_len = (4 + 4 + payload.len() + 4) as u32;
-    let mut frame = Vec::with_capacity(total_len as usize);
+    let mut frame = Vec::new();
+    build_rpc_frame_into(&mut frame, seq_no, payload, crc_mode);
+    frame
+}
+
+fn build_rpc_frame_into(
+    frame: &mut Vec<u8>,
+    seq_no: i32,
+    payload: &[u8],
+    crc_mode: RpcChecksumMode,
+) {
+    let total_len = 4 + 4 + payload.len() + 4;
+    frame.clear();
+    frame.reserve(total_len + 15);
+    let total_len = total_len as u32;
     frame.extend_from_slice(&total_len.to_le_bytes());
     frame.extend_from_slice(&seq_no.to_le_bytes());
     frame.extend_from_slice(payload);
     let c = rpc_crc(crc_mode, &frame);
     frame.extend_from_slice(&c.to_le_bytes());
-    frame
 }
 
 pub(crate) async fn read_rpc_frame_plaintext(
@@ -218,29 +256,65 @@ pub(crate) struct RpcWriter {
     pub(crate) iv: [u8; 16],
     pub(crate) seq_no: i32,
     pub(crate) crc_mode: RpcChecksumMode,
+    pub(crate) frame_buf: Vec<u8>,
 }
 
 impl RpcWriter {
     pub(crate) async fn send(&mut self, payload: &[u8]) -> Result<()> {
-        let frame = build_rpc_frame(self.seq_no, payload, self.crc_mode);
+        build_rpc_frame_into(&mut self.frame_buf, self.seq_no, payload, self.crc_mode);
         self.seq_no = self.seq_no.wrapping_add(1);
+        self.encrypt_and_write_frame().await
+    }
 
-        let pad = (16 - (frame.len() % 16)) % 16;
-        let mut buf = frame;
+    pub(crate) async fn send_proxy_req(&mut self, command: &ProxyReqCommand) -> Result<()> {
+        let payload_len = proxy_req_payload_len(
+            command.payload.len(),
+            command.proxy_tag.as_ref().map(|tag| tag.as_slice()),
+            command.proto_flags,
+        );
+        let total_len = 4 + 4 + payload_len + 4;
+        self.frame_buf.clear();
+        self.frame_buf.reserve(total_len + 15);
+        self.frame_buf
+            .extend_from_slice(&(total_len as u32).to_le_bytes());
+        self.frame_buf.extend_from_slice(&self.seq_no.to_le_bytes());
+        append_proxy_req_payload_into(
+            &mut self.frame_buf,
+            command.conn_id,
+            command.client_addr,
+            command.our_addr,
+            command.payload.as_ref(),
+            command.proxy_tag.as_ref().map(|tag| tag.as_slice()),
+            command.proto_flags,
+        );
+        let c = rpc_crc(self.crc_mode, &self.frame_buf);
+        self.frame_buf.extend_from_slice(&c.to_le_bytes());
+        self.seq_no = self.seq_no.wrapping_add(1);
+        self.encrypt_and_write_frame().await
+    }
+
+    async fn encrypt_and_write_frame(&mut self) -> Result<()> {
+        let pad = (16 - (self.frame_buf.len() % 16)) % 16;
         let pad_pattern: [u8; 4] = [0x04, 0x00, 0x00, 0x00];
         for i in 0..pad {
-            buf.push(pad_pattern[i % 4]);
+            self.frame_buf.push(pad_pattern[i % 4]);
         }
 
         let cipher = AesCbc::new(self.key, self.iv);
         cipher
-            .encrypt_in_place(&mut buf)
+            .encrypt_in_place(&mut self.frame_buf)
             .map_err(|e| ProxyError::Crypto(format!("{e}")))?;
 
-        if buf.len() >= 16 {
-            self.iv.copy_from_slice(&buf[buf.len() - 16..]);
+        if self.frame_buf.len() >= 16 {
+            self.iv
+                .copy_from_slice(&self.frame_buf[self.frame_buf.len() - 16..]);
         }
-        self.writer.write_all(&buf).await.map_err(ProxyError::Io)
+        let write_result = self.writer.write_all(&self.frame_buf).await;
+        if self.frame_buf.capacity() > RPC_WRITER_FRAME_BUF_SHRINK_THRESHOLD {
+            self.frame_buf.clear();
+            self.frame_buf.shrink_to(RPC_WRITER_FRAME_BUF_RETAIN);
+        }
+        write_result.map_err(ProxyError::Io)
     }
 
     pub(crate) async fn send_and_flush(&mut self, payload: &[u8]) -> Result<()> {

src/transport/middle_proxy/pool_init.rs

@@ -19,12 +19,14 @@ impl MePool {
             .me_reconnect_max_concurrent_per_dc
             .max(1) as usize;
         let ks = self.key_selector().await;
+        let me_servers = self.proxy_map_v4.read().await.len();
+        let secret_len = self.proxy_secret.read().await.secret.len();
         info!(
-            me_servers = self.proxy_map_v4.read().await.len(),
+            me_servers,
             pool_size,
             connect_concurrency,
             key_selector = format_args!("0x{ks:08x}"),
-            secret_len = self.proxy_secret.read().await.secret.len(),
+            secret_len,
             "Initializing ME pool"
         );
 

src/transport/middle_proxy/pool_writer.rs

@@ -5,7 +5,6 @@ use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, AtomicU8, AtomicU32, AtomicU64, Ordering};
 use std::time::{Duration, Instant};
 
-use bytes::Bytes;
 use bytes::BytesMut;
 use rand::RngExt;
 use tokio::sync::mpsc;
@@ -17,7 +16,7 @@ use crate::crypto::SecureRandom;
 use crate::error::{ProxyError, Result};
 use crate::protocol::constants::{RPC_CLOSE_EXT_U32, RPC_PING_U32};
 
-use super::codec::{RpcWriter, WriterCommand};
+use super::codec::{RpcWriter, WriterCommand, build_control_payload};
 use super::pool::{MePool, MeWriter, WriterContour};
 use super::reader::reader_loop;
 use super::wire::build_proxy_req_payload;
@@ -61,6 +60,12 @@ async fn writer_command_loop(
                     Some(WriterCommand::DataAndFlush(payload)) => {
                         rpc_writer.send_and_flush(&payload).await?;
                     }
+                    Some(WriterCommand::ProxyReq(command)) => {
+                        rpc_writer.send_proxy_req(&command).await?;
+                    }
+                    Some(WriterCommand::ControlAndFlush(payload)) => {
+                        rpc_writer.send_and_flush(&payload).await?;
+                    }
                     Some(WriterCommand::Close) | None => return Ok(()),
                 }
             }
@@ -130,9 +135,7 @@ async fn ping_loop(
             _ = tokio::time::sleep(wait) => {}
         }
         let sent_id = ping_id;
-        let mut p = Vec::with_capacity(12);
-        p.extend_from_slice(&RPC_PING_U32.to_le_bytes());
-        p.extend_from_slice(&sent_id.to_le_bytes());
+        let payload = build_control_payload(RPC_PING_U32, sent_id as u64);
         {
             let mut tracker = ping_tracker_ping.lock().await;
             cleanup_tick = cleanup_tick.wrapping_add(1);
@@ -149,7 +152,7 @@ async fn ping_loop(
         ping_id = ping_id.wrapping_add(1);
         stats_ping.increment_me_keepalive_sent();
         if tx_ping
-            .send(WriterCommand::DataAndFlush(Bytes::from(p)))
+            .send(WriterCommand::ControlAndFlush(payload))
             .await
             .is_err()
         {
@@ -253,12 +256,10 @@ async fn rpc_proxy_req_signal_loop(
             stats_signal.increment_me_rpc_proxy_req_signal_response_total();
         }
 
-        let mut close_payload = Vec::with_capacity(12);
-        close_payload.extend_from_slice(&RPC_CLOSE_EXT_U32.to_le_bytes());
-        close_payload.extend_from_slice(&conn_id.to_le_bytes());
+        let close_payload = build_control_payload(RPC_CLOSE_EXT_U32, conn_id);
 
         if tx_signal
-            .send(WriterCommand::DataAndFlush(Bytes::from(close_payload)))
+            .send(WriterCommand::ControlAndFlush(close_payload))
             .await
             .is_err()
         {
@@ -380,6 +381,7 @@ impl MePool {
             iv: hs.write_iv,
             seq_no: 0,
             crc_mode: hs.crc_mode,
+            frame_buf: Vec::new(),
         };
         let writer = MeWriter {
             id: writer_id,

src/transport/middle_proxy/reader.rs

@@ -19,7 +19,7 @@ use crate::error::{ProxyError, Result};
 use crate::protocol::constants::*;
 use crate::stats::Stats;
 
-use super::codec::{RpcChecksumMode, WriterCommand, rpc_crc};
+use super::codec::{RpcChecksumMode, WriterCommand, build_control_payload, rpc_crc};
 use super::fairness::{
     AdmissionDecision, DispatchAction, DispatchFeedback, PressureState, SchedulerDecision,
     WorkerFairnessConfig, WorkerFairnessSnapshot, WorkerFairnessState,
@@ -464,10 +464,8 @@ pub(crate) async fn reader_loop(
             } else if pt == RPC_PING_U32 && body.len() >= 8 {
                 let ping_id = i64::from_le_bytes(body[0..8].try_into().unwrap());
                 trace!(ping_id, "RPC_PING -> RPC_PONG");
-                let mut pong = Vec::with_capacity(12);
-                pong.extend_from_slice(&RPC_PONG_U32.to_le_bytes());
-                pong.extend_from_slice(&ping_id.to_le_bytes());
-                match tx.try_send(WriterCommand::DataAndFlush(Bytes::from(pong))) {
+                let pong = build_control_payload(RPC_PONG_U32, ping_id as u64);
+                match tx.try_send(WriterCommand::ControlAndFlush(pong)) {
                     Ok(()) => {}
                     Err(TrySendError::Full(_)) => {
                         debug!(ping_id, "PONG dropped: writer command channel is full");
@@ -667,10 +665,8 @@ mod tests {
 }
 
 async fn send_close_conn(tx: &mpsc::Sender<WriterCommand>, conn_id: u64) {
-    let mut p = Vec::with_capacity(12);
-    p.extend_from_slice(&RPC_CLOSE_CONN_U32.to_le_bytes());
-    p.extend_from_slice(&conn_id.to_le_bytes());
-    match tx.try_send(WriterCommand::DataAndFlush(Bytes::from(p))) {
+    let payload = build_control_payload(RPC_CLOSE_CONN_U32, conn_id);
+    match tx.try_send(WriterCommand::ControlAndFlush(payload)) {
         Ok(()) => {}
         Err(TrySendError::Full(_)) => {
             debug!(

src/transport/middle_proxy/registry.rs

@@ -2,14 +2,13 @@ use std::collections::{HashMap, HashSet};
 use std::net::SocketAddr;
 use std::sync::Arc;
 use std::sync::atomic::{AtomicU8, AtomicU64, Ordering};
-use std::time::{Duration, SystemTime, UNIX_EPOCH};
+use std::time::{SystemTime, UNIX_EPOCH};
 
 use dashmap::DashMap;
-use tokio::sync::mpsc::error::TrySendError;
 use tokio::sync::{Mutex, Semaphore, mpsc};
 
+use super::MeResponse;
 use super::codec::WriterCommand;
-use super::{MeResponse, RouteBytePermit};
 
 const ROUTE_BACKPRESSURE_BASE_TIMEOUT_MS: u64 = 25;
 const ROUTE_BACKPRESSURE_HIGH_TIMEOUT_MS: u64 = 120;
@@ -18,6 +17,8 @@ const ROUTE_QUEUED_BYTE_PERMIT_UNIT: usize = 16 * 1024;
 const ROUTE_QUEUED_PERMITS_PER_SLOT: usize = 4;
 const ROUTE_QUEUED_MAX_FRAME_PERMITS: usize = 1024;
 
