Merge 062464175e into 30ba41eb47

Update Cargo.toml

Cargo.lock

@@ -2611,7 +2611,7 @@ checksum = "7b2093cf4c8eb1e67749a6762251bc9cd836b6fc171623bd0a9d324d37af2417"
 
 [[package]]
 name = "telemt"
-version = "3.3.18"
+version = "3.3.20"
 dependencies = [
  "aes",
  "anyhow",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "telemt"
-version = "3.3.18"
+version = "3.3.21"
 edition = "2024"
 
 [dependencies]

config.toml

@@ -32,6 +32,7 @@ show = "*"
 port = 443
 # proxy_protocol = false           # Enable if behind HAProxy/nginx with PROXY protocol
 # metrics_port = 9090
+# metrics_listen = "0.0.0.0:9090"  # Listen address for metrics (overrides metrics_port)
 # metrics_whitelist = ["127.0.0.1", "::1", "0.0.0.0/0"]
 
 [server.api]

docs/CONFIG_PARAMS.en.md

@@ -0,0 +1,289 @@
+# Telemt Config Parameters Reference
+
+This document lists all configuration keys accepted by `config.toml`.
+
+> [!WARNING]
+> 
+> The configuration parameters detailed in this document are intended for advanced users and fine-tuning purposes. Modifying these settings without a clear understanding of their function may lead to application instability or other unexpected behavior. Please proceed with caution and at your own risk.
+
+## Top-level keys
+
+| Parameter | Type | Description |
+|---|---|---|
+| include | `String` (special directive) | Includes another TOML file with `include = "relative/or/absolute/path.toml"`; includes are processed recursively before parsing. |
+| show_link | `"*" \| String[]` | Legacy top-level link visibility selector (`"*"` for all users or explicit usernames list). |
+| dc_overrides | `Map<String, String[]>` | Overrides DC endpoints for non-standard DCs; key is DC id string, value is `ip:port` list. |
+| default_dc | `u8` | Default DC index used for unmapped non-standard DCs. |
+
+## [general]
+
+| Parameter | Type | Description |
+|---|---|---|
+| data_path | `String` | Optional runtime data directory path. |
+| prefer_ipv6 | `bool` | Prefer IPv6 where applicable in runtime logic. |
+| fast_mode | `bool` | Enables fast-path optimizations for traffic processing. |
+| use_middle_proxy | `bool` | Enables Middle Proxy mode. |
+| proxy_secret_path | `String` | Path to proxy secret binary; can be auto-downloaded if absent. |
+| proxy_config_v4_cache_path | `String` | Optional cache path for raw `getProxyConfig` (IPv4) snapshot. |
+| proxy_config_v6_cache_path | `String` | Optional cache path for raw `getProxyConfigV6` (IPv6) snapshot. |
+| ad_tag | `String` | Global fallback ad tag (32 hex characters). |
+| middle_proxy_nat_ip | `IpAddr` | Explicit public IP override for NAT environments. |
+| middle_proxy_nat_probe | `bool` | Enables NAT probing for Middle Proxy KDF/public address discovery. |
+| middle_proxy_nat_stun | `String` | Deprecated legacy single STUN server for NAT probing. |
+| middle_proxy_nat_stun_servers | `String[]` | Deprecated legacy STUN list for NAT probing fallback. |
+| stun_nat_probe_concurrency | `usize` | Maximum concurrent STUN probes during NAT detection. |
+| middle_proxy_pool_size | `usize` | Target size of active Middle Proxy writer pool. |
+| middle_proxy_warm_standby | `usize` | Number of warm standby Middle-End connections. |
+| me_init_retry_attempts | `u32` | Startup retries for ME pool initialization (`0` means unlimited). |
+| me2dc_fallback | `bool` | Allows fallback from ME mode to direct DC when ME startup fails. |
+| me_keepalive_enabled | `bool` | Enables ME keepalive padding frames. |
+| me_keepalive_interval_secs | `u64` | Keepalive interval in seconds. |
+| me_keepalive_jitter_secs | `u64` | Keepalive jitter in seconds. |
+| me_keepalive_payload_random | `bool` | Randomizes keepalive payload bytes instead of zero payload. |
+| rpc_proxy_req_every | `u64` | Interval for service `RPC_PROXY_REQ` activity signals (`0` disables). |
+| me_writer_cmd_channel_capacity | `usize` | Capacity of per-writer command channel. |
+| me_route_channel_capacity | `usize` | Capacity of per-connection ME response route channel. |
+| me_c2me_channel_capacity | `usize` | Capacity of per-client command queue (client reader -> ME sender). |
+| me_reader_route_data_wait_ms | `u64` | Bounded wait for routing ME DATA to per-connection queue (`0` = no wait). |
+| me_d2c_flush_batch_max_frames | `usize` | Max ME->client frames coalesced before flush. |
+| me_d2c_flush_batch_max_bytes | `usize` | Max ME->client payload bytes coalesced before flush. |
+| me_d2c_flush_batch_max_delay_us | `u64` | Max microsecond wait for coalescing more ME->client frames (`0` disables timed coalescing). |
+| me_d2c_ack_flush_immediate | `bool` | Flushes client writer immediately after quick-ack write. |
+| direct_relay_copy_buf_c2s_bytes | `usize` | Copy buffer size for client->DC direction in direct relay. |
+| direct_relay_copy_buf_s2c_bytes | `usize` | Copy buffer size for DC->client direction in direct relay. |
+| crypto_pending_buffer | `usize` | Max pending ciphertext buffer per client writer (bytes). |
+| max_client_frame | `usize` | Maximum allowed client MTProto frame size (bytes). |
+| desync_all_full | `bool` | Emits full crypto-desync forensic logs for every event. |
+| beobachten | `bool` | Enables per-IP forensic observation buckets. |
+| beobachten_minutes | `u64` | Retention window (minutes) for per-IP observation buckets. |
+| beobachten_flush_secs | `u64` | Snapshot flush interval (seconds) for observation output file. |
+| beobachten_file | `String` | Observation snapshot output file path. |
+| hardswap | `bool` | Enables hard-swap generation switching for ME pool updates. |
+| me_warmup_stagger_enabled | `bool` | Enables staggered warmup for extra ME writers. |
+| me_warmup_step_delay_ms | `u64` | Base delay between warmup connections (ms). |
+| me_warmup_step_jitter_ms | `u64` | Jitter for warmup delay (ms). |
+| me_reconnect_max_concurrent_per_dc | `u32` | Max concurrent reconnect attempts per DC. |
+| me_reconnect_backoff_base_ms | `u64` | Base reconnect backoff in ms. |
+| me_reconnect_backoff_cap_ms | `u64` | Cap reconnect backoff in ms. |
+| me_reconnect_fast_retry_count | `u32` | Number of fast retry attempts before backoff. |
+| me_single_endpoint_shadow_writers | `u8` | Additional reserve writers for one-endpoint DC groups. |
+| me_single_endpoint_outage_mode_enabled | `bool` | Enables aggressive outage recovery for one-endpoint DC groups. |
+| me_single_endpoint_outage_disable_quarantine | `bool` | Ignores endpoint quarantine in one-endpoint outage mode. |
+| me_single_endpoint_outage_backoff_min_ms | `u64` | Minimum reconnect backoff in outage mode (ms). |
+| me_single_endpoint_outage_backoff_max_ms | `u64` | Maximum reconnect backoff in outage mode (ms). |
+| me_single_endpoint_shadow_rotate_every_secs | `u64` | Periodic shadow writer rotation interval (`0` disables). |
+| me_floor_mode | `"static" \| "adaptive"` | Writer floor policy mode. |
+| me_adaptive_floor_idle_secs | `u64` | Idle time before adaptive floor may reduce one-endpoint target. |
+| me_adaptive_floor_min_writers_single_endpoint | `u8` | Minimum adaptive writer target for one-endpoint DC groups. |
+| me_adaptive_floor_min_writers_multi_endpoint | `u8` | Minimum adaptive writer target for multi-endpoint DC groups. |
+| me_adaptive_floor_recover_grace_secs | `u64` | Grace period to hold static floor after activity. |
+| me_adaptive_floor_writers_per_core_total | `u16` | Global writer budget per logical CPU core in adaptive mode. |
+| me_adaptive_floor_cpu_cores_override | `u16` | Manual CPU core count override (`0` uses auto-detection). |
+| me_adaptive_floor_max_extra_writers_single_per_core | `u16` | Per-core max extra writers above base floor for one-endpoint DCs. |
+| me_adaptive_floor_max_extra_writers_multi_per_core | `u16` | Per-core max extra writers above base floor for multi-endpoint DCs. |
+| me_adaptive_floor_max_active_writers_per_core | `u16` | Hard cap for active ME writers per logical CPU core. |
+| me_adaptive_floor_max_warm_writers_per_core | `u16` | Hard cap for warm ME writers per logical CPU core. |
+| me_adaptive_floor_max_active_writers_global | `u32` | Hard global cap for active ME writers. |
+| me_adaptive_floor_max_warm_writers_global | `u32` | Hard global cap for warm ME writers. |
+| upstream_connect_retry_attempts | `u32` | Connect attempts for selected upstream before error/fallback. |
+| upstream_connect_retry_backoff_ms | `u64` | Delay between upstream connect attempts (ms). |
+| upstream_connect_budget_ms | `u64` | Total wall-clock budget for one upstream connect request (ms). |
+| upstream_unhealthy_fail_threshold | `u32` | Consecutive failed requests before upstream is marked unhealthy. |
+| upstream_connect_failfast_hard_errors | `bool` | Skips additional retries for hard non-transient connect errors. |
+| stun_iface_mismatch_ignore | `bool` | Ignores STUN/interface mismatch and keeps Middle Proxy mode. |
+| unknown_dc_log_path | `String` | File path for unknown-DC request logging (`null` disables file path). |
+| unknown_dc_file_log_enabled | `bool` | Enables unknown-DC file logging. |
+| log_level | `"debug" \| "verbose" \| "normal" \| "silent"` | Runtime logging verbosity. |
+| disable_colors | `bool` | Disables ANSI colors in logs. |
+| me_socks_kdf_policy | `"strict" \| "compat"` | SOCKS-bound KDF fallback policy for ME handshake. |
+| me_route_backpressure_base_timeout_ms | `u64` | Base backpressure timeout for route-channel send (ms). |
+| me_route_backpressure_high_timeout_ms | `u64` | High backpressure timeout when queue occupancy exceeds watermark (ms). |
+| me_route_backpressure_high_watermark_pct | `u8` | Queue occupancy threshold (%) for high timeout mode. |
+| me_health_interval_ms_unhealthy | `u64` | Health monitor interval while writer coverage is degraded (ms). |
+| me_health_interval_ms_healthy | `u64` | Health monitor interval while writer coverage is healthy (ms). |
+| me_admission_poll_ms | `u64` | Poll interval for conditional-admission checks (ms). |
+| me_warn_rate_limit_ms | `u64` | Cooldown for repetitive ME warning logs (ms). |
+| me_route_no_writer_mode | `"async_recovery_failfast" \| "inline_recovery_legacy" \| "hybrid_async_persistent"` | Route behavior when no writer is immediately available. |
+| me_route_no_writer_wait_ms | `u64` | Max wait in async-recovery failfast mode (ms). |
+| me_route_inline_recovery_attempts | `u32` | Inline recovery attempts in legacy mode. |
+| me_route_inline_recovery_wait_ms | `u64` | Max inline recovery wait in legacy mode (ms). |
+| fast_mode_min_tls_record | `usize` | Minimum TLS record size when fast-mode coalescing is enabled (`0` disables). |
+| update_every | `u64` | Unified interval for config/secret updater tasks. |
+| me_reinit_every_secs | `u64` | Periodic ME pool reinitialization interval (seconds). |
+| me_hardswap_warmup_delay_min_ms | `u64` | Minimum delay between hardswap warmup connects (ms). |
+| me_hardswap_warmup_delay_max_ms | `u64` | Maximum delay between hardswap warmup connects (ms). |
+| me_hardswap_warmup_extra_passes | `u8` | Additional warmup passes per hardswap cycle. |
+| me_hardswap_warmup_pass_backoff_base_ms | `u64` | Base backoff between hardswap warmup passes (ms). |
+| me_config_stable_snapshots | `u8` | Number of identical config snapshots required before apply. |
+| me_config_apply_cooldown_secs | `u64` | Cooldown between applied ME map updates (seconds). |
+| me_snapshot_require_http_2xx | `bool` | Requires 2xx HTTP responses for applying config snapshots. |
+| me_snapshot_reject_empty_map | `bool` | Rejects empty config snapshots. |
+| me_snapshot_min_proxy_for_lines | `u32` | Minimum parsed `proxy_for` rows required to accept snapshot. |
+| proxy_secret_stable_snapshots | `u8` | Number of identical secret snapshots required before runtime rotation. |
+| proxy_secret_rotate_runtime | `bool` | Enables runtime proxy-secret rotation from remote source. |
+| me_secret_atomic_snapshot | `bool` | Keeps selector and secret bytes from the same snapshot atomically. |
+| proxy_secret_len_max | `usize` | Maximum allowed proxy-secret length (bytes). |
+| me_pool_drain_ttl_secs | `u64` | Drain TTL for stale ME writers after endpoint-map changes (seconds). |
+| me_pool_drain_threshold | `u64` | Max draining stale writers before batch force-close (`0` disables threshold cleanup). |
+| me_bind_stale_mode | `"never" \| "ttl" \| "always"` | Policy for new binds on stale draining writers. |
+| me_bind_stale_ttl_secs | `u64` | TTL for stale bind allowance when stale mode is `ttl`. |
+| me_pool_min_fresh_ratio | `f32` | Minimum desired-DC fresh coverage ratio before draining stale writers. |
+| me_reinit_drain_timeout_secs | `u64` | Force-close timeout for stale writers after endpoint-map changes (`0` disables force-close). |
+| proxy_secret_auto_reload_secs | `u64` | Deprecated legacy secret reload interval (fallback when `update_every` is not set). |
+| proxy_config_auto_reload_secs | `u64` | Deprecated legacy config reload interval (fallback when `update_every` is not set). |
+| me_reinit_singleflight | `bool` | Serializes ME reinit cycles across trigger sources. |
+| me_reinit_trigger_channel | `usize` | Trigger queue capacity for reinit scheduler. |
+| me_reinit_coalesce_window_ms | `u64` | Trigger coalescing window before starting reinit (ms). |
+| me_deterministic_writer_sort | `bool` | Enables deterministic candidate sort for writer binding path. |
+| me_writer_pick_mode | `"sorted_rr" \| "p2c"` | Writer selection mode for route bind path. |
+| me_writer_pick_sample_size | `u8` | Number of candidates sampled by picker in `p2c` mode. |
+| ntp_check | `bool` | Enables NTP drift check at startup. |
+| ntp_servers | `String[]` | NTP servers used for drift check. |
+| auto_degradation_enabled | `bool` | Enables automatic degradation from ME to direct DC. |
+| degradation_min_unavailable_dc_groups | `u8` | Minimum unavailable ME DC groups required before degrading. |
+
+## [general.modes]
+
+| Parameter | Type | Description |
+|---|---|---|
+| classic | `bool` | Enables classic MTProxy mode. |
+| secure | `bool` | Enables secure mode. |
+| tls | `bool` | Enables TLS mode. |
+
+## [general.links]
+
+| Parameter | Type | Description |
+|---|---|---|
+| show | `"*" \| String[]` | Selects users whose tg:// links are shown at startup. |
+| public_host | `String` | Public hostname/IP override for generated tg:// links. |
+| public_port | `u16` | Public port override for generated tg:// links. |
+
+## [general.telemetry]
+
+| Parameter | Type | Description |
+|---|---|---|
+| core_enabled | `bool` | Enables core hot-path telemetry counters. |
+| user_enabled | `bool` | Enables per-user telemetry counters. |
+| me_level | `"silent" \| "normal" \| "debug"` | Middle-End telemetry verbosity level. |
+
+## [network]
+
+| Parameter | Type | Description |
+|---|---|---|
+| ipv4 | `bool` | Enables IPv4 networking. |
+| ipv6 | `bool` | Enables/disables IPv6 (`null` = auto-detect availability). |
+| prefer | `u8` | Preferred IP family for selection (`4` or `6`). |
+| multipath | `bool` | Enables multipath behavior where supported. |
+| stun_use | `bool` | Global switch for STUN probing. |
+| stun_servers | `String[]` | STUN server list for public IP detection. |
+| stun_tcp_fallback | `bool` | Enables TCP STUN fallback when UDP STUN is blocked. |
+| http_ip_detect_urls | `String[]` | HTTP endpoints used as fallback public IP detectors. |
+| cache_public_ip_path | `String` | File path for caching detected public IP. |
+| dns_overrides | `String[]` | Runtime DNS overrides in `host:port:ip` format. |
+
+## [server]
+
+| Parameter | Type | Description |
+|---|---|---|
+| port | `u16` | Main proxy listen port. |
+| listen_addr_ipv4 | `String` | IPv4 bind address for TCP listener. |
+| listen_addr_ipv6 | `String` | IPv6 bind address for TCP listener. |
+| listen_unix_sock | `String` | Unix socket path for listener. |
+| listen_unix_sock_perm | `String` | Unix socket permissions in octal string (e.g., `"0666"`). |
+| listen_tcp | `bool` | Explicit TCP listener enable/disable override. |
+| proxy_protocol | `bool` | Enables HAProxy PROXY protocol parsing on incoming client connections. |
+| proxy_protocol_header_timeout_ms | `u64` | Timeout for PROXY protocol header read/parse (ms). |
+| metrics_port | `u16` | Metrics endpoint port (enables metrics listener). |
+| metrics_listen | `String` | Full metrics bind address (`IP:PORT`), overrides `metrics_port`. |
+| metrics_whitelist | `IpNetwork[]` | CIDR whitelist for metrics endpoint access. |
+| max_connections | `u32` | Max concurrent client connections (`0` = unlimited). |
+
+## [server.api]
+
+| Parameter | Type | Description |
+|---|---|---|
+| enabled | `bool` | Enables control-plane REST API. |
+| listen | `String` | API bind address in `IP:PORT` format. |
+| whitelist | `IpNetwork[]` | CIDR whitelist allowed to access API. |
+| auth_header | `String` | Exact expected `Authorization` header value (empty = disabled). |
+| request_body_limit_bytes | `usize` | Maximum accepted HTTP request body size. |
+| minimal_runtime_enabled | `bool` | Enables minimal runtime snapshots endpoint logic. |
+| minimal_runtime_cache_ttl_ms | `u64` | Cache TTL for minimal runtime snapshots (ms; `0` disables cache). |
+| runtime_edge_enabled | `bool` | Enables runtime edge endpoints. |
+| runtime_edge_cache_ttl_ms | `u64` | Cache TTL for runtime edge aggregation payloads (ms). |
+| runtime_edge_top_n | `usize` | Top-N size for edge connection leaderboard. |
+| runtime_edge_events_capacity | `usize` | Ring-buffer capacity for runtime edge events. |
+| read_only | `bool` | Rejects mutating API endpoints when enabled. |
+
+## [[server.listeners]]
+
+| Parameter | Type | Description |
+|---|---|---|
+| ip | `IpAddr` | Listener bind IP. |
+| announce | `String` | Public IP/domain announced in proxy links (priority over `announce_ip`). |
+| announce_ip | `IpAddr` | Deprecated legacy announce IP (migrated to `announce` if needed). |
+| proxy_protocol | `bool` | Per-listener override for PROXY protocol enable flag. |
+| reuse_allow | `bool` | Enables `SO_REUSEPORT` for multi-instance bind sharing. |
+
+## [timeouts]
+
+| Parameter | Type | Description |
+|---|---|---|
+| client_handshake | `u64` | Client handshake timeout. |
+| tg_connect | `u64` | Upstream Telegram connect timeout. |
+| client_keepalive | `u64` | Client keepalive timeout. |
+| client_ack | `u64` | Client ACK timeout. |
+| me_one_retry | `u8` | Quick ME reconnect attempts for single-address DC. |
+| me_one_timeout_ms | `u64` | Timeout per quick attempt for single-address DC (ms). |
+
+## [censorship]
+
+| Parameter | Type | Description |
+|---|---|---|
+| tls_domain | `String` | Primary TLS domain used in fake TLS handshake profile. |
+| tls_domains | `String[]` | Additional TLS domains for generating multiple links. |
+| mask | `bool` | Enables masking/fronting relay mode. |
+| mask_host | `String` | Upstream mask host for TLS fronting relay. |
+| mask_port | `u16` | Upstream mask port for TLS fronting relay. |
+| mask_unix_sock | `String` | Unix socket path for mask backend instead of TCP host/port. |
+| fake_cert_len | `usize` | Length of synthetic certificate payload when emulation data is unavailable. |
+| tls_emulation | `bool` | Enables certificate/TLS behavior emulation from cached real fronts. |
+| tls_front_dir | `String` | Directory path for TLS front cache storage. |
+| server_hello_delay_min_ms | `u64` | Minimum server_hello delay for anti-fingerprint behavior (ms). |
+| server_hello_delay_max_ms | `u64` | Maximum server_hello delay for anti-fingerprint behavior (ms). |
+| tls_new_session_tickets | `u8` | Number of `NewSessionTicket` messages to emit after handshake. |
+| tls_full_cert_ttl_secs | `u64` | TTL for sending full cert payload per (domain, client IP) tuple. |
+| alpn_enforce | `bool` | Enforces ALPN echo behavior based on client preference. |
+| mask_proxy_protocol | `u8` | PROXY protocol mode for mask backend (`0` disabled, `1` v1, `2` v2). |
+
+## [access]
+
+| Parameter | Type | Description |
+|---|---|---|
+| users | `Map<String, String>` | Username -> 32-hex secret mapping. |
+| user_ad_tags | `Map<String, String>` | Per-user ad tags (32 hex chars). |
+| user_max_tcp_conns | `Map<String, usize>` | Per-user maximum concurrent TCP connections. |
+| user_expirations | `Map<String, DateTime<Utc>>` | Per-user account expiration timestamps. |
+| user_data_quota | `Map<String, u64>` | Per-user data quota limits. |
+| user_max_unique_ips | `Map<String, usize>` | Per-user unique source IP limits. |
+| user_max_unique_ips_global_each | `usize` | Global fallback per-user unique IP limit when no per-user override exists. |
+| user_max_unique_ips_mode | `"active_window" \| "time_window" \| "combined"` | Unique source IP limit accounting mode. |
+| user_max_unique_ips_window_secs | `u64` | Recent-window size for unique IP accounting (seconds). |
+| replay_check_len | `usize` | Replay check storage length. |
+| replay_window_secs | `u64` | Replay protection time window in seconds. |
+| ignore_time_skew | `bool` | Ignores client/server timestamp skew in replay validation. |
+
+## [[upstreams]]
+
+| Parameter | Type | Description |
+|---|---|---|
+| type | `"direct" \| "socks4" \| "socks5"` | Upstream transport type selector. |
+| weight | `u16` | Weighted selection coefficient for this upstream. |
+| enabled | `bool` | Enables/disables this upstream entry. |
+| scopes | `String` | Comma-separated scope tags for routing. |
+| interface | `String` | Optional outgoing interface name (`direct`, `socks4`, `socks5`). |
+| bind_addresses | `String[]` | Optional source bind addresses for `direct` upstream. |
+| address | `String` | Upstream proxy address (`host:port`) for SOCKS upstreams. |
+| user_id | `String` | SOCKS4 user ID (only for `type = "socks4"`). |
+| username | `String` | SOCKS5 username (only for `type = "socks5"`). |
+| password | `String` | SOCKS5 password (only for `type = "socks5"`). |