+mod writer;
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum RouteResult {
     Routed,
@@ -218,737 +219,7 @@ impl ConnRegistry {
         );
         (id, rx)
     }
-
-    pub async fn register_writer(&self, writer_id: u64, tx: mpsc::Sender<WriterCommand>) {
-        let mut binding = self.binding.inner.lock().await;
-        binding.writers.insert(writer_id, tx.clone());
-        binding
-            .conns_for_writer
-            .entry(writer_id)
-            .or_insert_with(HashSet::new);
-        self.writers.map.insert(writer_id, tx);
-    }
-
-    /// Unregister connection, returning associated writer_id if any.
-    pub async fn unregister(&self, id: u64) -> Option<u64> {
-        self.routing.map.remove(&id);
-        self.routing.byte_budget.remove(&id);
-        self.hot_binding.map.remove(&id);
-        let mut binding = self.binding.inner.lock().await;
-        binding.meta.remove(&id);
-        if let Some(writer_id) = binding.writer_for_conn.remove(&id) {
-            let became_empty = if let Some(set) = binding.conns_for_writer.get_mut(&writer_id) {
-                set.remove(&id);
-                set.is_empty()
-            } else {
-                false
-            };
-            if became_empty {
-                binding
-                    .writer_idle_since_epoch_secs
-                    .insert(writer_id, Self::now_epoch_secs());
-            }
-            return Some(writer_id);
-        }
-        None
-    }
-
-    async fn attach_route_byte_permit(
-        &self,
-        id: u64,
-        resp: MeResponse,
-        timeout_ms: Option<u64>,
-    ) -> std::result::Result<MeResponse, RouteResult> {
-        let MeResponse::Data {
-            flags,
-            data,
-            route_permit,
-        } = resp
-        else {
-            return Ok(resp);
-        };
-
-        if route_permit.is_some() {
-            return Ok(MeResponse::Data {
-                flags,
-                data,
-                route_permit,
-            });
-        }
-
-        let Some(semaphore) = self
-            .routing
-            .byte_budget
-            .get(&id)
-            .map(|entry| entry.value().clone())
-        else {
-            return Err(RouteResult::NoConn);
-        };
-        let permits = Self::route_data_permits(data.len());
-        let permit = match timeout_ms {
-            Some(0) => semaphore
-                .try_acquire_many_owned(permits)
-                .map_err(|_| RouteResult::QueueFullHigh)?,
-            Some(timeout_ms) => {
-                let acquire = semaphore.acquire_many_owned(permits);
-                match tokio::time::timeout(Duration::from_millis(timeout_ms.max(1)), acquire).await
-                {
-                    Ok(Ok(permit)) => permit,
-                    Ok(Err(_)) => return Err(RouteResult::ChannelClosed),
-                    Err(_) => return Err(RouteResult::QueueFullHigh),
-                }
-            }
-            None => semaphore
-                .acquire_many_owned(permits)
-                .await
-                .map_err(|_| RouteResult::ChannelClosed)?,
-        };
-
-        Ok(MeResponse::Data {
-            flags,
-            data,
-            route_permit: Some(RouteBytePermit::new(permit)),
-        })
-    }
-
-    #[allow(dead_code)]
-    pub async fn route(&self, id: u64, resp: MeResponse) -> RouteResult {
-        let tx = self.routing.map.get(&id).map(|entry| entry.value().clone());
-
-        let Some(tx) = tx else {
-            return RouteResult::NoConn;
-        };
-
-        let base_timeout_ms = self
-            .route_backpressure_base_timeout_ms
-            .load(Ordering::Relaxed)
-            .max(1);
-        let resp = match self
-            .attach_route_byte_permit(id, resp, Some(base_timeout_ms))
-            .await
-        {
-            Ok(resp) => resp,
-            Err(result) => return result,
-        };
-
-        match tx.try_send(resp) {
-            Ok(()) => RouteResult::Routed,
-            Err(TrySendError::Closed(_)) => RouteResult::ChannelClosed,
-            Err(TrySendError::Full(resp)) => {
-                // Absorb short bursts without dropping/closing the session immediately.
-                let high_timeout_ms = self
-                    .route_backpressure_high_timeout_ms
-                    .load(Ordering::Relaxed)
-                    .max(base_timeout_ms);
-                let high_watermark_pct = self
-                    .route_backpressure_high_watermark_pct
-                    .load(Ordering::Relaxed)
-                    .clamp(1, 100);
-                let used = self.route_channel_capacity.saturating_sub(tx.capacity());
-                let used_pct = if self.route_channel_capacity == 0 {
-                    100
-                } else {
-                    (used.saturating_mul(100) / self.route_channel_capacity) as u8
-                };
-                let high_profile = used_pct >= high_watermark_pct;
-                let timeout_ms = if high_profile {
-                    high_timeout_ms
-                } else {
-                    base_timeout_ms
-                };
-                let timeout_dur = Duration::from_millis(timeout_ms);
-
-                match tokio::time::timeout(timeout_dur, tx.send(resp)).await {
-                    Ok(Ok(())) => RouteResult::Routed,
-                    Ok(Err(_)) => RouteResult::ChannelClosed,
-                    Err(_) => {
-                        if high_profile {
-                            RouteResult::QueueFullHigh
-                        } else {
-                            RouteResult::QueueFullBase
-                        }
-                    }
-                }
-            }
-        }
-    }
-
-    pub async fn route_nowait(&self, id: u64, resp: MeResponse) -> RouteResult {
-        let tx = self.routing.map.get(&id).map(|entry| entry.value().clone());
-
-        let Some(tx) = tx else {
-            return RouteResult::NoConn;
-        };
-        let resp = match self.attach_route_byte_permit(id, resp, Some(0)).await {
-            Ok(resp) => resp,
-            Err(result) => return result,
-        };
-
-        match tx.try_send(resp) {
-            Ok(()) => RouteResult::Routed,
-            Err(TrySendError::Closed(_)) => RouteResult::ChannelClosed,
-            Err(TrySendError::Full(_)) => RouteResult::QueueFullBase,
-        }
-    }
-
-    pub async fn route_with_timeout(
-        &self,
-        id: u64,
-        resp: MeResponse,
-        timeout_ms: u64,
-    ) -> RouteResult {
-        if timeout_ms == 0 {
-            return self.route_nowait(id, resp).await;
-        }
-
-        let tx = self.routing.map.get(&id).map(|entry| entry.value().clone());
-
-        let Some(tx) = tx else {
-            return RouteResult::NoConn;
-        };
-        let resp = match self
-            .attach_route_byte_permit(id, resp, Some(timeout_ms))
-            .await
-        {
-            Ok(resp) => resp,
-            Err(result) => return result,
-        };
-
-        match tx.try_send(resp) {
-            Ok(()) => RouteResult::Routed,
-            Err(TrySendError::Closed(_)) => RouteResult::ChannelClosed,
-            Err(TrySendError::Full(resp)) => {
-                let high_watermark_pct = self
-                    .route_backpressure_high_watermark_pct
-                    .load(Ordering::Relaxed)
-                    .clamp(1, 100);
-                let used = self.route_channel_capacity.saturating_sub(tx.capacity());
-                let used_pct = if self.route_channel_capacity == 0 {
-                    100
-                } else {
-                    (used.saturating_mul(100) / self.route_channel_capacity) as u8
-                };
-                let high_profile = used_pct >= high_watermark_pct;
-                let timeout_dur = Duration::from_millis(timeout_ms.max(1));
-
-                match tokio::time::timeout(timeout_dur, tx.send(resp)).await {
-                    Ok(Ok(())) => RouteResult::Routed,
-                    Ok(Err(_)) => RouteResult::ChannelClosed,
-                    Err(_) => {
-                        if high_profile {
-                            RouteResult::QueueFullHigh
-                        } else {
-                            RouteResult::QueueFullBase
-                        }
-                    }
-                }
-            }
-        }
-    }
-
-    pub async fn bind_writer(&self, conn_id: u64, writer_id: u64, meta: ConnMeta) -> bool {
-        let mut binding = self.binding.inner.lock().await;
-        // ROUTING IS THE SOURCE OF TRUTH:
-        // never keep/attach writer binding for a connection that is already
-        // absent from the routing table.
-        if !self.routing.map.contains_key(&conn_id) {
-            return false;
-        }
-        if !binding.writers.contains_key(&writer_id) {
-            return false;
-        }
-
-        let previous_writer_id = binding.writer_for_conn.insert(conn_id, writer_id);
-        if let Some(previous_writer_id) = previous_writer_id
-            && previous_writer_id != writer_id
-        {
-            let became_empty =
-                if let Some(set) = binding.conns_for_writer.get_mut(&previous_writer_id) {
-                    set.remove(&conn_id);
-                    set.is_empty()
-                } else {
-                    false
-                };
-            if became_empty {
-                binding
-                    .writer_idle_since_epoch_secs
-                    .insert(previous_writer_id, Self::now_epoch_secs());
-            }
-        }
-
-        binding.meta.insert(conn_id, meta.clone());
-        binding.last_meta_for_writer.insert(writer_id, meta.clone());
-        binding.writer_idle_since_epoch_secs.remove(&writer_id);
-        binding
-            .conns_for_writer
-            .entry(writer_id)
-            .or_insert_with(HashSet::new)
-            .insert(conn_id);
-        self.hot_binding
-            .map
-            .insert(conn_id, HotConnBinding { writer_id, meta });
-        true
-    }
-
-    pub async fn mark_writer_idle(&self, writer_id: u64) {
-        let mut binding = self.binding.inner.lock().await;
-        binding
-            .conns_for_writer
-            .entry(writer_id)
-            .or_insert_with(HashSet::new);
-        binding
-            .writer_idle_since_epoch_secs
-            .entry(writer_id)
-            .or_insert(Self::now_epoch_secs());
-    }
-
-    pub async fn get_last_writer_meta(&self, writer_id: u64) -> Option<ConnMeta> {
-        let binding = self.binding.inner.lock().await;
-        binding.last_meta_for_writer.get(&writer_id).cloned()
-    }
-
-    pub async fn writer_idle_since_snapshot(&self) -> HashMap<u64, u64> {
-        let binding = self.binding.inner.lock().await;
-        binding.writer_idle_since_epoch_secs.clone()
-    }
-
-    pub async fn writer_idle_since_for_writer_ids(&self, writer_ids: &[u64]) -> HashMap<u64, u64> {
-        let binding = self.binding.inner.lock().await;
-        let mut out = HashMap::<u64, u64>::with_capacity(writer_ids.len());
-        for writer_id in writer_ids {
-            if let Some(idle_since) = binding.writer_idle_since_epoch_secs.get(writer_id).copied() {
-                out.insert(*writer_id, idle_since);
-            }
-        }
-        out
-    }
-
-    pub(super) async fn writer_activity_snapshot(&self) -> WriterActivitySnapshot {
-        let binding = self.binding.inner.lock().await;
-        let mut bound_clients_by_writer = HashMap::<u64, usize>::new();
-        let mut active_sessions_by_target_dc = HashMap::<i16, usize>::new();
-
-        for (writer_id, conn_ids) in &binding.conns_for_writer {
-            bound_clients_by_writer.insert(*writer_id, conn_ids.len());
-        }
-        for conn_meta in binding.meta.values() {
-            if conn_meta.target_dc == 0 {
-                continue;
-            }
-            *active_sessions_by_target_dc
-                .entry(conn_meta.target_dc)
-                .or_insert(0) += 1;
-        }
-
-        WriterActivitySnapshot {
-            bound_clients_by_writer,
-            active_sessions_by_target_dc,
-        }
-    }
-
-    pub async fn get_writer(&self, conn_id: u64) -> Option<ConnWriter> {
-        if !self.routing.map.contains_key(&conn_id) {
-            return None;
-        }
-
-        let writer_id = self
-            .hot_binding
-            .map
-            .get(&conn_id)
-            .map(|entry| entry.writer_id)?;
-        let writer = self
-            .writers
-            .map
-            .get(&writer_id)
-            .map(|entry| entry.value().clone())?;
-        Some(ConnWriter {
-            writer_id,
-            tx: writer,
-        })
-    }
-
-    pub async fn active_conn_ids(&self) -> Vec<u64> {
-        let binding = self.binding.inner.lock().await;
-        binding.writer_for_conn.keys().copied().collect()
-    }
-
-    pub async fn writer_lost(&self, writer_id: u64) -> Vec<BoundConn> {
-        let mut binding = self.binding.inner.lock().await;
-        binding.writers.remove(&writer_id);
-        self.writers.map.remove(&writer_id);
-        binding.last_meta_for_writer.remove(&writer_id);
-        binding.writer_idle_since_epoch_secs.remove(&writer_id);
-        let conns = binding
-            .conns_for_writer
-            .remove(&writer_id)
-            .unwrap_or_default()
-            .into_iter()
-            .collect::<Vec<_>>();
-
-        let mut out = Vec::new();
-        for conn_id in conns {
-            if binding.writer_for_conn.get(&conn_id).copied() != Some(writer_id) {
-                continue;
-            }
-            binding.writer_for_conn.remove(&conn_id);
-            let remove_hot = self
-                .hot_binding
-                .map
-                .get(&conn_id)
-                .map(|hot| hot.writer_id == writer_id)
-                .unwrap_or(false);
-            if remove_hot {
-                self.hot_binding.map.remove(&conn_id);
-            }
-            if let Some(m) = binding.meta.get(&conn_id) {
-                out.push(BoundConn {
-                    conn_id,
-                    meta: m.clone(),
-                });
-            }
-        }
-        out
-    }
-
-    #[allow(dead_code)]
-    pub async fn get_meta(&self, conn_id: u64) -> Option<ConnMeta> {
-        self.hot_binding
-            .map
-            .get(&conn_id)
-            .map(|entry| entry.meta.clone())
-    }
-
-    pub async fn is_writer_empty(&self, writer_id: u64) -> bool {
-        let binding = self.binding.inner.lock().await;
-        binding
-            .conns_for_writer
-            .get(&writer_id)
-            .map(|s| s.is_empty())
-            .unwrap_or(true)
-    }
-
-    #[allow(dead_code)]
-    pub async fn unregister_writer_if_empty(&self, writer_id: u64) -> bool {
-        let mut binding = self.binding.inner.lock().await;
-        let Some(conn_ids) = binding.conns_for_writer.get(&writer_id) else {
-            // Writer is already absent from the registry.
-            return true;
-        };
-        if !conn_ids.is_empty() {
-            return false;
-        }
-
-        binding.writers.remove(&writer_id);
-        self.writers.map.remove(&writer_id);
-        binding.last_meta_for_writer.remove(&writer_id);
-        binding.writer_idle_since_epoch_secs.remove(&writer_id);
-        binding.conns_for_writer.remove(&writer_id);
-        true
-    }
-
-    #[allow(dead_code)]
-    pub(super) async fn non_empty_writer_ids(&self, writer_ids: &[u64]) -> HashSet<u64> {
-        let binding = self.binding.inner.lock().await;
-        let mut out = HashSet::<u64>::with_capacity(writer_ids.len());
-        for writer_id in writer_ids {
-            if let Some(conns) = binding.conns_for_writer.get(writer_id)
-                && !conns.is_empty()
-            {
-                out.insert(*writer_id);
-            }
-        }
-        out
-    }
 }
 