docs/FAQ.en.md

@@ -83,6 +83,13 @@ To specify a domain in the links, add to the `[general.links]` section of the co
 public_host = "proxy.example.com"
 ```
 
+### Server connection limit
+Limits the total number of open connections to the server:
+```toml
+[server]
+max_connections = 10000    # 0 - unlimited, 10000 - default
+```
+
 ### Upstream Manager
 To specify an upstream, add to the `[[upstreams]]` section of the config.toml file:
 #### Binding to IP

docs/FAQ.ru.md

@@ -83,6 +83,13 @@ metrics_whitelist = ["127.0.0.1/32", "::1/128", "0.0.0.0/0"]
 public_host = "proxy.example.com"
 ```
 
+### Общий лимит подключений к серверу
+Ограничивает общее число открытых подключений к серверу:
+```toml
+[server]
+max_connections = 10000    # 0 - unlimited, 10000 - default
+```
+
 ### Upstream Manager
 Чтобы указать апстрим, добавьте в секцию `[[upstreams]]` файла config.toml:
 #### Привязка к IP

docs/fronting-splitting/TLS-F-TCP-S.ru.md

@@ -38,8 +38,9 @@ umweltschutz.de -> A-запись 198.18.88.88
 
 В конфигурации Telemt:
 
-```
-tls_domain = umweltschutz.de
+```toml
+[censorship]
+tls_domain = "umweltschutz.de"
 ```
 
 Этот домен используется клиентом как SNI в ClientHello
@@ -56,8 +57,9 @@ tls_domain = umweltschutz.de
 
 В конфигурации Telemt:
 
-```
-mask_host = 127.0.0.1
+```toml
+[censorship]
+mask_host = "127.0.0.1"
 mask_port = 8443
 ```
 
@@ -151,16 +153,18 @@ mask_host:mask_port
 
 Например:
 
-```
-tls_domain = github.com
-mask_host = github.com
+```toml
+[censorship]
+tls_domain = "github.com"
+mask_host = "github.com"
 mask_port = 443
 ```
 
 или
 
-```
-mask_host = 140.82.121.4
+```toml
+[censorship]
+mask_host = "140.82.121.4"
 ```
 
 В этом случае:

src/api/model.rs

@@ -196,6 +196,8 @@ pub(super) struct ZeroPoolData {
     pub(super) pool_swap_total: u64,
     pub(super) pool_drain_active: u64,
     pub(super) pool_force_close_total: u64,
+    pub(super) pool_drain_soft_evict_total: u64,
+    pub(super) pool_drain_soft_evict_writer_total: u64,
     pub(super) pool_stale_pick_total: u64,
     pub(super) writer_removed_total: u64,
     pub(super) writer_removed_unexpected_total: u64,
@@ -236,6 +238,7 @@ pub(super) struct MeWritersSummary {
     pub(super) available_pct: f64,
     pub(super) required_writers: usize,
     pub(super) alive_writers: usize,
+    pub(super) coverage_ratio: f64,
     pub(super) coverage_pct: f64,
     pub(super) fresh_alive_writers: usize,
     pub(super) fresh_coverage_pct: f64,
@@ -284,6 +287,7 @@ pub(super) struct DcStatus {
     pub(super) floor_max: usize,
     pub(super) floor_capped: bool,
     pub(super) alive_writers: usize,
+    pub(super) coverage_ratio: f64,
     pub(super) coverage_pct: f64,
     pub(super) fresh_alive_writers: usize,
     pub(super) fresh_coverage_pct: f64,
@@ -361,6 +365,11 @@ pub(super) struct MinimalMeRuntimeData {
     pub(super) me_reconnect_backoff_cap_ms: u64,
     pub(super) me_reconnect_fast_retry_count: u32,
     pub(super) me_pool_drain_ttl_secs: u64,
+    pub(super) me_pool_drain_soft_evict_enabled: bool,
+    pub(super) me_pool_drain_soft_evict_grace_secs: u64,
+    pub(super) me_pool_drain_soft_evict_per_writer: u8,
+    pub(super) me_pool_drain_soft_evict_budget_per_core: u16,
+    pub(super) me_pool_drain_soft_evict_cooldown_ms: u64,
     pub(super) me_pool_force_close_secs: u64,
     pub(super) me_pool_min_fresh_ratio: f32,
     pub(super) me_bind_stale_mode: &'static str,

src/api/runtime_min.rs

@@ -113,6 +113,7 @@ pub(super) struct RuntimeMeQualityDcRttData {
     pub(super) rtt_ema_ms: Option<f64>,
     pub(super) alive_writers: usize,
     pub(super) required_writers: usize,
+    pub(super) coverage_ratio: f64,
     pub(super) coverage_pct: f64,
 }
 
@@ -389,6 +390,7 @@ pub(super) async fn build_runtime_me_quality_data(shared: &ApiShared) -> Runtime
                     rtt_ema_ms: dc.rtt_ms,
                     alive_writers: dc.alive_writers,
                     required_writers: dc.required_writers,
+                    coverage_ratio: dc.coverage_ratio,
                     coverage_pct: dc.coverage_pct,
                 })
                 .collect(),

src/api/runtime_stats.rs

@@ -96,6 +96,8 @@ pub(super) fn build_zero_all_data(stats: &Stats, configured_users: usize) -> Zer
             pool_swap_total: stats.get_pool_swap_total(),
             pool_drain_active: stats.get_pool_drain_active(),
             pool_force_close_total: stats.get_pool_force_close_total(),
+            pool_drain_soft_evict_total: stats.get_pool_drain_soft_evict_total(),
+            pool_drain_soft_evict_writer_total: stats.get_pool_drain_soft_evict_writer_total(),
             pool_stale_pick_total: stats.get_pool_stale_pick_total(),
             writer_removed_total: stats.get_me_writer_removed_total(),
             writer_removed_unexpected_total: stats.get_me_writer_removed_unexpected_total(),
@@ -315,6 +317,7 @@ async fn get_minimal_payload_cached(
             available_pct: status.available_pct,
             required_writers: status.required_writers,
             alive_writers: status.alive_writers,
+            coverage_ratio: status.coverage_ratio,
             coverage_pct: status.coverage_pct,
             fresh_alive_writers: status.fresh_alive_writers,
             fresh_coverage_pct: status.fresh_coverage_pct,
@@ -372,6 +375,7 @@ async fn get_minimal_payload_cached(
                 floor_max: entry.floor_max,
                 floor_capped: entry.floor_capped,
                 alive_writers: entry.alive_writers,
+                coverage_ratio: entry.coverage_ratio,
                 coverage_pct: entry.coverage_pct,
                 fresh_alive_writers: entry.fresh_alive_writers,
                 fresh_coverage_pct: entry.fresh_coverage_pct,
@@ -425,6 +429,11 @@ async fn get_minimal_payload_cached(
         me_reconnect_backoff_cap_ms: runtime.me_reconnect_backoff_cap_ms,
         me_reconnect_fast_retry_count: runtime.me_reconnect_fast_retry_count,
         me_pool_drain_ttl_secs: runtime.me_pool_drain_ttl_secs,
+        me_pool_drain_soft_evict_enabled: runtime.me_pool_drain_soft_evict_enabled,
+        me_pool_drain_soft_evict_grace_secs: runtime.me_pool_drain_soft_evict_grace_secs,
+        me_pool_drain_soft_evict_per_writer: runtime.me_pool_drain_soft_evict_per_writer,
+        me_pool_drain_soft_evict_budget_per_core: runtime.me_pool_drain_soft_evict_budget_per_core,
+        me_pool_drain_soft_evict_cooldown_ms: runtime.me_pool_drain_soft_evict_cooldown_ms,
         me_pool_force_close_secs: runtime.me_pool_force_close_secs,
         me_pool_min_fresh_ratio: runtime.me_pool_min_fresh_ratio,
         me_bind_stale_mode: runtime.me_bind_stale_mode,
@@ -493,6 +502,7 @@ fn disabled_me_writers(now_epoch_secs: u64, reason: &'static str) -> MeWritersDa
             available_pct: 0.0,
             required_writers: 0,
             alive_writers: 0,
+            coverage_ratio: 0.0,
             coverage_pct: 0.0,
             fresh_alive_writers: 0,
             fresh_coverage_pct: 0.0,

src/config/defaults.rs

@@ -27,8 +27,8 @@ const DEFAULT_ME_C2ME_CHANNEL_CAPACITY: usize = 1024;
 const DEFAULT_ME_READER_ROUTE_DATA_WAIT_MS: u64 = 2;
 const DEFAULT_ME_D2C_FLUSH_BATCH_MAX_FRAMES: usize = 32;
 const DEFAULT_ME_D2C_FLUSH_BATCH_MAX_BYTES: usize = 128 * 1024;
-const DEFAULT_ME_D2C_FLUSH_BATCH_MAX_DELAY_US: u64 = 1500;
-const DEFAULT_ME_D2C_ACK_FLUSH_IMMEDIATE: bool = false;
+const DEFAULT_ME_D2C_FLUSH_BATCH_MAX_DELAY_US: u64 = 500;
+const DEFAULT_ME_D2C_ACK_FLUSH_IMMEDIATE: bool = true;
 const DEFAULT_DIRECT_RELAY_COPY_BUF_C2S_BYTES: usize = 64 * 1024;
 const DEFAULT_DIRECT_RELAY_COPY_BUF_S2C_BYTES: usize = 256 * 1024;
 const DEFAULT_ME_WRITER_PICK_SAMPLE_SIZE: u8 = 3;
@@ -36,6 +36,11 @@ const DEFAULT_ME_HEALTH_INTERVAL_MS_UNHEALTHY: u64 = 1000;
 const DEFAULT_ME_HEALTH_INTERVAL_MS_HEALTHY: u64 = 3000;
 const DEFAULT_ME_ADMISSION_POLL_MS: u64 = 1000;
 const DEFAULT_ME_WARN_RATE_LIMIT_MS: u64 = 5000;
+const DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_ENABLED: bool = true;
+const DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_GRACE_SECS: u64 = 30;
+const DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_PER_WRITER: u8 = 1;
+const DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_BUDGET_PER_CORE: u16 = 8;
+const DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_COOLDOWN_MS: u64 = 5000;
 const DEFAULT_USER_MAX_UNIQUE_IPS_WINDOW_SECS: u64 = 30;
 const DEFAULT_UPSTREAM_CONNECT_RETRY_ATTEMPTS: u32 = 2;
 const DEFAULT_UPSTREAM_UNHEALTHY_FAIL_THRESHOLD: u32 = 5;
@@ -85,11 +90,11 @@ pub(crate) fn default_connect_timeout() -> u64 {
 }
 
 pub(crate) fn default_keepalive() -> u64 {
-    60
+    15
 }
 
 pub(crate) fn default_ack_timeout() -> u64 {
-    300
+    90
 }
 pub(crate) fn default_me_one_retry() -> u8 {
     12
@@ -592,6 +597,26 @@ pub(crate) fn default_me_pool_drain_threshold() -> u64 {
     128
 }
 
+pub(crate) fn default_me_pool_drain_soft_evict_enabled() -> bool {
+    DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_ENABLED
+}
+
+pub(crate) fn default_me_pool_drain_soft_evict_grace_secs() -> u64 {
+    DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_GRACE_SECS
+}
+
+pub(crate) fn default_me_pool_drain_soft_evict_per_writer() -> u8 {
+    DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_PER_WRITER
+}
+
+pub(crate) fn default_me_pool_drain_soft_evict_budget_per_core() -> u16 {
+    DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_BUDGET_PER_CORE
+}
+
+pub(crate) fn default_me_pool_drain_soft_evict_cooldown_ms() -> u64 {
+    DEFAULT_ME_POOL_DRAIN_SOFT_EVICT_COOLDOWN_MS
+}
+
 pub(crate) fn default_me_bind_stale_ttl_secs() -> u64 {
     default_me_pool_drain_ttl_secs()
 }

src/config/hot_reload.rs

@@ -56,6 +56,11 @@ pub struct HotFields {
     pub hardswap:                bool,
     pub me_pool_drain_ttl_secs:  u64,
     pub me_pool_drain_threshold: u64,
+    pub me_pool_drain_soft_evict_enabled: bool,
+    pub me_pool_drain_soft_evict_grace_secs: u64,
+    pub me_pool_drain_soft_evict_per_writer: u8,
+    pub me_pool_drain_soft_evict_budget_per_core: u16,
+    pub me_pool_drain_soft_evict_cooldown_ms: u64,
     pub me_pool_min_fresh_ratio: f32,
     pub me_reinit_drain_timeout_secs: u64,
     pub me_hardswap_warmup_delay_min_ms: u64,
@@ -138,6 +143,15 @@ impl HotFields {
             hardswap:                cfg.general.hardswap,
             me_pool_drain_ttl_secs:  cfg.general.me_pool_drain_ttl_secs,
             me_pool_drain_threshold: cfg.general.me_pool_drain_threshold,
+            me_pool_drain_soft_evict_enabled: cfg.general.me_pool_drain_soft_evict_enabled,
+            me_pool_drain_soft_evict_grace_secs: cfg.general.me_pool_drain_soft_evict_grace_secs,
+            me_pool_drain_soft_evict_per_writer: cfg.general.me_pool_drain_soft_evict_per_writer,
+            me_pool_drain_soft_evict_budget_per_core: cfg
+                .general
+                .me_pool_drain_soft_evict_budget_per_core,
+            me_pool_drain_soft_evict_cooldown_ms: cfg
+                .general
+                .me_pool_drain_soft_evict_cooldown_ms,
             me_pool_min_fresh_ratio: cfg.general.me_pool_min_fresh_ratio,
             me_reinit_drain_timeout_secs: cfg.general.me_reinit_drain_timeout_secs,
             me_hardswap_warmup_delay_min_ms: cfg.general.me_hardswap_warmup_delay_min_ms,
@@ -455,6 +469,15 @@ fn overlay_hot_fields(old: &ProxyConfig, new: &ProxyConfig) -> ProxyConfig {
     cfg.general.hardswap = new.general.hardswap;
     cfg.general.me_pool_drain_ttl_secs = new.general.me_pool_drain_ttl_secs;
     cfg.general.me_pool_drain_threshold = new.general.me_pool_drain_threshold;
+    cfg.general.me_pool_drain_soft_evict_enabled = new.general.me_pool_drain_soft_evict_enabled;
+    cfg.general.me_pool_drain_soft_evict_grace_secs =
+        new.general.me_pool_drain_soft_evict_grace_secs;
+    cfg.general.me_pool_drain_soft_evict_per_writer =
+        new.general.me_pool_drain_soft_evict_per_writer;
+    cfg.general.me_pool_drain_soft_evict_budget_per_core =
+        new.general.me_pool_drain_soft_evict_budget_per_core;
+    cfg.general.me_pool_drain_soft_evict_cooldown_ms =
+        new.general.me_pool_drain_soft_evict_cooldown_ms;
     cfg.general.me_pool_min_fresh_ratio = new.general.me_pool_min_fresh_ratio;
     cfg.general.me_reinit_drain_timeout_secs = new.general.me_reinit_drain_timeout_secs;
     cfg.general.me_hardswap_warmup_delay_min_ms = new.general.me_hardswap_warmup_delay_min_ms;
@@ -835,6 +858,25 @@ fn log_changes(
             old_hot.me_pool_drain_threshold, new_hot.me_pool_drain_threshold,
         );
     }
+    if old_hot.me_pool_drain_soft_evict_enabled != new_hot.me_pool_drain_soft_evict_enabled
+        || old_hot.me_pool_drain_soft_evict_grace_secs
+            != new_hot.me_pool_drain_soft_evict_grace_secs
+        || old_hot.me_pool_drain_soft_evict_per_writer
+            != new_hot.me_pool_drain_soft_evict_per_writer
+        || old_hot.me_pool_drain_soft_evict_budget_per_core
+            != new_hot.me_pool_drain_soft_evict_budget_per_core
+        || old_hot.me_pool_drain_soft_evict_cooldown_ms
+            != new_hot.me_pool_drain_soft_evict_cooldown_ms
+    {
+        info!(
+            "config reload: me_pool_drain_soft_evict: enabled={} grace={}s per_writer={} budget_per_core={} cooldown={}ms",
+            new_hot.me_pool_drain_soft_evict_enabled,
+            new_hot.me_pool_drain_soft_evict_grace_secs,
+            new_hot.me_pool_drain_soft_evict_per_writer,
+            new_hot.me_pool_drain_soft_evict_budget_per_core,
+            new_hot.me_pool_drain_soft_evict_cooldown_ms
+        );
+    }
 
     if (old_hot.me_pool_min_fresh_ratio - new_hot.me_pool_min_fresh_ratio).abs() > f32::EPSILON {
         info!(

src/config/load.rs

@@ -440,6 +440,35 @@ impl ProxyConfig {
             ));
         }
 
+        if config.general.me_pool_drain_soft_evict_grace_secs > 3600 {
+            return Err(ProxyError::Config(
+                "general.me_pool_drain_soft_evict_grace_secs must be within [0, 3600]".to_string(),
+            ));
+        }
+
+        if config.general.me_pool_drain_soft_evict_per_writer == 0
+            || config.general.me_pool_drain_soft_evict_per_writer > 16
+        {
+            return Err(ProxyError::Config(
+                "general.me_pool_drain_soft_evict_per_writer must be within [1, 16]".to_string(),
+            ));
+        }
+
+        if config.general.me_pool_drain_soft_evict_budget_per_core == 0
+            || config.general.me_pool_drain_soft_evict_budget_per_core > 64
+        {
+            return Err(ProxyError::Config(
+                "general.me_pool_drain_soft_evict_budget_per_core must be within [1, 64]"
+                    .to_string(),
+            ));
+        }
+
+        if config.general.me_pool_drain_soft_evict_cooldown_ms == 0 {
+            return Err(ProxyError::Config(
+                "general.me_pool_drain_soft_evict_cooldown_ms must be > 0".to_string(),
+            ));
+        }
+
         if config.access.user_max_unique_ips_window_secs == 0 {
             return Err(ProxyError::Config(
                 "access.user_max_unique_ips_window_secs must be > 0".to_string(),

src/config/types.rs

@@ -803,6 +803,26 @@ pub struct GeneralConfig {
     #[serde(default = "default_me_pool_drain_threshold")]
     pub me_pool_drain_threshold: u64,
 