 #[cfg(test)]
-mod tests {
-    use std::net::{IpAddr, Ipv4Addr, SocketAddr};
-
-    use bytes::Bytes;
-
-    use super::{ConnMeta, ConnRegistry, RouteResult};
-    use crate::transport::middle_proxy::MeResponse;
-
-    #[tokio::test]
-    async fn writer_activity_snapshot_tracks_writer_and_dc_load() {
-        let registry = ConnRegistry::new();
-
-        let (conn_a, _rx_a) = registry.register().await;
-        let (conn_b, _rx_b) = registry.register().await;
-        let (conn_c, _rx_c) = registry.register().await;
-        let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
-        let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
-        registry.register_writer(10, writer_tx_a.clone()).await;
-        registry.register_writer(20, writer_tx_b.clone()).await;
-
-        let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
-        assert!(
-            registry
-                .bind_writer(
-                    conn_a,
-                    10,
-                    ConnMeta {
-                        target_dc: 2,
-                        client_addr: addr,
-                        our_addr: addr,
-                        proto_flags: 0,
-                    },
-                )
-                .await
-        );
-        assert!(
-            registry
-                .bind_writer(
-                    conn_b,
-                    10,
-                    ConnMeta {
-                        target_dc: -2,
-                        client_addr: addr,
-                        our_addr: addr,
-                        proto_flags: 0,
-                    },
-                )
-                .await
-        );
-        assert!(
-            registry
-                .bind_writer(
-                    conn_c,
-                    20,
-                    ConnMeta {
-                        target_dc: 4,
-                        client_addr: addr,
-                        our_addr: addr,
-                        proto_flags: 0,
-                    },
-                )
-                .await
-        );
-
-        let snapshot = registry.writer_activity_snapshot().await;
-        assert_eq!(snapshot.bound_clients_by_writer.get(&10), Some(&2));
-        assert_eq!(snapshot.bound_clients_by_writer.get(&20), Some(&1));
-        assert_eq!(snapshot.active_sessions_by_target_dc.get(&2), Some(&1));
-        assert_eq!(snapshot.active_sessions_by_target_dc.get(&-2), Some(&1));
-        assert_eq!(snapshot.active_sessions_by_target_dc.get(&4), Some(&1));
-    }
-
-    #[tokio::test]
-    async fn route_data_is_bounded_by_byte_permits_before_channel_capacity() {
-        let registry = ConnRegistry::with_route_byte_permits_for_tests(4, 1);
-        let (conn_id, mut rx) = registry.register().await;
-        let routed = registry
-            .route_nowait(
-                conn_id,
-                MeResponse::Data {
-                    flags: 0,
-                    data: Bytes::from_static(&[0xAA]),
-                    route_permit: None,
-                },
-            )
-            .await;
-        assert!(matches!(routed, RouteResult::Routed));
-
-        let blocked = registry
-            .route_nowait(
-                conn_id,
-                MeResponse::Data {
-                    flags: 0,
-                    data: Bytes::from_static(&[0xBB]),
-                    route_permit: None,
-                },
-            )
-            .await;
-        assert!(
-            matches!(blocked, RouteResult::QueueFullHigh),
-            "byte budget must reject data before count capacity is exhausted"
-        );
-
-        drop(rx.recv().await);
-
-        let routed_after_drain = registry
-            .route_nowait(
-                conn_id,
-                MeResponse::Data {
-                    flags: 0,
-                    data: Bytes::from_static(&[0xCC]),
-                    route_permit: None,
-                },
-            )
-            .await;
-        assert!(
-            matches!(routed_after_drain, RouteResult::Routed),
-            "receiving queued data must release byte permits"
-        );
-    }
-
-    #[tokio::test]
-    async fn bind_writer_rebinds_conn_atomically() {
-        let registry = ConnRegistry::new();
-        let (conn_id, _rx) = registry.register().await;
-        let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
-        let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
-        registry.register_writer(10, writer_tx_a).await;
-        registry.register_writer(20, writer_tx_b).await;
-
-        let client_addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
-        let first_our_addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::new(1, 1, 1, 1)), 443);
-        let second_our_addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::new(2, 2, 2, 2)), 443);
-
-        assert!(
-            registry
-                .bind_writer(
-                    conn_id,
-                    10,
-                    ConnMeta {
-                        target_dc: 2,
-                        client_addr,
-                        our_addr: first_our_addr,
-                        proto_flags: 1,
-                    },
-                )
-                .await
-        );
-        assert!(
-            registry
-                .bind_writer(
-                    conn_id,
-                    20,
-                    ConnMeta {
-                        target_dc: 2,
-                        client_addr,
-                        our_addr: second_our_addr,
-                        proto_flags: 2,
-                    },
-                )
-                .await
-        );
-
-        let writer = registry.get_writer(conn_id).await.expect("writer binding");
-        assert_eq!(writer.writer_id, 20);
-
-        let meta = registry.get_meta(conn_id).await.expect("conn meta");
-        assert_eq!(meta.our_addr, second_our_addr);
-        assert_eq!(meta.proto_flags, 2);
-
-        let snapshot = registry.writer_activity_snapshot().await;
-        assert_eq!(snapshot.bound_clients_by_writer.get(&10), Some(&0));
-        assert_eq!(snapshot.bound_clients_by_writer.get(&20), Some(&1));
-        assert!(
-            registry
-                .writer_idle_since_snapshot()
-                .await
-                .contains_key(&10)
-        );
-    }
-
-    #[tokio::test]
-    async fn writer_lost_does_not_drop_rebound_conn() {
-        let registry = ConnRegistry::new();
-        let (conn_id, _rx) = registry.register().await;
-        let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
-        let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
-        registry.register_writer(10, writer_tx_a).await;
-        registry.register_writer(20, writer_tx_b).await;
-
-        let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
-        assert!(
-            registry
-                .bind_writer(
-                    conn_id,
-                    10,
-                    ConnMeta {
-                        target_dc: 2,
-                        client_addr: addr,
-                        our_addr: addr,
-                        proto_flags: 0,
-                    },
-                )
-                .await
-        );
-        assert!(
-            registry
-                .bind_writer(
-                    conn_id,
-                    20,
-                    ConnMeta {
-                        target_dc: 2,
-                        client_addr: addr,
-                        our_addr: addr,
-                        proto_flags: 1,
-                    },
-                )
-                .await
-        );
-
-        let lost = registry.writer_lost(10).await;
-        assert!(lost.is_empty());
-        assert_eq!(
-            registry
-                .get_writer(conn_id)
-                .await
-                .expect("writer")
-                .writer_id,
-            20
-        );
-
-        let removed_writer = registry.unregister(conn_id).await;
-        assert_eq!(removed_writer, Some(20));
-        assert!(registry.is_writer_empty(20).await);
-    }
-
-    #[tokio::test]
-    async fn bind_writer_rejects_unregistered_writer() {
-        let registry = ConnRegistry::new();
-        let (conn_id, _rx) = registry.register().await;
-        let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
-
-        assert!(
-            !registry
-                .bind_writer(
-                    conn_id,
-                    10,
-                    ConnMeta {
-                        target_dc: 2,
-                        client_addr: addr,
-                        our_addr: addr,
-                        proto_flags: 0,
-                    },
-                )
-                .await
-        );
-        assert!(registry.get_writer(conn_id).await.is_none());
-    }
-
-    #[tokio::test]
-    async fn non_empty_writer_ids_returns_only_writers_with_bound_clients() {
-        let registry = ConnRegistry::new();
-        let (conn_id, _rx) = registry.register().await;
-        let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
-        let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
-        registry.register_writer(10, writer_tx_a).await;
-        registry.register_writer(20, writer_tx_b).await;
-
-        let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
-        assert!(
-            registry
-                .bind_writer(
-                    conn_id,
-                    10,
-                    ConnMeta {
-                        target_dc: 2,
-                        client_addr: addr,
-                        our_addr: addr,
-                        proto_flags: 0,
-                    },
-                )
-                .await
-        );
-
-        let non_empty = registry.non_empty_writer_ids(&[10, 20, 30]).await;
-        assert!(non_empty.contains(&10));
-        assert!(!non_empty.contains(&20));
-        assert!(!non_empty.contains(&30));
-    }
-}
+mod tests;

src/transport/middle_proxy/registry/tests.rs

@@ -0,0 +1,288 @@
+use std::net::{IpAddr, Ipv4Addr, SocketAddr};
+
+use bytes::Bytes;
+
+use super::{ConnMeta, ConnRegistry, RouteResult};
+use crate::transport::middle_proxy::MeResponse;
+
+#[tokio::test]
+async fn writer_activity_snapshot_tracks_writer_and_dc_load() {
+    let registry = ConnRegistry::new();
+
+    let (conn_a, _rx_a) = registry.register().await;
+    let (conn_b, _rx_b) = registry.register().await;
+    let (conn_c, _rx_c) = registry.register().await;
+    let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
+    let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
+    registry.register_writer(10, writer_tx_a.clone()).await;
+    registry.register_writer(20, writer_tx_b.clone()).await;
+
+    let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
+    assert!(
+        registry
+            .bind_writer(
+                conn_a,
+                10,
+                ConnMeta {
+                    target_dc: 2,
+                    client_addr: addr,
+                    our_addr: addr,
+                    proto_flags: 0,
+                },
+            )
+            .await
+    );
+    assert!(
+        registry
+            .bind_writer(
+                conn_b,
+                10,
+                ConnMeta {
+                    target_dc: -2,
+                    client_addr: addr,
+                    our_addr: addr,
+                    proto_flags: 0,
+                },
+            )
+            .await
+    );
+    assert!(
+        registry
+            .bind_writer(
+                conn_c,
+                20,
+                ConnMeta {
+                    target_dc: 4,
+                    client_addr: addr,
+                    our_addr: addr,
+                    proto_flags: 0,
+                },
+            )
+            .await
+    );
+
+    let snapshot = registry.writer_activity_snapshot().await;
+    assert_eq!(snapshot.bound_clients_by_writer.get(&10), Some(&2));
+    assert_eq!(snapshot.bound_clients_by_writer.get(&20), Some(&1));
+    assert_eq!(snapshot.active_sessions_by_target_dc.get(&2), Some(&1));
+    assert_eq!(snapshot.active_sessions_by_target_dc.get(&-2), Some(&1));
+    assert_eq!(snapshot.active_sessions_by_target_dc.get(&4), Some(&1));
+}
+
+#[tokio::test]
+async fn route_data_is_bounded_by_byte_permits_before_channel_capacity() {
+    let registry = ConnRegistry::with_route_byte_permits_for_tests(4, 1);
+    let (conn_id, mut rx) = registry.register().await;
+    let routed = registry
+        .route_nowait(
+            conn_id,
+            MeResponse::Data {
+                flags: 0,
+                data: Bytes::from_static(&[0xAA]),
+                route_permit: None,
+            },
+        )
+        .await;
+    assert!(matches!(routed, RouteResult::Routed));
+
+    let blocked = registry
+        .route_nowait(
+            conn_id,
+            MeResponse::Data {
+                flags: 0,
+                data: Bytes::from_static(&[0xBB]),
+                route_permit: None,
+            },
+        )
+        .await;
+    assert!(
+        matches!(blocked, RouteResult::QueueFullHigh),
+        "byte budget must reject data before count capacity is exhausted"
+    );
+
+    drop(rx.recv().await);
+
+    let routed_after_drain = registry
+        .route_nowait(
+            conn_id,
+            MeResponse::Data {
+                flags: 0,
+                data: Bytes::from_static(&[0xCC]),
+                route_permit: None,
+            },
+        )
+        .await;
+    assert!(
+        matches!(routed_after_drain, RouteResult::Routed),
+        "receiving queued data must release byte permits"
+    );
+}
+
+#[tokio::test]
+async fn bind_writer_rebinds_conn_atomically() {
+    let registry = ConnRegistry::new();
+    let (conn_id, _rx) = registry.register().await;
+    let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
+    let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
+    registry.register_writer(10, writer_tx_a).await;
+    registry.register_writer(20, writer_tx_b).await;
+
+    let client_addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
+    let first_our_addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::new(1, 1, 1, 1)), 443);
+    let second_our_addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::new(2, 2, 2, 2)), 443);
+
+    assert!(
+        registry
+            .bind_writer(
+                conn_id,
+                10,
+                ConnMeta {
+                    target_dc: 2,
+                    client_addr,
+                    our_addr: first_our_addr,
+                    proto_flags: 1,
+                },
+            )
+            .await
+    );
+    assert!(
+        registry
+            .bind_writer(
+                conn_id,
+                20,
+                ConnMeta {
+                    target_dc: 2,
+                    client_addr,
+                    our_addr: second_our_addr,
+                    proto_flags: 2,
+                },
+            )
+            .await
+    );
+
+    let writer = registry.get_writer(conn_id).await.expect("writer binding");
+    assert_eq!(writer.writer_id, 20);
+
+    let meta = registry.get_meta(conn_id).await.expect("conn meta");
+    assert_eq!(meta.our_addr, second_our_addr);
+    assert_eq!(meta.proto_flags, 2);
+
+    let snapshot = registry.writer_activity_snapshot().await;
+    assert_eq!(snapshot.bound_clients_by_writer.get(&10), Some(&0));
+    assert_eq!(snapshot.bound_clients_by_writer.get(&20), Some(&1));
+    assert!(
+        registry
+            .writer_idle_since_snapshot()
+            .await
+            .contains_key(&10)
+    );
+}
+
+#[tokio::test]
+async fn writer_lost_does_not_drop_rebound_conn() {
+    let registry = ConnRegistry::new();
+    let (conn_id, _rx) = registry.register().await;
+    let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
+    let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
+    registry.register_writer(10, writer_tx_a).await;
+    registry.register_writer(20, writer_tx_b).await;
+
+    let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
+    assert!(
+        registry
+            .bind_writer(
+                conn_id,
+                10,
+                ConnMeta {
+                    target_dc: 2,
+                    client_addr: addr,
+                    our_addr: addr,
+                    proto_flags: 0,
+                },
+            )
+            .await
+    );
+    assert!(
+        registry
+            .bind_writer(
+                conn_id,
+                20,
+                ConnMeta {
+                    target_dc: 2,
+                    client_addr: addr,
+                    our_addr: addr,
+                    proto_flags: 1,
+                },
+            )
+            .await
+    );
+
+    let lost = registry.writer_lost(10).await;
+    assert!(lost.is_empty());
+    assert_eq!(
+        registry
+            .get_writer(conn_id)
+            .await
+            .expect("writer")
+            .writer_id,
+        20
+    );
+
+    let removed_writer = registry.unregister(conn_id).await;
+    assert_eq!(removed_writer, Some(20));
+    assert!(registry.is_writer_empty(20).await);
+}
+
+#[tokio::test]
+async fn bind_writer_rejects_unregistered_writer() {
+    let registry = ConnRegistry::new();
+    let (conn_id, _rx) = registry.register().await;
+    let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
+
+    assert!(
+        !registry
+            .bind_writer(
+                conn_id,
+                10,
+                ConnMeta {
+                    target_dc: 2,
+                    client_addr: addr,
+                    our_addr: addr,
+                    proto_flags: 0,
+                },
+            )
+            .await
+    );
+    assert!(registry.get_writer(conn_id).await.is_none());
+}
+
+#[tokio::test]
+async fn non_empty_writer_ids_returns_only_writers_with_bound_clients() {
+    let registry = ConnRegistry::new();
+    let (conn_id, _rx) = registry.register().await;
+    let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
+    let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
+    registry.register_writer(10, writer_tx_a).await;
+    registry.register_writer(20, writer_tx_b).await;
+
+    let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
+    assert!(
+        registry
+            .bind_writer(
+                conn_id,
+                10,
+                ConnMeta {
+                    target_dc: 2,
+                    client_addr: addr,
+                    our_addr: addr,
+                    proto_flags: 0,
+                },
+            )
+            .await
+    );
+
+    let non_empty = registry.non_empty_writer_ids(&[10, 20, 30]).await;
+    assert!(non_empty.contains(&10));
+    assert!(!non_empty.contains(&20));
+    assert!(!non_empty.contains(&30));
+}