+    /// Enable staged client eviction for draining ME writers that remain non-empty past TTL.
+    #[serde(default = "default_me_pool_drain_soft_evict_enabled")]
+    pub me_pool_drain_soft_evict_enabled: bool,
+
+    /// Extra grace in seconds after drain TTL before soft-eviction stage starts.
+    #[serde(default = "default_me_pool_drain_soft_evict_grace_secs")]
+    pub me_pool_drain_soft_evict_grace_secs: u64,
+
+    /// Maximum number of client sessions to evict from one draining writer per health tick.
+    #[serde(default = "default_me_pool_drain_soft_evict_per_writer")]
+    pub me_pool_drain_soft_evict_per_writer: u8,
+
+    /// Soft-eviction budget per CPU core for one health tick.
+    #[serde(default = "default_me_pool_drain_soft_evict_budget_per_core")]
+    pub me_pool_drain_soft_evict_budget_per_core: u16,
+
+    /// Cooldown for repetitive soft-eviction on the same writer in milliseconds.
+    #[serde(default = "default_me_pool_drain_soft_evict_cooldown_ms")]
+    pub me_pool_drain_soft_evict_cooldown_ms: u64,
+
     /// Policy for new binds on stale draining writers.
     #[serde(default)]
     pub me_bind_stale_mode: MeBindStaleMode,
@@ -1000,6 +1020,13 @@ impl Default for GeneralConfig {
             proxy_secret_len_max: default_proxy_secret_len_max(),
             me_pool_drain_ttl_secs: default_me_pool_drain_ttl_secs(),
             me_pool_drain_threshold: default_me_pool_drain_threshold(),
+            me_pool_drain_soft_evict_enabled: default_me_pool_drain_soft_evict_enabled(),
+            me_pool_drain_soft_evict_grace_secs: default_me_pool_drain_soft_evict_grace_secs(),
+            me_pool_drain_soft_evict_per_writer: default_me_pool_drain_soft_evict_per_writer(),
+            me_pool_drain_soft_evict_budget_per_core:
+                default_me_pool_drain_soft_evict_budget_per_core(),
+            me_pool_drain_soft_evict_cooldown_ms:
+                default_me_pool_drain_soft_evict_cooldown_ms(),
             me_bind_stale_mode: MeBindStaleMode::default(),
             me_bind_stale_ttl_secs: default_me_bind_stale_ttl_secs(),
             me_pool_min_fresh_ratio: default_me_pool_min_fresh_ratio(),
@@ -1174,9 +1201,17 @@ pub struct ServerConfig {
     #[serde(default = "default_proxy_protocol_header_timeout_ms")]
     pub proxy_protocol_header_timeout_ms: u64,
 
+    /// Port for the Prometheus-compatible metrics endpoint.
+    /// Enables metrics when set; binds on all interfaces (dual-stack) by default.
     #[serde(default)]
     pub metrics_port: Option<u16>,
 
+    /// Listen address for metrics in `IP:PORT` format (e.g. `"127.0.0.1:9090"`).
+    /// When set, takes precedence over `metrics_port` and binds on the specified address only.
+    #[serde(default)]
+    pub metrics_listen: Option<String>,
+
+    /// CIDR whitelist for the metrics endpoint.
     #[serde(default = "default_metrics_whitelist")]
     pub metrics_whitelist: Vec<IpNetwork>,
 
@@ -1204,6 +1239,7 @@ impl Default for ServerConfig {
             proxy_protocol: false,
             proxy_protocol_header_timeout_ms: default_proxy_protocol_header_timeout_ms(),
             metrics_port: None,
+            metrics_listen: None,
             metrics_whitelist: default_metrics_whitelist(),
             api: ApiConfig::default(),
             listeners: Vec::new(),