src/transport/middle_proxy/registry/writer.rs

@@ -0,0 +1,481 @@
+use std::collections::{HashMap, HashSet};
+use std::sync::atomic::Ordering;
+use std::time::Duration;
+
+use tokio::sync::mpsc;
+use tokio::sync::mpsc::error::TrySendError;
+
+use super::super::codec::WriterCommand;
+use super::super::{MeResponse, RouteBytePermit};
+use super::{
+    BoundConn, ConnMeta, ConnRegistry, ConnWriter, HotConnBinding, RouteResult,
+    WriterActivitySnapshot,
+};
+
+impl ConnRegistry {
+    pub async fn register_writer(&self, writer_id: u64, tx: mpsc::Sender<WriterCommand>) {
+        let mut binding = self.binding.inner.lock().await;
+        binding.writers.insert(writer_id, tx.clone());
+        binding
+            .conns_for_writer
+            .entry(writer_id)
+            .or_insert_with(HashSet::new);
+        self.writers.map.insert(writer_id, tx);
+    }
+
+    /// Unregister connection, returning associated writer_id if any.
+    pub async fn unregister(&self, id: u64) -> Option<u64> {
+        self.routing.map.remove(&id);
+        self.routing.byte_budget.remove(&id);
+        self.hot_binding.map.remove(&id);
+        let mut binding = self.binding.inner.lock().await;
+        binding.meta.remove(&id);
+        if let Some(writer_id) = binding.writer_for_conn.remove(&id) {
+            let became_empty = if let Some(set) = binding.conns_for_writer.get_mut(&writer_id) {
+                set.remove(&id);
+                set.is_empty()
+            } else {
+                false
+            };
+            if became_empty {
+                binding
+                    .writer_idle_since_epoch_secs
+                    .insert(writer_id, Self::now_epoch_secs());
+            }
+            return Some(writer_id);
+        }
+        None
+    }
+
+    async fn attach_route_byte_permit(
+        &self,
+        id: u64,
+        resp: MeResponse,
+        timeout_ms: Option<u64>,
+    ) -> std::result::Result<MeResponse, RouteResult> {
+        let MeResponse::Data {
+            flags,
+            data,
+            route_permit,
+        } = resp
+        else {
+            return Ok(resp);
+        };
+
+        if route_permit.is_some() {
+            return Ok(MeResponse::Data {
+                flags,
+                data,
+                route_permit,
+            });
+        }
+
+        let Some(semaphore) = self
+            .routing
+            .byte_budget
+            .get(&id)
+            .map(|entry| entry.value().clone())
+        else {
+            return Err(RouteResult::NoConn);
+        };
+        let permits = Self::route_data_permits(data.len());
+        let permit = match timeout_ms {
+            Some(0) => semaphore
+                .try_acquire_many_owned(permits)
+                .map_err(|_| RouteResult::QueueFullHigh)?,
+            Some(timeout_ms) => {
+                let acquire = semaphore.acquire_many_owned(permits);
+                match tokio::time::timeout(Duration::from_millis(timeout_ms.max(1)), acquire).await
+                {
+                    Ok(Ok(permit)) => permit,
+                    Ok(Err(_)) => return Err(RouteResult::ChannelClosed),
+                    Err(_) => return Err(RouteResult::QueueFullHigh),
+                }
+            }
+            None => semaphore
+                .acquire_many_owned(permits)
+                .await
+                .map_err(|_| RouteResult::ChannelClosed)?,
+        };
+
+        Ok(MeResponse::Data {
+            flags,
+            data,
+            route_permit: Some(RouteBytePermit::new(permit)),
+        })
+    }
+
+    #[allow(dead_code)]
+    pub async fn route(&self, id: u64, resp: MeResponse) -> RouteResult {
+        let tx = self.routing.map.get(&id).map(|entry| entry.value().clone());
+
+        let Some(tx) = tx else {
+            return RouteResult::NoConn;
+        };
+
+        let base_timeout_ms = self
+            .route_backpressure_base_timeout_ms
+            .load(Ordering::Relaxed)
+            .max(1);
+        let resp = match self
+            .attach_route_byte_permit(id, resp, Some(base_timeout_ms))
+            .await
+        {
+            Ok(resp) => resp,
+            Err(result) => return result,
+        };
+
+        match tx.try_send(resp) {
+            Ok(()) => RouteResult::Routed,
+            Err(TrySendError::Closed(_)) => RouteResult::ChannelClosed,
+            Err(TrySendError::Full(resp)) => {
+                // Absorb short bursts without dropping/closing the session immediately.
+                let high_timeout_ms = self
+                    .route_backpressure_high_timeout_ms
+                    .load(Ordering::Relaxed)
+                    .max(base_timeout_ms);
+                let high_watermark_pct = self
+                    .route_backpressure_high_watermark_pct
+                    .load(Ordering::Relaxed)
+                    .clamp(1, 100);
+                let used = self.route_channel_capacity.saturating_sub(tx.capacity());
+                let used_pct = if self.route_channel_capacity == 0 {
+                    100
+                } else {
+                    (used.saturating_mul(100) / self.route_channel_capacity) as u8
+                };
+                let high_profile = used_pct >= high_watermark_pct;
+                let timeout_ms = if high_profile {
+                    high_timeout_ms
+                } else {
+                    base_timeout_ms
+                };
+                let timeout_dur = Duration::from_millis(timeout_ms);
+
+                match tokio::time::timeout(timeout_dur, tx.send(resp)).await {
+                    Ok(Ok(())) => RouteResult::Routed,
+                    Ok(Err(_)) => RouteResult::ChannelClosed,
+                    Err(_) => {
+                        if high_profile {
+                            RouteResult::QueueFullHigh
+                        } else {
+                            RouteResult::QueueFullBase
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    pub async fn route_nowait(&self, id: u64, resp: MeResponse) -> RouteResult {
+        let tx = self.routing.map.get(&id).map(|entry| entry.value().clone());
+
+        let Some(tx) = tx else {
+            return RouteResult::NoConn;
+        };
+        let resp = match self.attach_route_byte_permit(id, resp, Some(0)).await {
+            Ok(resp) => resp,
+            Err(result) => return result,
+        };
+
+        match tx.try_send(resp) {
+            Ok(()) => RouteResult::Routed,
+            Err(TrySendError::Closed(_)) => RouteResult::ChannelClosed,
+            Err(TrySendError::Full(_)) => RouteResult::QueueFullBase,
+        }
+    }
+
+    pub async fn route_with_timeout(
+        &self,
+        id: u64,
+        resp: MeResponse,
+        timeout_ms: u64,
+    ) -> RouteResult {
+        if timeout_ms == 0 {
+            return self.route_nowait(id, resp).await;
+        }
+
+        let tx = self.routing.map.get(&id).map(|entry| entry.value().clone());
+
+        let Some(tx) = tx else {
+            return RouteResult::NoConn;
+        };
+        let resp = match self
+            .attach_route_byte_permit(id, resp, Some(timeout_ms))
+            .await
+        {
+            Ok(resp) => resp,
+            Err(result) => return result,
+        };
+
+        match tx.try_send(resp) {
+            Ok(()) => RouteResult::Routed,
+            Err(TrySendError::Closed(_)) => RouteResult::ChannelClosed,
+            Err(TrySendError::Full(resp)) => {
+                let high_watermark_pct = self
+                    .route_backpressure_high_watermark_pct
+                    .load(Ordering::Relaxed)
+                    .clamp(1, 100);
+                let used = self.route_channel_capacity.saturating_sub(tx.capacity());
+                let used_pct = if self.route_channel_capacity == 0 {
+                    100
+                } else {
+                    (used.saturating_mul(100) / self.route_channel_capacity) as u8
+                };
+                let high_profile = used_pct >= high_watermark_pct;
+                let timeout_dur = Duration::from_millis(timeout_ms.max(1));
+
+                match tokio::time::timeout(timeout_dur, tx.send(resp)).await {
+                    Ok(Ok(())) => RouteResult::Routed,
+                    Ok(Err(_)) => RouteResult::ChannelClosed,
+                    Err(_) => {
+                        if high_profile {
+                            RouteResult::QueueFullHigh
+                        } else {
+                            RouteResult::QueueFullBase
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    pub async fn bind_writer(&self, conn_id: u64, writer_id: u64, meta: ConnMeta) -> bool {
+        let mut binding = self.binding.inner.lock().await;
+        // ROUTING IS THE SOURCE OF TRUTH:
+        // never keep/attach writer binding for a connection that is already
+        // absent from the routing table.
+        if !self.routing.map.contains_key(&conn_id) {
+            return false;
+        }
+        if !binding.writers.contains_key(&writer_id) {
+            return false;
+        }
+
+        let previous_writer_id = binding.writer_for_conn.insert(conn_id, writer_id);
+        if let Some(previous_writer_id) = previous_writer_id
+            && previous_writer_id != writer_id
+        {
+            let became_empty =
+                if let Some(set) = binding.conns_for_writer.get_mut(&previous_writer_id) {
+                    set.remove(&conn_id);
+                    set.is_empty()
+                } else {
+                    false
+                };
+            if became_empty {
+                binding
+                    .writer_idle_since_epoch_secs
+                    .insert(previous_writer_id, Self::now_epoch_secs());
+            }
+        }
+
+        binding.meta.insert(conn_id, meta.clone());
+        binding.last_meta_for_writer.insert(writer_id, meta.clone());
+        binding.writer_idle_since_epoch_secs.remove(&writer_id);
+        binding
+            .conns_for_writer
+            .entry(writer_id)
+            .or_insert_with(HashSet::new)
+            .insert(conn_id);
+        self.hot_binding
+            .map
+            .insert(conn_id, HotConnBinding { writer_id, meta });
+        true
+    }
+
+    pub async fn mark_writer_idle(&self, writer_id: u64) {
+        let mut binding = self.binding.inner.lock().await;
+        binding
+            .conns_for_writer
+            .entry(writer_id)
+            .or_insert_with(HashSet::new);
+        binding
+            .writer_idle_since_epoch_secs
+            .entry(writer_id)
+            .or_insert(Self::now_epoch_secs());
+    }
+
+    pub async fn get_last_writer_meta(&self, writer_id: u64) -> Option<ConnMeta> {
+        let binding = self.binding.inner.lock().await;
+        binding.last_meta_for_writer.get(&writer_id).cloned()
+    }
+
+    pub async fn writer_idle_since_snapshot(&self) -> HashMap<u64, u64> {
+        let binding = self.binding.inner.lock().await;
+        binding.writer_idle_since_epoch_secs.clone()
+    }
+
+    pub async fn writer_idle_since_for_writer_ids(&self, writer_ids: &[u64]) -> HashMap<u64, u64> {
+        let binding = self.binding.inner.lock().await;
+        let mut out = HashMap::<u64, u64>::with_capacity(writer_ids.len());
+        for writer_id in writer_ids {
+            if let Some(idle_since) = binding.writer_idle_since_epoch_secs.get(writer_id).copied() {
+                out.insert(*writer_id, idle_since);
+            }
+        }
+        out
+    }
+
+    pub(in crate::transport::middle_proxy) async fn writer_activity_snapshot(
+        &self,
+    ) -> WriterActivitySnapshot {
+        let binding = self.binding.inner.lock().await;
+        let mut bound_clients_by_writer = HashMap::<u64, usize>::new();
+        let mut active_sessions_by_target_dc = HashMap::<i16, usize>::new();
+
+        for (writer_id, conn_ids) in &binding.conns_for_writer {
+            bound_clients_by_writer.insert(*writer_id, conn_ids.len());
+        }
+        for conn_meta in binding.meta.values() {
+            if conn_meta.target_dc == 0 {
+                continue;
+            }
+            *active_sessions_by_target_dc
+                .entry(conn_meta.target_dc)
+                .or_insert(0) += 1;
+        }
+
+        WriterActivitySnapshot {
+            bound_clients_by_writer,
+            active_sessions_by_target_dc,
+        }
+    }
+
+    pub async fn get_writer(&self, conn_id: u64) -> Option<ConnWriter> {
+        if !self.routing.map.contains_key(&conn_id) {
+            return None;
+        }
+
+        let writer_id = self
+            .hot_binding
+            .map
+            .get(&conn_id)
+            .map(|entry| entry.writer_id)?;
+        let writer = self
+            .writers
+            .map
+            .get(&writer_id)
+            .map(|entry| entry.value().clone())?;
+        Some(ConnWriter {
+            writer_id,
+            tx: writer,
+        })
+    }
+
+    /// Returns the active writer and routing metadata from one hot-binding lookup.
+    pub async fn get_writer_with_meta(&self, conn_id: u64) -> Option<(ConnWriter, ConnMeta)> {
+        if !self.routing.map.contains_key(&conn_id) {
+            return None;
+        }
+
+        let hot = self.hot_binding.map.get(&conn_id)?;
+        let writer_id = hot.writer_id;
+        let meta = hot.meta.clone();
+        let writer = self
+            .writers
+            .map
+            .get(&writer_id)
+            .map(|entry| entry.value().clone())?;
+        Some((
+            ConnWriter {
+                writer_id,
+                tx: writer,
+            },
+            meta,
+        ))
+    }
+
+    pub async fn active_conn_ids(&self) -> Vec<u64> {
+        let binding = self.binding.inner.lock().await;
+        binding.writer_for_conn.keys().copied().collect()
+    }
+
+    pub async fn writer_lost(&self, writer_id: u64) -> Vec<BoundConn> {
+        let mut binding = self.binding.inner.lock().await;
+        binding.writers.remove(&writer_id);
+        self.writers.map.remove(&writer_id);
+        binding.last_meta_for_writer.remove(&writer_id);
+        binding.writer_idle_since_epoch_secs.remove(&writer_id);
+        let conns = binding
+            .conns_for_writer
+            .remove(&writer_id)
+            .unwrap_or_default()
+            .into_iter()
+            .collect::<Vec<_>>();
+
+        let mut out = Vec::new();
+        for conn_id in conns {
+            if binding.writer_for_conn.get(&conn_id).copied() != Some(writer_id) {
+                continue;
+            }
+            binding.writer_for_conn.remove(&conn_id);
+            let remove_hot = self
+                .hot_binding
+                .map
+                .get(&conn_id)
+                .map(|hot| hot.writer_id == writer_id)
+                .unwrap_or(false);
+            if remove_hot {
+                self.hot_binding.map.remove(&conn_id);
+            }
+            if let Some(m) = binding.meta.get(&conn_id) {
+                out.push(BoundConn {
+                    conn_id,
+                    meta: m.clone(),
+                });
+            }
+        }
+        out
+    }
+
+    #[allow(dead_code)]
+    pub async fn get_meta(&self, conn_id: u64) -> Option<ConnMeta> {
+        self.hot_binding
+            .map
+            .get(&conn_id)
+            .map(|entry| entry.meta.clone())
+    }
+
+    pub async fn is_writer_empty(&self, writer_id: u64) -> bool {
+        let binding = self.binding.inner.lock().await;
+        binding
+            .conns_for_writer
+            .get(&writer_id)
+            .map(|s| s.is_empty())
+            .unwrap_or(true)
+    }
+
+    #[allow(dead_code)]
+    pub async fn unregister_writer_if_empty(&self, writer_id: u64) -> bool {
+        let mut binding = self.binding.inner.lock().await;
+        let Some(conn_ids) = binding.conns_for_writer.get(&writer_id) else {
+            // Writer is already absent from the registry.
+            return true;
+        };
+        if !conn_ids.is_empty() {
+            return false;
+        }
+
+        binding.writers.remove(&writer_id);
+        self.writers.map.remove(&writer_id);
+        binding.last_meta_for_writer.remove(&writer_id);
+        binding.writer_idle_since_epoch_secs.remove(&writer_id);
+        binding.conns_for_writer.remove(&writer_id);
+        true
+    }
+
+    #[allow(dead_code)]
+    pub(super) async fn non_empty_writer_ids(&self, writer_ids: &[u64]) -> HashSet<u64> {
+        let binding = self.binding.inner.lock().await;
+        let mut out = HashSet::<u64>::with_capacity(writer_ids.len());
+        for writer_id in writer_ids {
+            if let Some(conns) = binding.conns_for_writer.get(writer_id)
+                && !conns.is_empty()
+            {
+                out.insert(*writer_id);
+            }
+        }
+        out
+    }
+}

src/transport/middle_proxy/send.rs

@@ -1,26 +1,22 @@
 #![allow(clippy::too_many_arguments)]
 
 use std::cmp::Reverse;
-use std::collections::{HashMap, HashSet};
 use std::net::SocketAddr;
 use std::sync::Arc;
 use std::sync::atomic::Ordering;
 use std::time::{Duration, Instant};
 
-use bytes::Bytes;
 use tokio::sync::mpsc::error::TrySendError;
 use tracing::{debug, warn};
 
+use super::MePool;
+use super::codec::{ProxyReqCommand, WriterCommand};
+use super::registry::ConnMeta;
+use super::wire::build_proxy_req_payload;
 use crate::config::{MeRouteNoWriterMode, MeWriterPickMode};
 use crate::error::{ProxyError, Result};
 use crate::network::IpFamily;
-use crate::protocol::constants::{RPC_CLOSE_CONN_U32, RPC_CLOSE_EXT_U32};
-
-use super::MePool;
-use super::codec::WriterCommand;
-use super::pool::WriterContour;
-use super::registry::ConnMeta;
-use super::wire::build_proxy_req_payload;
+use crate::stream::PooledBuffer;
 use rand::seq::SliceRandom;
 
 const IDLE_WRITER_PENALTY_MID_SECS: u64 = 45;
@@ -34,6 +30,21 @@ const PICK_PENALTY_DRAINING: u64 = 600;
 const PICK_PENALTY_STALE: u64 = 300;
 const PICK_PENALTY_DEGRADED: u64 = 250;
 
+mod close;
+mod recovery;
+mod selection;
+
+fn proxy_tag_array(tag: Option<&[u8]>) -> Option<[u8; 16]> {
+    tag.and_then(|tag| <[u8; 16]>::try_from(tag).ok())
+}
+
+fn proxy_req_payload_from_command(cmd: WriterCommand) -> Option<PooledBuffer> {
+    match cmd {
+        WriterCommand::ProxyReq(command) => Some(command.payload),
+        _ => None,
+    }
+}
+
 impl MePool {
     /// Send RPC_PROXY_REQ. `tag_override`: per-user ad_tag (from access.user_ad_tags); if None, uses pool default.
     pub async fn send_proxy_req(
@@ -47,12 +58,6 @@ impl MePool {
         tag_override: Option<&[u8]>,
     ) -> Result<()> {
         let tag = tag_override.or(self.proxy_tag.as_deref());
-        let fallback_meta = ConnMeta {
-            target_dc,
-            client_addr,
-            our_addr,
-            proto_flags,
-        };
         let build_routed_payload = |effective_our_addr: SocketAddr| {
             (
                 build_proxy_req_payload(
@@ -91,17 +96,10 @@ impl MePool {
         let mut hybrid_wait_current = hybrid_wait_step;
 
         loop {
-            let current_meta = self
-                .registry
-                .get_meta(conn_id)
-                .await
-                .unwrap_or_else(|| fallback_meta.clone());
-            let (current_payload, _) = build_routed_payload(current_meta.our_addr);
-            if let Some(current) = self.registry.get_writer(conn_id).await {
-                match current
-                    .tx
-                    .try_send(WriterCommand::Data(current_payload.clone()))
-                {
+            if let Some((current, current_meta)) = self.registry.get_writer_with_meta(conn_id).await
+            {
+                let (current_payload, _) = build_routed_payload(current_meta.our_addr);
+                match current.tx.try_send(WriterCommand::Data(current_payload)) {
                     Ok(()) => {
                         self.note_hybrid_route_success();
                         return Ok(());
@@ -452,7 +450,7 @@ impl MePool {
                             self.remove_writer_and_close_clients(w.id).await;
                             continue;
                         }
-                        permit.send(WriterCommand::Data(payload.clone()));
+                        permit.send(WriterCommand::Data(payload));
                         self.stats
                             .increment_me_writer_pick_success_try_total(pick_mode);
                         if w.generation < self.current_generation() {
@@ -520,7 +518,7 @@ impl MePool {
                         self.remove_writer_and_close_clients(w.id).await;
                         continue;
                     }
-                    permit.send(WriterCommand::Data(payload.clone()));
+                    permit.send(WriterCommand::Data(payload));
                     self.stats
                         .increment_me_writer_pick_success_fallback_total(pick_mode);
                     if w.generation < self.current_generation() {
@@ -538,405 +536,93 @@ impl MePool {
         }
     }
 