src/ip_tracker_regression_tests.rs

@@ -0,0 +1,450 @@
+use std::collections::HashMap;
+use std::net::{IpAddr, Ipv4Addr};
+use std::sync::Arc;
+use std::time::Duration;
+
+use crate::config::UserMaxUniqueIpsMode;
+use crate::ip_tracker::UserIpTracker;
+
+fn ip_from_idx(idx: u32) -> IpAddr {
+    let a = 10u8;
+    let b = ((idx / 65_536) % 256) as u8;
+    let c = ((idx / 256) % 256) as u8;
+    let d = (idx % 256) as u8;
+    IpAddr::V4(Ipv4Addr::new(a, b, c, d))
+}
+
+#[tokio::test]
+async fn active_window_enforces_large_unique_ip_burst() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("burst_user", 64).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::ActiveWindow, 30)
+        .await;
+
+    for idx in 0..64 {
+        assert!(tracker.check_and_add("burst_user", ip_from_idx(idx)).await.is_ok());
+    }
+    assert!(tracker.check_and_add("burst_user", ip_from_idx(9_999)).await.is_err());
+    assert_eq!(tracker.get_active_ip_count("burst_user").await, 64);
+}
+
+#[tokio::test]
+async fn global_limit_applies_across_many_users() {
+    let tracker = UserIpTracker::new();
+    tracker.load_limits(3, &HashMap::new()).await;
+
+    for user_idx in 0..150u32 {
+        let user = format!("u{}", user_idx);
+        assert!(tracker.check_and_add(&user, ip_from_idx(user_idx * 10)).await.is_ok());
+        assert!(tracker
+            .check_and_add(&user, ip_from_idx(user_idx * 10 + 1))
+            .await
+            .is_ok());
+        assert!(tracker
+            .check_and_add(&user, ip_from_idx(user_idx * 10 + 2))
+            .await
+            .is_ok());
+        assert!(tracker
+            .check_and_add(&user, ip_from_idx(user_idx * 10 + 3))
+            .await
+            .is_err());
+    }
+
+    assert_eq!(tracker.get_stats().await.len(), 150);
+}
+
+#[tokio::test]
+async fn user_zero_override_falls_back_to_global_limit() {
+    let tracker = UserIpTracker::new();
+    let mut limits = HashMap::new();
+    limits.insert("target".to_string(), 0);
+    tracker.load_limits(2, &limits).await;
+
+    assert!(tracker.check_and_add("target", ip_from_idx(1)).await.is_ok());
+    assert!(tracker.check_and_add("target", ip_from_idx(2)).await.is_ok());
+    assert!(tracker.check_and_add("target", ip_from_idx(3)).await.is_err());
+    assert_eq!(tracker.get_user_limit("target").await, Some(2));
+}
+
+#[tokio::test]
+async fn remove_ip_is_idempotent_after_counter_reaches_zero() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("u", 2).await;
+    let ip = ip_from_idx(42);
+
+    tracker.check_and_add("u", ip).await.unwrap();
+    tracker.remove_ip("u", ip).await;
+    tracker.remove_ip("u", ip).await;
+    tracker.remove_ip("u", ip).await;
+
+    assert_eq!(tracker.get_active_ip_count("u").await, 0);
+    assert!(!tracker.is_ip_active("u", ip).await);
+}
+
+#[tokio::test]
+async fn clear_user_ips_resets_active_and_recent() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("u", 10).await;
+
+    for idx in 0..6 {
+        tracker.check_and_add("u", ip_from_idx(idx)).await.unwrap();
+    }
+
+    tracker.clear_user_ips("u").await;
+
+    assert_eq!(tracker.get_active_ip_count("u").await, 0);
+    let counts = tracker
+        .get_recent_counts_for_users(&["u".to_string()])
+        .await;
+    assert_eq!(counts.get("u").copied().unwrap_or(0), 0);
+}
+
+#[tokio::test]
+async fn clear_all_resets_multi_user_state() {
+    let tracker = UserIpTracker::new();
+
+    for user_idx in 0..80u32 {
+        let user = format!("u{}", user_idx);
+        for ip_idx in 0..3 {
+            tracker
+                .check_and_add(&user, ip_from_idx(user_idx * 100 + ip_idx))
+                .await
+                .unwrap();
+        }
+    }
+
+    tracker.clear_all().await;
+
+    assert!(tracker.get_stats().await.is_empty());
+    let users = (0..80u32)
+        .map(|idx| format!("u{}", idx))
+        .collect::<Vec<_>>();
+    let recent = tracker.get_recent_counts_for_users(&users).await;
+    assert!(recent.values().all(|count| *count == 0));
+}
+
+#[tokio::test]
+async fn get_active_ips_for_users_are_sorted() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("user", 10).await;
+
+    tracker
+        .check_and_add("user", IpAddr::V4(Ipv4Addr::new(10, 0, 0, 9)))
+        .await
+        .unwrap();
+    tracker
+        .check_and_add("user", IpAddr::V4(Ipv4Addr::new(10, 0, 0, 1)))
+        .await
+        .unwrap();
+    tracker
+        .check_and_add("user", IpAddr::V4(Ipv4Addr::new(10, 0, 0, 5)))
+        .await
+        .unwrap();
+
+    let map = tracker
+        .get_active_ips_for_users(&["user".to_string()])
+        .await;
+    let ips = map.get("user").cloned().unwrap_or_default();
+
+    assert_eq!(
+        ips,
+        vec![
+            IpAddr::V4(Ipv4Addr::new(10, 0, 0, 1)),
+            IpAddr::V4(Ipv4Addr::new(10, 0, 0, 5)),
+            IpAddr::V4(Ipv4Addr::new(10, 0, 0, 9)),
+        ]
+    );
+}
+
+#[tokio::test]
+async fn get_recent_ips_for_users_are_sorted() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("user", 10).await;
+
+    tracker
+        .check_and_add("user", IpAddr::V4(Ipv4Addr::new(10, 1, 0, 9)))
+        .await
+        .unwrap();
+    tracker
+        .check_and_add("user", IpAddr::V4(Ipv4Addr::new(10, 1, 0, 1)))
+        .await
+        .unwrap();
+    tracker
+        .check_and_add("user", IpAddr::V4(Ipv4Addr::new(10, 1, 0, 5)))
+        .await
+        .unwrap();
+
+    let map = tracker
+        .get_recent_ips_for_users(&["user".to_string()])
+        .await;
+    let ips = map.get("user").cloned().unwrap_or_default();
+
+    assert_eq!(
+        ips,
+        vec![
+            IpAddr::V4(Ipv4Addr::new(10, 1, 0, 1)),
+            IpAddr::V4(Ipv4Addr::new(10, 1, 0, 5)),
+            IpAddr::V4(Ipv4Addr::new(10, 1, 0, 9)),
+        ]
+    );
+}
+
+#[tokio::test]
+async fn time_window_expires_for_large_rotation() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("tw", 1).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::TimeWindow, 1)
+        .await;
+
+    tracker.check_and_add("tw", ip_from_idx(1)).await.unwrap();
+    tracker.remove_ip("tw", ip_from_idx(1)).await;
+    assert!(tracker.check_and_add("tw", ip_from_idx(2)).await.is_err());
+
+    tokio::time::sleep(Duration::from_millis(1_100)).await;
+    assert!(tracker.check_and_add("tw", ip_from_idx(2)).await.is_ok());
+}
+
+#[tokio::test]
+async fn combined_mode_blocks_recent_after_disconnect() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("cmb", 1).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::Combined, 2)
+        .await;
+
+    tracker.check_and_add("cmb", ip_from_idx(11)).await.unwrap();
+    tracker.remove_ip("cmb", ip_from_idx(11)).await;
+
+    assert!(tracker.check_and_add("cmb", ip_from_idx(12)).await.is_err());
+}
+
+#[tokio::test]
+async fn load_limits_replaces_large_limit_map() {
+    let tracker = UserIpTracker::new();
+    let mut first = HashMap::new();
+    let mut second = HashMap::new();
+
+    for idx in 0..300usize {
+        first.insert(format!("u{}", idx), 2usize);
+    }
+    for idx in 150..450usize {
+        second.insert(format!("u{}", idx), 4usize);
+    }
+
+    tracker.load_limits(0, &first).await;
+    tracker.load_limits(0, &second).await;
+
+    assert_eq!(tracker.get_user_limit("u20").await, None);
+    assert_eq!(tracker.get_user_limit("u200").await, Some(4));
+    assert_eq!(tracker.get_user_limit("u420").await, Some(4));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn concurrent_same_user_unique_ip_pressure_stays_bounded() {
+    let tracker = Arc::new(UserIpTracker::new());
+    tracker.set_user_limit("hot", 32).await;
+    tracker
+        .set_limit_policy(UserMaxUniqueIpsMode::ActiveWindow, 30)
+        .await;
+
+    let mut handles = Vec::new();
+    for worker in 0..16u32 {
+        let tracker_cloned = tracker.clone();
+        handles.push(tokio::spawn(async move {
+            let base = worker * 200;
+            for step in 0..200u32 {
+                let _ = tracker_cloned
+                    .check_and_add("hot", ip_from_idx(base + step))
+                    .await;
+            }
+        }));
+    }
+
+    for handle in handles {
+        handle.await.unwrap();
+    }
+
+    assert!(tracker.get_active_ip_count("hot").await <= 32);
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn concurrent_many_users_isolate_limits() {
+    let tracker = Arc::new(UserIpTracker::new());
+    tracker.load_limits(4, &HashMap::new()).await;
+
+    let mut handles = Vec::new();
+    for user_idx in 0..120u32 {
+        let tracker_cloned = tracker.clone();
+        handles.push(tokio::spawn(async move {
+            let user = format!("u{}", user_idx);
+            for ip_idx in 0..10u32 {
+                let _ = tracker_cloned
+                    .check_and_add(&user, ip_from_idx(user_idx * 1_000 + ip_idx))
+                    .await;
+            }
+        }));
+    }
+
+    for handle in handles {
+        handle.await.unwrap();
+    }
+
+    let stats = tracker.get_stats().await;
+    assert_eq!(stats.len(), 120);
+    assert!(stats.iter().all(|(_, active, limit)| *active <= 4 && *limit == 4));
+}
+
+#[tokio::test]
+async fn same_ip_reconnect_high_frequency_keeps_single_unique() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("same", 2).await;
+    let ip = ip_from_idx(9);
+
+    for _ in 0..2_000 {
+        tracker.check_and_add("same", ip).await.unwrap();
+    }
+
+    assert_eq!(tracker.get_active_ip_count("same").await, 1);
+    assert!(tracker.is_ip_active("same", ip).await);
+}
+
+#[tokio::test]
+async fn format_stats_contains_expected_limited_and_unlimited_markers() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("limited", 2).await;
+    tracker.check_and_add("limited", ip_from_idx(1)).await.unwrap();
+    tracker.check_and_add("open", ip_from_idx(2)).await.unwrap();
+
+    let text = tracker.format_stats().await;
+
+    assert!(text.contains("limited"));
+    assert!(text.contains("open"));
+    assert!(text.contains("unlimited"));
+}
+
+#[tokio::test]
+async fn stats_report_global_default_for_users_without_override() {
+    let tracker = UserIpTracker::new();
+    tracker.load_limits(5, &HashMap::new()).await;
+
+    tracker.check_and_add("a", ip_from_idx(1)).await.unwrap();
+    tracker.check_and_add("b", ip_from_idx(2)).await.unwrap();
+
+    let stats = tracker.get_stats().await;
+    assert!(stats.iter().any(|(user, _, limit)| user == "a" && *limit == 5));
+    assert!(stats.iter().any(|(user, _, limit)| user == "b" && *limit == 5));
+}
+
+#[tokio::test]
+async fn stress_cycle_add_remove_clear_preserves_empty_end_state() {
+    let tracker = UserIpTracker::new();
+
+    for cycle in 0..50u32 {
+        let user = format!("cycle{}", cycle);
+        tracker.set_user_limit(&user, 128).await;
+
+        for ip_idx in 0..128u32 {
+            tracker
+                .check_and_add(&user, ip_from_idx(cycle * 10_000 + ip_idx))
+                .await
+                .unwrap();
+        }
+
+        for ip_idx in 0..128u32 {
+            tracker
+                .remove_ip(&user, ip_from_idx(cycle * 10_000 + ip_idx))
+                .await;
+        }
+
+        tracker.clear_user_ips(&user).await;
+    }
+
+    assert!(tracker.get_stats().await.is_empty());
+}
+
+#[tokio::test]
+async fn remove_unknown_user_or_ip_does_not_corrupt_state() {
+    let tracker = UserIpTracker::new();
+
+    tracker.remove_ip("no_user", ip_from_idx(1)).await;
+    tracker.check_and_add("x", ip_from_idx(2)).await.unwrap();
+    tracker.remove_ip("x", ip_from_idx(3)).await;
+
+    assert_eq!(tracker.get_active_ip_count("x").await, 1);
+    assert!(tracker.is_ip_active("x", ip_from_idx(2)).await);
+}
+
+#[tokio::test]
+async fn active_and_recent_views_match_after_mixed_workload() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("mix", 16).await;
+
+    for ip_idx in 0..12u32 {
+        tracker.check_and_add("mix", ip_from_idx(ip_idx)).await.unwrap();
+    }
+    for ip_idx in 0..6u32 {
+        tracker.remove_ip("mix", ip_from_idx(ip_idx)).await;
+    }
+
+    let active = tracker
+        .get_active_ips_for_users(&["mix".to_string()])
+        .await
+        .get("mix")
+        .cloned()
+        .unwrap_or_default();
+    let recent_count = tracker
+        .get_recent_counts_for_users(&["mix".to_string()])
+        .await
+        .get("mix")
+        .copied()
+        .unwrap_or(0);
+
+    assert_eq!(active.len(), 6);
+    assert!(recent_count >= active.len());
+    assert!(recent_count <= 12);
+}
+
+#[tokio::test]
+async fn global_limit_switch_updates_enforcement_immediately() {
+    let tracker = UserIpTracker::new();
+    tracker.load_limits(2, &HashMap::new()).await;
+
+    assert!(tracker.check_and_add("u", ip_from_idx(1)).await.is_ok());
+    assert!(tracker.check_and_add("u", ip_from_idx(2)).await.is_ok());
+    assert!(tracker.check_and_add("u", ip_from_idx(3)).await.is_err());
+
+    tracker.clear_user_ips("u").await;
+    tracker.load_limits(4, &HashMap::new()).await;
+
+    assert!(tracker.check_and_add("u", ip_from_idx(1)).await.is_ok());
+    assert!(tracker.check_and_add("u", ip_from_idx(2)).await.is_ok());
+    assert!(tracker.check_and_add("u", ip_from_idx(3)).await.is_ok());
+    assert!(tracker.check_and_add("u", ip_from_idx(4)).await.is_ok());
+    assert!(tracker.check_and_add("u", ip_from_idx(5)).await.is_err());
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn concurrent_reconnect_and_disconnect_preserves_non_negative_counts() {
+    let tracker = Arc::new(UserIpTracker::new());
+    tracker.set_user_limit("cc", 8).await;
+
+    let mut handles = Vec::new();
+    for worker in 0..8u32 {
+        let tracker_cloned = tracker.clone();
+        handles.push(tokio::spawn(async move {
+            let ip = ip_from_idx(50 + worker);
+            for _ in 0..500u32 {
+                let _ = tracker_cloned.check_and_add("cc", ip).await;
+                tracker_cloned.remove_ip("cc", ip).await;
+            }
+        }));
+    }
+
+    for handle in handles {
+        handle.await.unwrap();
+    }
+
+    assert!(tracker.get_active_ip_count("cc").await <= 8);
+}

src/maestro/me_startup.rs

@@ -238,6 +238,11 @@ pub(crate) async fn initialize_me_pool(
                     config.general.hardswap,
                     config.general.me_pool_drain_ttl_secs,
                     config.general.me_pool_drain_threshold,
+                    config.general.me_pool_drain_soft_evict_enabled,
+                    config.general.me_pool_drain_soft_evict_grace_secs,
+                    config.general.me_pool_drain_soft_evict_per_writer,
+                    config.general.me_pool_drain_soft_evict_budget_per_core,
+                    config.general.me_pool_drain_soft_evict_cooldown_ms,
                     config.general.effective_me_pool_force_close_secs(),
                     config.general.me_pool_min_fresh_ratio,
                     config.general.me_hardswap_warmup_delay_min_ms,

src/maestro/mod.rs

@@ -476,7 +476,7 @@ pub async fn run() -> std::result::Result<(), Box<dyn std::error::Error>> {
         Duration::from_secs(config.access.replay_window_secs),
     ));
 