-    async fn wait_for_writer_until(&self, deadline: Instant) -> bool {
-        let mut rx = self.writer_epoch.subscribe();
-        if !self.writers.read().await.is_empty() {
-            return true;
-        }
-        let now = Instant::now();
-        if now >= deadline {
-            return !self.writers.read().await.is_empty();
-        }
-        let timeout = deadline.saturating_duration_since(now);
-        if tokio::time::timeout(timeout, rx.changed()).await.is_ok() {
-            return !self.writers.read().await.is_empty();
-        }
-        !self.writers.read().await.is_empty()
-    }
-
-    async fn wait_for_candidate_until(&self, routed_dc: i32, deadline: Instant) -> bool {
-        let mut rx = self.writer_epoch.subscribe();
-        loop {
-            if self.has_candidate_for_target_dc(routed_dc).await {
-                return true;
-            }
-
-            let now = Instant::now();
-            if now >= deadline {
-                return self.has_candidate_for_target_dc(routed_dc).await;
-            }
-
-            if self.has_candidate_for_target_dc(routed_dc).await {
-                return true;
-            }
-            let remaining = deadline.saturating_duration_since(Instant::now());
-            if remaining.is_zero() {
-                return self.has_candidate_for_target_dc(routed_dc).await;
-            }
-            if tokio::time::timeout(remaining, rx.changed()).await.is_err() {
-                return self.has_candidate_for_target_dc(routed_dc).await;
-            }
-        }
-    }
-
-    async fn has_candidate_for_target_dc(&self, routed_dc: i32) -> bool {
-        let writers_snapshot = {
-            let ws = self.writers.read().await;
-            if ws.is_empty() {
-                return false;
-            }
-            ws.clone()
-        };
-        let mut candidate_indices = self
-            .candidate_indices_for_dc(&writers_snapshot, routed_dc, false)
-            .await;
-        if candidate_indices.is_empty() {
-            candidate_indices = self
-                .candidate_indices_for_dc(&writers_snapshot, routed_dc, true)
-                .await;
-        }
-        !candidate_indices.is_empty()
-    }
-
-    async fn trigger_async_recovery_for_target_dc(self: &Arc<Self>, routed_dc: i32) -> bool {
-        let endpoints = self.endpoint_candidates_for_target_dc(routed_dc).await;
-        if endpoints.is_empty() {
-            return false;
-        }
-        self.stats.increment_me_async_recovery_trigger_total();
-        for addr in endpoints.into_iter().take(8) {
-            self.trigger_immediate_refill_for_dc(addr, routed_dc);
-        }
-        true
-    }
-
-    async fn trigger_async_recovery_global(self: &Arc<Self>) {
-        self.stats.increment_me_async_recovery_trigger_total();
-        let mut seen = HashSet::<(i32, SocketAddr)>::new();
-        for family in self.family_order() {
-            let map_guard = match family {
-                IpFamily::V4 => self.proxy_map_v4.read().await,
-                IpFamily::V6 => self.proxy_map_v6.read().await,
-            };
-            for (dc, addrs) in map_guard.iter() {
-                for (ip, port) in addrs {
-                    let addr = SocketAddr::new(*ip, *port);
-                    if seen.insert((*dc, addr)) {
-                        self.trigger_immediate_refill_for_dc(addr, *dc);
-                    }
-                    if seen.len() >= 8 {
-                        return;
-                    }
-                }
-            }
-        }
-    }
-
-    async fn endpoint_candidates_for_target_dc(&self, routed_dc: i32) -> Vec<SocketAddr> {
-        self.preferred_endpoints_for_dc(routed_dc).await
-    }
-
-    async fn maybe_trigger_hybrid_recovery(
+    /// Send RPC_PROXY_REQ while keeping the first bound-writer path allocation-light.
+    pub async fn send_proxy_req_pooled(
         self: &Arc<Self>,
-        routed_dc: i32,
-        hybrid_recovery_round: &mut u32,
-        hybrid_last_recovery_at: &mut Option<Instant>,
-        hybrid_wait_step: Duration,
-    ) {
-        if !self.try_consume_hybrid_recovery_trigger_slot(HYBRID_RECOVERY_TRIGGER_MIN_INTERVAL_MS) {
-            return;
-        }
-        if let Some(last) = *hybrid_last_recovery_at
-            && last.elapsed() < hybrid_wait_step
-        {
-            return;
-        }
+        conn_id: u64,
+        target_dc: i16,
+        client_addr: SocketAddr,
+        our_addr: SocketAddr,
+        payload: PooledBuffer,
+        proto_flags: u32,
+        tag_override: Option<[u8; 16]>,
+    ) -> Result<()> {
+        let tag = tag_override.or_else(|| proxy_tag_array(self.proxy_tag.as_deref()));
 
-        let round = *hybrid_recovery_round;
-        let target_triggered = self.trigger_async_recovery_for_target_dc(routed_dc).await;
-        if !target_triggered || round.is_multiple_of(HYBRID_GLOBAL_BURST_PERIOD_ROUNDS) {
-            self.trigger_async_recovery_global().await;
-        }
-        *hybrid_recovery_round = round.saturating_add(1);
-        *hybrid_last_recovery_at = Some(Instant::now());
-    }
-
-    fn hybrid_total_wait_budget(&self) -> Duration {
-        let base = self
-            .route_runtime
-            .me_route_hybrid_max_wait
-            .max(Duration::from_millis(50));
-        let now_ms = Self::now_epoch_millis();
-        let last_success_ms = self
-            .route_runtime
-            .me_route_last_success_epoch_ms
-            .load(Ordering::Relaxed);
-        if last_success_ms != 0
-            && now_ms.saturating_sub(last_success_ms) <= HYBRID_RECENT_SUCCESS_WINDOW_MS
-        {
-            return base.saturating_mul(2);
-        }
-        base
-    }
-
-    fn note_hybrid_route_success(&self) {
-        self.route_runtime
-            .me_route_last_success_epoch_ms
-            .store(Self::now_epoch_millis(), Ordering::Relaxed);
-    }
-
-    fn on_hybrid_timeout(&self, deadline: Instant, routed_dc: i32) {
-        self.stats.increment_me_hybrid_timeout_total();
-        let now_ms = Self::now_epoch_millis();
-        let mut last_warn_ms = self
-            .route_runtime
-            .me_route_hybrid_timeout_warn_epoch_ms
-            .load(Ordering::Relaxed);
-        while now_ms.saturating_sub(last_warn_ms) >= HYBRID_TIMEOUT_WARN_RATE_LIMIT_MS {
-            match self
-                .route_runtime
-                .me_route_hybrid_timeout_warn_epoch_ms
-                .compare_exchange_weak(last_warn_ms, now_ms, Ordering::AcqRel, Ordering::Relaxed)
-            {
-                Ok(_) => {
-                    warn!(
-                        routed_dc,
-                        budget_ms = self.hybrid_total_wait_budget().as_millis() as u64,
-                        elapsed_ms = deadline.elapsed().as_millis() as u64,
-                        "ME hybrid route timeout reached"
-                    );
-                    break;
+        if let Some((current, current_meta)) = self.registry.get_writer_with_meta(conn_id).await {
+            let command = WriterCommand::ProxyReq(ProxyReqCommand {
+                conn_id,
+                client_addr,
+                our_addr: current_meta.our_addr,
+                proto_flags,
+                proxy_tag: tag,
+                payload,
+            });
+            match current.tx.try_send(command) {
+                Ok(()) => {
+                    self.note_hybrid_route_success();
+                    return Ok(());
                 }
-                Err(actual) => last_warn_ms = actual,
-            }
-        }
-    }
-
-    fn try_consume_hybrid_recovery_trigger_slot(&self, min_interval_ms: u64) -> bool {
-        let now_ms = Self::now_epoch_millis();
-        let mut last_trigger_ms = self
-            .route_runtime
-            .me_async_recovery_last_trigger_epoch_ms
-            .load(Ordering::Relaxed);
-        loop {
-            if now_ms.saturating_sub(last_trigger_ms) < min_interval_ms {
-                return false;
-            }
-            match self
-                .route_runtime
-                .me_async_recovery_last_trigger_epoch_ms
-                .compare_exchange_weak(last_trigger_ms, now_ms, Ordering::AcqRel, Ordering::Relaxed)
-            {
-                Ok(_) => return true,
-                Err(actual) => last_trigger_ms = actual,
-            }
-        }
-    }
-
-    pub async fn send_close(self: &Arc<Self>, conn_id: u64) -> Result<()> {
-        if let Some(w) = self.registry.get_writer(conn_id).await {
-            let mut p = Vec::with_capacity(12);
-            p.extend_from_slice(&RPC_CLOSE_EXT_U32.to_le_bytes());
-            p.extend_from_slice(&conn_id.to_le_bytes());
-            if w.tx
-                .send(WriterCommand::DataAndFlush(Bytes::from(p)))
-                .await
-                .is_err()
-            {
-                debug!("ME close write failed");
-                self.remove_writer_and_close_clients(w.writer_id).await;
-            }
-        } else {
-            debug!(conn_id, "ME close skipped (writer missing)");
-        }
-
-        self.registry.unregister(conn_id).await;
-        Ok(())
-    }
-
-    pub async fn send_close_conn(self: &Arc<Self>, conn_id: u64) -> Result<()> {
-        if let Some(w) = self.registry.get_writer(conn_id).await {
-            let mut p = Vec::with_capacity(12);
-            p.extend_from_slice(&RPC_CLOSE_CONN_U32.to_le_bytes());
-            p.extend_from_slice(&conn_id.to_le_bytes());
-            match w.tx.try_send(WriterCommand::DataAndFlush(Bytes::from(p))) {
-                Ok(()) => {}
-                Err(TrySendError::Full(cmd)) => {
-                    let _ = tokio::time::timeout(Duration::from_millis(50), w.tx.send(cmd)).await;
-                }
-                Err(TrySendError::Closed(_)) => {
-                    debug!(conn_id, "ME close_conn skipped: writer channel closed");
+                Err(TrySendError::Full(cmd)) => match current.tx.send(cmd).await {
+                    Ok(()) => {
+                        self.note_hybrid_route_success();
+                        return Ok(());
+                    }
+                    Err(send_err) => {
+                        let Some(payload) = proxy_req_payload_from_command(send_err.0) else {
+                            return Err(ProxyError::Proxy(
+                                "ME writer rejected unexpected command type".into(),
+                            ));
+                        };
+                        warn!(writer_id = current.writer_id, "ME writer channel closed");
+                        self.remove_writer_and_close_clients(current.writer_id)
+                            .await;
+                        return self
+                            .send_proxy_req(
+                                conn_id,
+                                target_dc,
+                                client_addr,
+                                our_addr,
+                                payload.as_ref(),
+                                proto_flags,
+                                tag.as_ref().map(|tag| tag.as_slice()),
+                            )
+                            .await;
+                    }
+                },
+                Err(TrySendError::Closed(cmd)) => {
+                    let Some(payload) = proxy_req_payload_from_command(cmd) else {
+                        return Err(ProxyError::Proxy(
+                            "ME writer rejected unexpected command type".into(),
+                        ));
+                    };
+                    warn!(writer_id = current.writer_id, "ME writer channel closed");
+                    self.remove_writer_and_close_clients(current.writer_id)
+                        .await;
+                    return self
+                        .send_proxy_req(
+                            conn_id,
+                            target_dc,
+                            client_addr,
+                            our_addr,
+                            payload.as_ref(),
+                            proto_flags,
+                            tag.as_ref().map(|tag| tag.as_slice()),
+                        )
+                        .await;
                 }
             }
-        } else {
-            debug!(conn_id, "ME close_conn skipped (writer missing)");
         }
 