-    let buffer_pool = Arc::new(BufferPool::with_config(16 * 1024, 4096));
+    let buffer_pool = Arc::new(BufferPool::with_config(64 * 1024, 4096));
 
     connectivity::run_startup_connectivity(
         &config,

src/maestro/runtime_tasks.rs

@@ -279,11 +279,32 @@ pub(crate) async fn spawn_metrics_if_configured(
     ip_tracker: Arc<UserIpTracker>,
     config_rx: watch::Receiver<Arc<ProxyConfig>>,
 ) {
-    if let Some(port) = config.server.metrics_port {
+    // metrics_listen takes precedence; fall back to metrics_port for backward compat.
+    let metrics_target: Option<(u16, Option<String>)> =
+        if let Some(ref listen) = config.server.metrics_listen {
+            match listen.parse::<std::net::SocketAddr>() {
+                Ok(addr) => Some((addr.port(), Some(listen.clone()))),
+                Err(e) => {
+                    startup_tracker
+                        .skip_component(
+                            COMPONENT_METRICS_START,
+                            Some(format!("invalid metrics_listen \"{}\": {}", listen, e)),
+                        )
+                        .await;
+                    None
+                }
+            }
+        } else {
+            config.server.metrics_port.map(|p| (p, None))
+        };
+
+    if let Some((port, listen)) = metrics_target {
+        let fallback_label = format!("port {}", port);
+        let label = listen.as_deref().unwrap_or(&fallback_label);
         startup_tracker
             .start_component(
                 COMPONENT_METRICS_START,
-                Some(format!("spawn metrics endpoint on {}", port)),
+                Some(format!("spawn metrics endpoint on {}", label)),
             )
             .await;
         let stats = stats.clone();
@@ -294,6 +315,7 @@ pub(crate) async fn spawn_metrics_if_configured(
         tokio::spawn(async move {
             metrics::serve(
                 port,
+                listen,
                 stats,
                 beobachten,
                 ip_tracker_metrics,
@@ -308,7 +330,7 @@ pub(crate) async fn spawn_metrics_if_configured(
                 Some("metrics task spawned".to_string()),
             )
             .await;
-    } else {
+    } else if config.server.metrics_listen.is_none() {
         startup_tracker
             .skip_component(
                 COMPONENT_METRICS_START,

src/main.rs

@@ -6,6 +6,8 @@ mod config;
 mod crypto;
 mod error;
 mod ip_tracker;
+#[cfg(test)]
+mod ip_tracker_regression_tests;
 mod maestro;
 mod metrics;
 mod network;

src/metrics.rs

@@ -21,6 +21,7 @@ use crate::transport::{ListenOptions, create_listener};
 
 pub async fn serve(
     port: u16,
+    listen: Option<String>,
     stats: Arc<Stats>,
     beobachten: Arc<BeobachtenStore>,
     ip_tracker: Arc<UserIpTracker>,
@@ -28,6 +29,33 @@ pub async fn serve(
     whitelist: Vec<IpNetwork>,
 ) {
     let whitelist = Arc::new(whitelist);
+
+    // If `metrics_listen` is set, bind on that single address only.
+    if let Some(ref listen_addr) = listen {
+        let addr: SocketAddr = match listen_addr.parse() {
+            Ok(a) => a,
+            Err(e) => {
+                warn!(error = %e, "Invalid metrics_listen address: {}", listen_addr);
+                return;
+            }
+        };
+        let is_ipv6 = addr.is_ipv6();
+        match bind_metrics_listener(addr, is_ipv6) {
+            Ok(listener) => {
+                info!("Metrics endpoint: http://{}/metrics and /beobachten", addr);
+                serve_listener(
+                    listener, stats, beobachten, ip_tracker, config_rx, whitelist,
+                )
+                .await;
+            }
+            Err(e) => {
+                warn!(error = %e, "Failed to bind metrics on {}", addr);
+            }
+        }
+        return;
+    }
+
+    // Fallback: bind on 0.0.0.0 and [::] using metrics_port.
     let mut listener_v4 = None;
     let mut listener_v6 = None;
 
@@ -264,6 +292,109 @@ async fn render_metrics(stats: &Stats, config: &ProxyConfig, ip_tracker: &UserIp
         "telemt_connections_bad_total {}",
         if core_enabled { stats.get_connects_bad() } else { 0 }
     );
+    let _ = writeln!(out, "# HELP telemt_connections_current Current active connections");
+    let _ = writeln!(out, "# TYPE telemt_connections_current gauge");
+    let _ = writeln!(
+        out,
+        "telemt_connections_current {}",
+        if core_enabled {
+            stats.get_current_connections_total()
+        } else {
+            0
+        }
+    );
+    let _ = writeln!(out, "# HELP telemt_connections_direct_current Current active direct connections");
+    let _ = writeln!(out, "# TYPE telemt_connections_direct_current gauge");
+    let _ = writeln!(
+        out,
+        "telemt_connections_direct_current {}",
+        if core_enabled {
+            stats.get_current_connections_direct()
+        } else {
+            0
+        }
+    );
+    let _ = writeln!(out, "# HELP telemt_connections_me_current Current active middle-end connections");
+    let _ = writeln!(out, "# TYPE telemt_connections_me_current gauge");
+    let _ = writeln!(
+        out,
+        "telemt_connections_me_current {}",
+        if core_enabled {
+            stats.get_current_connections_me()
+        } else {
+            0
+        }
+    );
+    let _ = writeln!(
+        out,
+        "# HELP telemt_relay_adaptive_promotions_total Adaptive relay tier promotions"
+    );
+    let _ = writeln!(out, "# TYPE telemt_relay_adaptive_promotions_total counter");
+    let _ = writeln!(
+        out,
+        "telemt_relay_adaptive_promotions_total {}",
+        if core_enabled {
+            stats.get_relay_adaptive_promotions_total()
+        } else {
+            0
+        }
+    );
+    let _ = writeln!(
+        out,
+        "# HELP telemt_relay_adaptive_demotions_total Adaptive relay tier demotions"
+    );
+    let _ = writeln!(out, "# TYPE telemt_relay_adaptive_demotions_total counter");
+    let _ = writeln!(
+        out,
+        "telemt_relay_adaptive_demotions_total {}",
+        if core_enabled {
+            stats.get_relay_adaptive_demotions_total()
+        } else {
+            0
+        }
+    );
+    let _ = writeln!(
+        out,
+        "# HELP telemt_relay_adaptive_hard_promotions_total Adaptive relay hard promotions triggered by write pressure"
+    );
+    let _ = writeln!(
+        out,
+        "# TYPE telemt_relay_adaptive_hard_promotions_total counter"
+    );
+    let _ = writeln!(
+        out,
+        "telemt_relay_adaptive_hard_promotions_total {}",
+        if core_enabled {
+            stats.get_relay_adaptive_hard_promotions_total()
+        } else {
+            0
+        }
+    );
+    let _ = writeln!(out, "# HELP telemt_reconnect_evict_total Reconnect-driven session evictions");
+    let _ = writeln!(out, "# TYPE telemt_reconnect_evict_total counter");
+    let _ = writeln!(
+        out,
+        "telemt_reconnect_evict_total {}",
+        if core_enabled {
+            stats.get_reconnect_evict_total()
+        } else {
+            0
+        }
+    );
+    let _ = writeln!(
+        out,
+        "# HELP telemt_reconnect_stale_close_total Sessions closed because they became stale after reconnect"
+    );
+    let _ = writeln!(out, "# TYPE telemt_reconnect_stale_close_total counter");
+    let _ = writeln!(
+        out,
+        "telemt_reconnect_stale_close_total {}",
+        if core_enabled {
+            stats.get_reconnect_stale_close_total()
+        } else {
+            0
+        }
+    );
 
     let _ = writeln!(out, "# HELP telemt_handshake_timeouts_total Handshake timeouts");
     let _ = writeln!(out, "# TYPE telemt_handshake_timeouts_total counter");
@@ -1519,6 +1650,36 @@ async fn render_metrics(stats: &Stats, config: &ProxyConfig, ip_tracker: &UserIp
         }
     );
 
+    let _ = writeln!(
+        out,
+        "# HELP telemt_pool_drain_soft_evict_total Soft-evicted client sessions on stuck draining writers"
+    );
+    let _ = writeln!(out, "# TYPE telemt_pool_drain_soft_evict_total counter");
+    let _ = writeln!(
+        out,
+        "telemt_pool_drain_soft_evict_total {}",
+        if me_allows_normal {
+            stats.get_pool_drain_soft_evict_total()
+        } else {
+            0
+        }
+    );
+
+    let _ = writeln!(
+        out,
+        "# HELP telemt_pool_drain_soft_evict_writer_total Draining writers with at least one soft eviction"
+    );
+    let _ = writeln!(out, "# TYPE telemt_pool_drain_soft_evict_writer_total counter");
+    let _ = writeln!(
+        out,
+        "telemt_pool_drain_soft_evict_writer_total {}",
+        if me_allows_normal {
+            stats.get_pool_drain_soft_evict_writer_total()
+        } else {
+            0
+        }
+    );
+
     let _ = writeln!(out, "# HELP telemt_pool_stale_pick_total Stale writer fallback picks for new binds");
     let _ = writeln!(out, "# TYPE telemt_pool_stale_pick_total counter");
     let _ = writeln!(
@@ -1836,6 +1997,8 @@ mod tests {
         stats.increment_connects_all();
         stats.increment_connects_all();
         stats.increment_connects_bad();
+        stats.increment_current_connections_direct();
+        stats.increment_current_connections_me();
         stats.increment_handshake_timeouts();
         stats.increment_upstream_connect_attempt_total();
         stats.increment_upstream_connect_attempt_total();
@@ -1867,6 +2030,9 @@ mod tests {
 