-        self.registry.unregister(conn_id).await;
-        Ok(())
-    }
-
-    pub async fn shutdown_send_close_conn_all(self: &Arc<Self>) -> usize {
-        let conn_ids = self.registry.active_conn_ids().await;
-        let total = conn_ids.len();
-        for conn_id in conn_ids {
-            let _ = self.send_close_conn(conn_id).await;
-        }
-        total
-    }
-
-    pub fn connection_count(&self) -> usize {
-        self.conn_count.load(Ordering::Relaxed)
-    }
-
-    pub(super) async fn candidate_indices_for_dc(
-        &self,
-        writers: &[super::pool::MeWriter],
-        routed_dc: i32,
-        include_warm: bool,
-    ) -> Vec<usize> {
-        let preferred = self.preferred_endpoints_for_dc(routed_dc).await;
-        if preferred.is_empty() {
-            return Vec::new();
-        }
-
-        let mut out = Vec::new();
-        for (idx, w) in writers.iter().enumerate() {
-            if !self.writer_eligible_for_selection(w, include_warm) {
-                continue;
-            }
-            if w.writer_dc == routed_dc && preferred.contains(&w.addr) {
-                out.push(idx);
-            }
-        }
-        out
-    }
-
-    fn writer_eligible_for_selection(
-        &self,
-        writer: &super::pool::MeWriter,
-        include_warm: bool,
-    ) -> bool {
-        if !self.writer_accepts_new_binding(writer) {
-            return false;
-        }
-
-        match WriterContour::from_u8(writer.contour.load(Ordering::Relaxed)) {
-            WriterContour::Active => true,
-            WriterContour::Warm => include_warm,
-            WriterContour::Draining => true,
-        }
-    }
-
-    fn writer_contour_rank_for_selection(&self, writer: &super::pool::MeWriter) -> usize {
-        match WriterContour::from_u8(writer.contour.load(Ordering::Relaxed)) {
-            WriterContour::Active => 0,
-            WriterContour::Warm => 1,
-            WriterContour::Draining => 2,
-        }
-    }
-
-    fn writer_idle_rank_for_selection(
-        &self,
-        writer: &super::pool::MeWriter,
-        idle_since_by_writer: &HashMap<u64, u64>,
-        now_epoch_secs: u64,
-    ) -> usize {
-        let Some(idle_since) = idle_since_by_writer.get(&writer.id).copied() else {
-            return 0;
-        };
-        let idle_age_secs = now_epoch_secs.saturating_sub(idle_since);
-        if idle_age_secs >= IDLE_WRITER_PENALTY_HIGH_SECS {
-            2
-        } else if idle_age_secs >= IDLE_WRITER_PENALTY_MID_SECS {
-            1
-        } else {
-            0
-        }
-    }
-
-    fn writer_pick_score(
-        &self,
-        writer: &super::pool::MeWriter,
-        idle_since_by_writer: &HashMap<u64, u64>,
-        now_epoch_secs: u64,
-    ) -> u64 {
-        let contour_penalty = match WriterContour::from_u8(writer.contour.load(Ordering::Relaxed)) {
-            WriterContour::Active => 0,
-            WriterContour::Warm => PICK_PENALTY_WARM,
-            WriterContour::Draining => PICK_PENALTY_DRAINING,
-        };
-        let stale_penalty = if writer.generation < self.current_generation() {
-            PICK_PENALTY_STALE
-        } else {
-            0
-        };
-        let degraded_penalty = if writer.degraded.load(Ordering::Relaxed) {
-            PICK_PENALTY_DEGRADED
-        } else {
-            0
-        };
-        let idle_penalty =
-            (self.writer_idle_rank_for_selection(writer, idle_since_by_writer, now_epoch_secs)
-                as u64)
-                * 100;
-        let queue_cap = self.writer_lifecycle.writer_cmd_channel_capacity.max(1) as u64;
-        let queue_remaining = writer.tx.capacity() as u64;
-        let queue_used = queue_cap.saturating_sub(queue_remaining.min(queue_cap));
-        let queue_util_pct = queue_used.saturating_mul(100) / queue_cap;
-        let queue_penalty = queue_util_pct.saturating_mul(4);
-        let rtt_penalty =
-            ((writer.rtt_ema_ms_x10.load(Ordering::Relaxed) as u64).saturating_add(5) / 10)
-                .min(400);
-
-        contour_penalty
-            .saturating_add(stale_penalty)
-            .saturating_add(degraded_penalty)
-            .saturating_add(idle_penalty)
-            .saturating_add(queue_penalty)
-            .saturating_add(rtt_penalty)
-    }
-
-    fn p2c_ordered_candidate_indices(
-        &self,
-        candidate_indices: &[usize],
-        writers_snapshot: &[super::pool::MeWriter],
-        idle_since_by_writer: &HashMap<u64, u64>,
-        now_epoch_secs: u64,
-        start: usize,
-        sample_size: usize,
-    ) -> Vec<usize> {
-        let total = candidate_indices.len();
-        if total == 0 {
-            return Vec::new();
-        }
-
-        let mut sampled = Vec::<usize>::with_capacity(sample_size.min(total));
-        let mut seen = HashSet::<usize>::with_capacity(total);
-        for offset in 0..sample_size.min(total) {
-            let idx = candidate_indices[(start + offset) % total];
-            if seen.insert(idx) {
-                sampled.push(idx);
-            }
-        }
-
-        sampled.sort_by_key(|idx| {
-            let writer = &writers_snapshot[*idx];
-            (
-                self.writer_pick_score(writer, idle_since_by_writer, now_epoch_secs),
-                writer.addr,
-                writer.id,
-            )
-        });
-
-        let mut ordered = Vec::<usize>::with_capacity(total);
-        ordered.extend(sampled.iter().copied());
-        for offset in 0..total {
-            let idx = candidate_indices[(start + offset) % total];
-            if seen.insert(idx) {
-                ordered.push(idx);
-            }
-        }
-        ordered
+        self.send_proxy_req(
+            conn_id,
+            target_dc,
+            client_addr,
+            our_addr,
+            payload.as_ref(),
+            proto_flags,
+            tag.as_ref().map(|tag| tag.as_slice()),
+        )
+        .await
     }
 }

src/transport/middle_proxy/send/close.rs

@@ -0,0 +1,66 @@
+use std::sync::Arc;
+use std::sync::atomic::Ordering;
+use std::time::Duration;
+
+use tokio::sync::mpsc::error::TrySendError;
+use tracing::debug;
+
+use crate::error::Result;
+use crate::protocol::constants::{RPC_CLOSE_CONN_U32, RPC_CLOSE_EXT_U32};
+
+use super::super::MePool;
+use super::super::codec::{WriterCommand, build_control_payload};
+
+impl MePool {
+    pub async fn send_close(self: &Arc<Self>, conn_id: u64) -> Result<()> {
+        if let Some(w) = self.registry.get_writer(conn_id).await {
+            let payload = build_control_payload(RPC_CLOSE_EXT_U32, conn_id);
+            if w.tx
+                .send(WriterCommand::ControlAndFlush(payload))
+                .await
+                .is_err()
+            {
+                debug!("ME close write failed");
+                self.remove_writer_and_close_clients(w.writer_id).await;
+            }
+        } else {
+            debug!(conn_id, "ME close skipped (writer missing)");
+        }
+
+        self.registry.unregister(conn_id).await;
+        Ok(())
+    }
+
+    pub async fn send_close_conn(self: &Arc<Self>, conn_id: u64) -> Result<()> {
+        if let Some(w) = self.registry.get_writer(conn_id).await {
+            let payload = build_control_payload(RPC_CLOSE_CONN_U32, conn_id);
+            match w.tx.try_send(WriterCommand::ControlAndFlush(payload)) {
+                Ok(()) => {}
+                Err(TrySendError::Full(cmd)) => {
+                    let _ = tokio::time::timeout(Duration::from_millis(50), w.tx.send(cmd)).await;
+                }
+                Err(TrySendError::Closed(_)) => {
+                    debug!(conn_id, "ME close_conn skipped: writer channel closed");
+                }
+            }
+        } else {
+            debug!(conn_id, "ME close_conn skipped (writer missing)");
+        }
+
+        self.registry.unregister(conn_id).await;
+        Ok(())
+    }
+
+    pub async fn shutdown_send_close_conn_all(self: &Arc<Self>) -> usize {
+        let conn_ids = self.registry.active_conn_ids().await;
+        let total = conn_ids.len();
+        for conn_id in conn_ids {
+            let _ = self.send_close_conn(conn_id).await;
+        }
+        total
+    }
+
+    pub fn connection_count(&self) -> usize {
+        self.conn_count.load(Ordering::Relaxed)
+    }
+}

src/transport/middle_proxy/send/recovery.rs

@@ -0,0 +1,218 @@
+use std::collections::HashSet;
+use std::net::SocketAddr;
+use std::sync::Arc;
+use std::sync::atomic::Ordering;
+use std::time::{Duration, Instant};
+
+use tracing::warn;
+
+use crate::network::IpFamily;
+
+use super::super::MePool;
+use super::{
+    HYBRID_GLOBAL_BURST_PERIOD_ROUNDS, HYBRID_RECENT_SUCCESS_WINDOW_MS,
+    HYBRID_RECOVERY_TRIGGER_MIN_INTERVAL_MS, HYBRID_TIMEOUT_WARN_RATE_LIMIT_MS,
+};
+
+impl MePool {
+    pub(super) async fn wait_for_writer_until(&self, deadline: Instant) -> bool {
+        let mut rx = self.writer_epoch.subscribe();
+        if !self.writers.read().await.is_empty() {
+            return true;
+        }
+        let now = Instant::now();
+        if now >= deadline {
+            return !self.writers.read().await.is_empty();
+        }
+        let timeout = deadline.saturating_duration_since(now);
+        if tokio::time::timeout(timeout, rx.changed()).await.is_ok() {
+            return !self.writers.read().await.is_empty();
+        }
+        !self.writers.read().await.is_empty()
+    }
+
+    pub(super) async fn wait_for_candidate_until(&self, routed_dc: i32, deadline: Instant) -> bool {
+        let mut rx = self.writer_epoch.subscribe();
+        loop {
+            if self.has_candidate_for_target_dc(routed_dc).await {
+                return true;
+            }
+
+            let now = Instant::now();
+            if now >= deadline {
+                return self.has_candidate_for_target_dc(routed_dc).await;
+            }
+
+            if self.has_candidate_for_target_dc(routed_dc).await {
+                return true;
+            }
+            let remaining = deadline.saturating_duration_since(Instant::now());
+            if remaining.is_zero() {
+                return self.has_candidate_for_target_dc(routed_dc).await;
+            }
+            if tokio::time::timeout(remaining, rx.changed()).await.is_err() {
+                return self.has_candidate_for_target_dc(routed_dc).await;
+            }
+        }
+    }
+
+    pub(super) async fn has_candidate_for_target_dc(&self, routed_dc: i32) -> bool {
+        let writers_snapshot = {
+            let ws = self.writers.read().await;
+            if ws.is_empty() {
+                return false;
+            }
+            ws.clone()
+        };
+        let mut candidate_indices = self
+            .candidate_indices_for_dc(&writers_snapshot, routed_dc, false)
+            .await;
+        if candidate_indices.is_empty() {
+            candidate_indices = self
+                .candidate_indices_for_dc(&writers_snapshot, routed_dc, true)
+                .await;
+        }
+        !candidate_indices.is_empty()
+    }
+
+    pub(super) async fn trigger_async_recovery_for_target_dc(
+        self: &Arc<Self>,
+        routed_dc: i32,
+    ) -> bool {
+        let endpoints = self.endpoint_candidates_for_target_dc(routed_dc).await;
+        if endpoints.is_empty() {
+            return false;
+        }
+        self.stats.increment_me_async_recovery_trigger_total();
+        for addr in endpoints.into_iter().take(8) {
+            self.trigger_immediate_refill_for_dc(addr, routed_dc);
+        }
+        true
+    }
+
+    pub(super) async fn trigger_async_recovery_global(self: &Arc<Self>) {
+        self.stats.increment_me_async_recovery_trigger_total();
+        let mut seen = HashSet::<(i32, SocketAddr)>::new();
+        for family in self.family_order() {
+            let map_guard = match family {
+                IpFamily::V4 => self.proxy_map_v4.read().await,
+                IpFamily::V6 => self.proxy_map_v6.read().await,
+            };
+            for (dc, addrs) in map_guard.iter() {
+                for (ip, port) in addrs {
+                    let addr = SocketAddr::new(*ip, *port);
+                    if seen.insert((*dc, addr)) {
+                        self.trigger_immediate_refill_for_dc(addr, *dc);
+                    }
+                    if seen.len() >= 8 {
+                        return;
+                    }
+                }
+            }
+        }
+    }
+
+    pub(super) async fn endpoint_candidates_for_target_dc(
+        &self,
+        routed_dc: i32,
+    ) -> Vec<SocketAddr> {
+        self.preferred_endpoints_for_dc(routed_dc).await
+    }
+
+    pub(super) async fn maybe_trigger_hybrid_recovery(
+        self: &Arc<Self>,
+        routed_dc: i32,
+        hybrid_recovery_round: &mut u32,
+        hybrid_last_recovery_at: &mut Option<Instant>,
+        hybrid_wait_step: Duration,
+    ) {
+        if !self.try_consume_hybrid_recovery_trigger_slot(HYBRID_RECOVERY_TRIGGER_MIN_INTERVAL_MS) {
+            return;
+        }
+        if let Some(last) = *hybrid_last_recovery_at
+            && last.elapsed() < hybrid_wait_step
+        {
+            return;
+        }
+
+        let round = *hybrid_recovery_round;
+        let target_triggered = self.trigger_async_recovery_for_target_dc(routed_dc).await;
+        if !target_triggered || round.is_multiple_of(HYBRID_GLOBAL_BURST_PERIOD_ROUNDS) {
+            self.trigger_async_recovery_global().await;
+        }
+        *hybrid_recovery_round = round.saturating_add(1);
+        *hybrid_last_recovery_at = Some(Instant::now());
+    }
+
+    pub(super) fn hybrid_total_wait_budget(&self) -> Duration {
+        let base = self
+            .route_runtime
+            .me_route_hybrid_max_wait
+            .max(Duration::from_millis(50));
+        let now_ms = Self::now_epoch_millis();
+        let last_success_ms = self
+            .route_runtime
+            .me_route_last_success_epoch_ms
+            .load(Ordering::Relaxed);
+        if last_success_ms != 0
+            && now_ms.saturating_sub(last_success_ms) <= HYBRID_RECENT_SUCCESS_WINDOW_MS
+        {
+            return base.saturating_mul(2);
+        }
+        base
+    }
+
+    pub(super) fn note_hybrid_route_success(&self) {
+        self.route_runtime
+            .me_route_last_success_epoch_ms
+            .store(Self::now_epoch_millis(), Ordering::Relaxed);
+    }
+
+    pub(super) fn on_hybrid_timeout(&self, deadline: Instant, routed_dc: i32) {
+        self.stats.increment_me_hybrid_timeout_total();
+        let now_ms = Self::now_epoch_millis();
+        let mut last_warn_ms = self
+            .route_runtime
+            .me_route_hybrid_timeout_warn_epoch_ms
+            .load(Ordering::Relaxed);
+        while now_ms.saturating_sub(last_warn_ms) >= HYBRID_TIMEOUT_WARN_RATE_LIMIT_MS {
+            match self
+                .route_runtime
+                .me_route_hybrid_timeout_warn_epoch_ms
+                .compare_exchange_weak(last_warn_ms, now_ms, Ordering::AcqRel, Ordering::Relaxed)
+            {
+                Ok(_) => {
+                    warn!(
+                        routed_dc,
+                        budget_ms = self.hybrid_total_wait_budget().as_millis() as u64,
+                        elapsed_ms = deadline.elapsed().as_millis() as u64,
+                        "ME hybrid route timeout reached"
+                    );
+                    break;
+                }
+                Err(actual) => last_warn_ms = actual,
+            }
+        }
+    }
+
+    pub(super) fn try_consume_hybrid_recovery_trigger_slot(&self, min_interval_ms: u64) -> bool {
+        let now_ms = Self::now_epoch_millis();
+        let mut last_trigger_ms = self
+            .route_runtime
+            .me_async_recovery_last_trigger_epoch_ms
+            .load(Ordering::Relaxed);
+        loop {
+            if now_ms.saturating_sub(last_trigger_ms) < min_interval_ms {
+                return false;
+            }
+            match self
+                .route_runtime
+                .me_async_recovery_last_trigger_epoch_ms
+                .compare_exchange_weak(last_trigger_ms, now_ms, Ordering::AcqRel, Ordering::Relaxed)
+            {
+                Ok(_) => return true,
+                Err(actual) => last_trigger_ms = actual,
+            }
+        }
+    }
+}

src/transport/middle_proxy/send/selection.rs

@@ -0,0 +1,165 @@
+use std::collections::{HashMap, HashSet};
+use std::sync::atomic::Ordering;
+
+use super::super::MePool;
+use super::super::pool::WriterContour;
+use super::{
+    IDLE_WRITER_PENALTY_HIGH_SECS, IDLE_WRITER_PENALTY_MID_SECS, PICK_PENALTY_DEGRADED,
+    PICK_PENALTY_DRAINING, PICK_PENALTY_STALE, PICK_PENALTY_WARM,
+};
+
+impl MePool {
+    pub(super) async fn candidate_indices_for_dc(
+        &self,
+        writers: &[super::super::pool::MeWriter],
+        routed_dc: i32,
+        include_warm: bool,
+    ) -> Vec<usize> {
+        let preferred = self.preferred_endpoints_for_dc(routed_dc).await;
+        if preferred.is_empty() {
+            return Vec::new();
+        }
+
+        let mut out = Vec::new();
+        for (idx, w) in writers.iter().enumerate() {
+            if !self.writer_eligible_for_selection(w, include_warm) {
+                continue;
+            }
+            if w.writer_dc == routed_dc && preferred.contains(&w.addr) {
+                out.push(idx);
+            }
+        }
+        out
+    }
+
+    pub(super) fn writer_eligible_for_selection(
+        &self,
+        writer: &super::super::pool::MeWriter,
+        include_warm: bool,
+    ) -> bool {
+        if !self.writer_accepts_new_binding(writer) {
+            return false;
+        }
+
+        match WriterContour::from_u8(writer.contour.load(Ordering::Relaxed)) {
+            WriterContour::Active => true,
+            WriterContour::Warm => include_warm,
+            WriterContour::Draining => true,
+        }
+    }
+
+    pub(super) fn writer_contour_rank_for_selection(
+        &self,
+        writer: &super::super::pool::MeWriter,
+    ) -> usize {
+        match WriterContour::from_u8(writer.contour.load(Ordering::Relaxed)) {
+            WriterContour::Active => 0,
+            WriterContour::Warm => 1,
+            WriterContour::Draining => 2,
+        }
+    }
+
+    pub(super) fn writer_idle_rank_for_selection(
+        &self,
+        writer: &super::super::pool::MeWriter,
+        idle_since_by_writer: &HashMap<u64, u64>,
+        now_epoch_secs: u64,
+    ) -> usize {
+        let Some(idle_since) = idle_since_by_writer.get(&writer.id).copied() else {
+            return 0;
+        };
+        let idle_age_secs = now_epoch_secs.saturating_sub(idle_since);
+        if idle_age_secs >= IDLE_WRITER_PENALTY_HIGH_SECS {
+            2
+        } else if idle_age_secs >= IDLE_WRITER_PENALTY_MID_SECS {
+            1
+        } else {
+            0
+        }
+    }
+
+    pub(super) fn writer_pick_score(
+        &self,
+        writer: &super::super::pool::MeWriter,
+        idle_since_by_writer: &HashMap<u64, u64>,
+        now_epoch_secs: u64,
+    ) -> u64 {
+        let contour_penalty = match WriterContour::from_u8(writer.contour.load(Ordering::Relaxed)) {
+            WriterContour::Active => 0,
+            WriterContour::Warm => PICK_PENALTY_WARM,
+            WriterContour::Draining => PICK_PENALTY_DRAINING,
+        };
+        let stale_penalty = if writer.generation < self.current_generation() {
+            PICK_PENALTY_STALE
+        } else {
+            0
+        };
+        let degraded_penalty = if writer.degraded.load(Ordering::Relaxed) {
+            PICK_PENALTY_DEGRADED
+        } else {
+            0
+        };
+        let idle_penalty =
+            (self.writer_idle_rank_for_selection(writer, idle_since_by_writer, now_epoch_secs)
+                as u64)
+                * 100;
+        let queue_cap = self.writer_lifecycle.writer_cmd_channel_capacity.max(1) as u64;
+        let queue_remaining = writer.tx.capacity() as u64;
+        let queue_used = queue_cap.saturating_sub(queue_remaining.min(queue_cap));
+        let queue_util_pct = queue_used.saturating_mul(100) / queue_cap;
+        let queue_penalty = queue_util_pct.saturating_mul(4);
+        let rtt_penalty =
+            ((writer.rtt_ema_ms_x10.load(Ordering::Relaxed) as u64).saturating_add(5) / 10)
+                .min(400);
+
+        contour_penalty
+            .saturating_add(stale_penalty)
+            .saturating_add(degraded_penalty)
+            .saturating_add(idle_penalty)
+            .saturating_add(queue_penalty)
+            .saturating_add(rtt_penalty)
+    }
+
+    pub(super) fn p2c_ordered_candidate_indices(
+        &self,
+        candidate_indices: &[usize],
+        writers_snapshot: &[super::super::pool::MeWriter],
+        idle_since_by_writer: &HashMap<u64, u64>,
+        now_epoch_secs: u64,
+        start: usize,
+        sample_size: usize,
+    ) -> Vec<usize> {
+        let total = candidate_indices.len();
+        if total == 0 {
+            return Vec::new();
+        }
+
+        let mut sampled = Vec::<usize>::with_capacity(sample_size.min(total));
+        let mut seen = HashSet::<usize>::with_capacity(total);
+        for offset in 0..sample_size.min(total) {
+            let idx = candidate_indices[(start + offset) % total];
+            if seen.insert(idx) {
+                sampled.push(idx);
+            }
+        }
+
+        sampled.sort_by_key(|idx| {
+            let writer = &writers_snapshot[*idx];
+            (
+                self.writer_pick_score(writer, idle_since_by_writer, now_epoch_secs),
+                writer.addr,
+                writer.id,
+            )
+        });
+
+        let mut ordered = Vec::<usize>::with_capacity(total);
+        ordered.extend(sampled.iter().copied());
+        for offset in 0..total {
+            let idx = candidate_indices[(start + offset) % total];
+            if seen.insert(idx) {
+                ordered.push(idx);
+            }
+        }
+        ordered
+    }
+}

src/transport/middle_proxy/tests/send_adversarial_tests.rs

@@ -165,6 +165,8 @@ async fn recv_data_count(rx: &mut mpsc::Receiver<WriterCommand>, budget: Duratio
         match tokio::time::timeout(remaining.min(Duration::from_millis(10)), rx.recv()).await {
             Ok(Some(WriterCommand::Data(_))) => data_count += 1,
             Ok(Some(WriterCommand::DataAndFlush(_))) => data_count += 1,
+            Ok(Some(WriterCommand::ProxyReq(_))) => data_count += 1,
+            Ok(Some(WriterCommand::ControlAndFlush(_))) => data_count += 1,
             Ok(Some(WriterCommand::Close)) => {}
             Ok(None) => break,
             Err(_) => break,

src/transport/middle_proxy/wire.rs

@@ -42,22 +42,45 @@ fn append_mapped_addr_and_port(buf: &mut Vec<u8>, addr: SocketAddr) {
     buf.extend_from_slice(&(addr.port() as u32).to_le_bytes());
 }
 
-pub(crate) fn build_proxy_req_payload(
+fn proxy_tag_wire_len(tag: &[u8]) -> usize {
+    if tag.len() < 254 {
+        4 + 1 + tag.len() + ((4 - ((1 + tag.len()) % 4)) % 4)
+    } else {
+        4 + 4 + tag.len() + ((4 - (tag.len() % 4)) % 4)
+    }
+}
+
+/// Returns the exact unencrypted RPC_PROXY_REQ payload length for pre-sizing frame buffers.
+pub(crate) fn proxy_req_payload_len(
+    data_len: usize,
+    proxy_tag: Option<&[u8]>,
+    proto_flags: u32,
+) -> usize {
+    let base_len = 4 + 4 + 8 + 20 + 20;
+    let extra_len = if proto_flags & RPC_FLAG_HAS_AD_TAG != 0 {
+        4 + proxy_tag.map(proxy_tag_wire_len).unwrap_or(0)
+    } else {
+        0
+    };
+    base_len + extra_len + data_len
+}
+
+/// Appends RPC_PROXY_REQ payload bytes without allocating an intermediate payload buffer.
+pub(crate) fn append_proxy_req_payload_into(
+    b: &mut Vec<u8>,
     conn_id: u64,
     client_addr: SocketAddr,
     our_addr: SocketAddr,
     data: &[u8],
     proxy_tag: Option<&[u8]>,
     proto_flags: u32,
-) -> Bytes {
-    let mut b = Vec::with_capacity(128 + data.len());
-
+) {
     b.extend_from_slice(&RPC_PROXY_REQ_U32.to_le_bytes());
     b.extend_from_slice(&proto_flags.to_le_bytes());
     b.extend_from_slice(&conn_id.to_le_bytes());
 
-    append_mapped_addr_and_port(&mut b, client_addr);
-    append_mapped_addr_and_port(&mut b, our_addr);
+    append_mapped_addr_and_port(b, client_addr);
+    append_mapped_addr_and_port(b, our_addr);
 
     if proto_flags & RPC_FLAG_HAS_AD_TAG != 0 {
         let extra_start = b.len();
@@ -86,6 +109,26 @@ pub(crate) fn build_proxy_req_payload(
     }
 
     b.extend_from_slice(data);
+}
+
+pub(crate) fn build_proxy_req_payload(
+    conn_id: u64,
+    client_addr: SocketAddr,
+    our_addr: SocketAddr,
+    data: &[u8],
+    proxy_tag: Option<&[u8]>,
+    proto_flags: u32,
+) -> Bytes {
+    let mut b = Vec::with_capacity(128 + data.len());
+    append_proxy_req_payload_into(
+        &mut b,
+        conn_id,
+        client_addr,
+        our_addr,
+        data,
+        proxy_tag,
+        proto_flags,
+    );
     Bytes::from(b)
 }