         assert!(output.contains("telemt_connections_total 2"));
         assert!(output.contains("telemt_connections_bad_total 1"));
+        assert!(output.contains("telemt_connections_current 2"));
+        assert!(output.contains("telemt_connections_direct_current 1"));
+        assert!(output.contains("telemt_connections_me_current 1"));
         assert!(output.contains("telemt_handshake_timeouts_total 1"));
         assert!(output.contains("telemt_upstream_connect_attempt_total 2"));
         assert!(output.contains("telemt_upstream_connect_success_total 1"));
@@ -1909,6 +2075,9 @@ mod tests {
         let output = render_metrics(&stats, &config, &tracker).await;
         assert!(output.contains("telemt_connections_total 0"));
         assert!(output.contains("telemt_connections_bad_total 0"));
+        assert!(output.contains("telemt_connections_current 0"));
+        assert!(output.contains("telemt_connections_direct_current 0"));
+        assert!(output.contains("telemt_connections_me_current 0"));
         assert!(output.contains("telemt_handshake_timeouts_total 0"));
         assert!(output.contains("telemt_user_unique_ips_current{user="));
         assert!(output.contains("telemt_user_unique_ips_recent_window{user="));
@@ -1942,11 +2111,21 @@ mod tests {
         assert!(output.contains("# TYPE telemt_uptime_seconds gauge"));
         assert!(output.contains("# TYPE telemt_connections_total counter"));
         assert!(output.contains("# TYPE telemt_connections_bad_total counter"));
+        assert!(output.contains("# TYPE telemt_connections_current gauge"));
+        assert!(output.contains("# TYPE telemt_connections_direct_current gauge"));
+        assert!(output.contains("# TYPE telemt_connections_me_current gauge"));
+        assert!(output.contains("# TYPE telemt_relay_adaptive_promotions_total counter"));
+        assert!(output.contains("# TYPE telemt_relay_adaptive_demotions_total counter"));
+        assert!(output.contains("# TYPE telemt_relay_adaptive_hard_promotions_total counter"));
+        assert!(output.contains("# TYPE telemt_reconnect_evict_total counter"));
+        assert!(output.contains("# TYPE telemt_reconnect_stale_close_total counter"));
         assert!(output.contains("# TYPE telemt_handshake_timeouts_total counter"));
         assert!(output.contains("# TYPE telemt_upstream_connect_attempt_total counter"));
         assert!(output.contains("# TYPE telemt_me_rpc_proxy_req_signal_sent_total counter"));
         assert!(output.contains("# TYPE telemt_me_idle_close_by_peer_total counter"));
         assert!(output.contains("# TYPE telemt_me_writer_removed_total counter"));
+        assert!(output.contains("# TYPE telemt_pool_drain_soft_evict_total counter"));
+        assert!(output.contains("# TYPE telemt_pool_drain_soft_evict_writer_total counter"));
         assert!(output.contains(
             "# TYPE telemt_me_writer_removed_unexpected_minus_restored_total gauge"
         ));

src/proxy/adaptive_buffers.rs

@@ -0,0 +1,383 @@
+use dashmap::DashMap;
+use std::cmp::max;
+use std::sync::OnceLock;
+use std::time::{Duration, Instant};
+
+const EMA_ALPHA: f64 = 0.2;
+const PROFILE_TTL: Duration = Duration::from_secs(300);
+const THROUGHPUT_UP_BPS: f64 = 8_000_000.0;
+const THROUGHPUT_DOWN_BPS: f64 = 2_000_000.0;
+const RATIO_CONFIRM_THRESHOLD: f64 = 1.12;
+const TIER1_HOLD_TICKS: u32 = 8;
+const TIER2_HOLD_TICKS: u32 = 4;
+const QUIET_DEMOTE_TICKS: u32 = 480;
+const HARD_COOLDOWN_TICKS: u32 = 20;
+const HARD_PENDING_THRESHOLD: u32 = 3;
+const HARD_PARTIAL_RATIO_THRESHOLD: f64 = 0.25;
+const DIRECT_C2S_CAP_BYTES: usize = 128 * 1024;
+const DIRECT_S2C_CAP_BYTES: usize = 512 * 1024;
+const ME_FRAMES_CAP: usize = 96;
+const ME_BYTES_CAP: usize = 384 * 1024;
+const ME_DELAY_MIN_US: u64 = 150;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
+pub enum AdaptiveTier {
+    Base = 0,
+    Tier1 = 1,
+    Tier2 = 2,
+    Tier3 = 3,
+}
+
+impl AdaptiveTier {
+    pub fn promote(self) -> Self {
+        match self {
+            Self::Base => Self::Tier1,
+            Self::Tier1 => Self::Tier2,
+            Self::Tier2 => Self::Tier3,
+            Self::Tier3 => Self::Tier3,
+        }
+    }
+
+    pub fn demote(self) -> Self {
+        match self {
+            Self::Base => Self::Base,
+            Self::Tier1 => Self::Base,
+            Self::Tier2 => Self::Tier1,
+            Self::Tier3 => Self::Tier2,
+        }
+    }
+
+    fn ratio(self) -> (usize, usize) {
+        match self {
+            Self::Base => (1, 1),
+            Self::Tier1 => (5, 4),
+            Self::Tier2 => (3, 2),
+            Self::Tier3 => (2, 1),
+        }
+    }
+
+    pub fn as_u8(self) -> u8 {
+        self as u8
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum TierTransitionReason {
+    SoftConfirmed,
+    HardPressure,
+    QuietDemotion,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct TierTransition {
+    pub from: AdaptiveTier,
+    pub to: AdaptiveTier,
+    pub reason: TierTransitionReason,
+}
+
+#[derive(Debug, Clone, Copy, Default)]
+pub struct RelaySignalSample {
+    pub c2s_bytes: u64,
+    pub s2c_requested_bytes: u64,
+    pub s2c_written_bytes: u64,
+    pub s2c_write_ops: u64,
+    pub s2c_partial_writes: u64,
+    pub s2c_consecutive_pending_writes: u32,
+}
+
+#[derive(Debug, Clone, Copy)]
+pub struct SessionAdaptiveController {
+    tier: AdaptiveTier,
+    max_tier_seen: AdaptiveTier,
+    throughput_ema_bps: f64,
+    incoming_ema_bps: f64,
+    outgoing_ema_bps: f64,
+    tier1_hold_ticks: u32,
+    tier2_hold_ticks: u32,
+    quiet_ticks: u32,
+    hard_cooldown_ticks: u32,
+}
+
+impl SessionAdaptiveController {
+    pub fn new(initial_tier: AdaptiveTier) -> Self {
+        Self {
+            tier: initial_tier,
+            max_tier_seen: initial_tier,
+            throughput_ema_bps: 0.0,
+            incoming_ema_bps: 0.0,
+            outgoing_ema_bps: 0.0,
+            tier1_hold_ticks: 0,
+            tier2_hold_ticks: 0,
+            quiet_ticks: 0,
+            hard_cooldown_ticks: 0,
+        }
+    }
+
+    pub fn max_tier_seen(&self) -> AdaptiveTier {
+        self.max_tier_seen
+    }
+
+    pub fn observe(&mut self, sample: RelaySignalSample, tick_secs: f64) -> Option<TierTransition> {
+        if tick_secs <= f64::EPSILON {
+            return None;
+        }
+
+        if self.hard_cooldown_ticks > 0 {
+            self.hard_cooldown_ticks -= 1;
+        }
+
+        let c2s_bps = (sample.c2s_bytes as f64 * 8.0) / tick_secs;
+        let incoming_bps = (sample.s2c_requested_bytes as f64 * 8.0) / tick_secs;
+        let outgoing_bps = (sample.s2c_written_bytes as f64 * 8.0) / tick_secs;
+        let throughput = c2s_bps.max(outgoing_bps);
+
+        self.throughput_ema_bps = ema(self.throughput_ema_bps, throughput);
+        self.incoming_ema_bps = ema(self.incoming_ema_bps, incoming_bps);
+        self.outgoing_ema_bps = ema(self.outgoing_ema_bps, outgoing_bps);
+
+        let tier1_now = self.throughput_ema_bps >= THROUGHPUT_UP_BPS;
+        if tier1_now {
+            self.tier1_hold_ticks = self.tier1_hold_ticks.saturating_add(1);
+        } else {
+            self.tier1_hold_ticks = 0;
+        }
+
+        let ratio = if self.outgoing_ema_bps <= f64::EPSILON {
+            0.0
+        } else {
+            self.incoming_ema_bps / self.outgoing_ema_bps
+        };
+        let tier2_now = ratio >= RATIO_CONFIRM_THRESHOLD;
+        if tier2_now {
+            self.tier2_hold_ticks = self.tier2_hold_ticks.saturating_add(1);
+        } else {
+            self.tier2_hold_ticks = 0;
+        }
+
+        let partial_ratio = if sample.s2c_write_ops == 0 {
+            0.0
+        } else {
+            sample.s2c_partial_writes as f64 / sample.s2c_write_ops as f64
+        };
+        let hard_now = sample.s2c_consecutive_pending_writes >= HARD_PENDING_THRESHOLD
+            || partial_ratio >= HARD_PARTIAL_RATIO_THRESHOLD;
+
+        if hard_now && self.hard_cooldown_ticks == 0 {
+            return self.promote(TierTransitionReason::HardPressure, HARD_COOLDOWN_TICKS);
+        }
+
+        if self.tier1_hold_ticks >= TIER1_HOLD_TICKS && self.tier2_hold_ticks >= TIER2_HOLD_TICKS {
+            return self.promote(TierTransitionReason::SoftConfirmed, 0);
+        }
+
+        let demote_candidate = self.throughput_ema_bps < THROUGHPUT_DOWN_BPS && !tier2_now && !hard_now;
+        if demote_candidate {
+            self.quiet_ticks = self.quiet_ticks.saturating_add(1);
+            if self.quiet_ticks >= QUIET_DEMOTE_TICKS {
+                self.quiet_ticks = 0;
+                return self.demote(TierTransitionReason::QuietDemotion);
+            }
+        } else {
+            self.quiet_ticks = 0;
+        }
+
+        None
+    }
+
+    fn promote(
+        &mut self,
+        reason: TierTransitionReason,
+        hard_cooldown_ticks: u32,
+    ) -> Option<TierTransition> {
+        let from = self.tier;
+        let to = from.promote();
+        if from == to {
+            return None;
+        }
+        self.tier = to;
+        self.max_tier_seen = max(self.max_tier_seen, to);
+        self.hard_cooldown_ticks = hard_cooldown_ticks;
+        self.tier1_hold_ticks = 0;
+        self.tier2_hold_ticks = 0;
+        self.quiet_ticks = 0;
+        Some(TierTransition { from, to, reason })
+    }
+
+    fn demote(&mut self, reason: TierTransitionReason) -> Option<TierTransition> {
+        let from = self.tier;
+        let to = from.demote();
+        if from == to {
+            return None;
+        }
+        self.tier = to;
+        self.tier1_hold_ticks = 0;
+        self.tier2_hold_ticks = 0;
+        Some(TierTransition { from, to, reason })
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+struct UserAdaptiveProfile {
+    tier: AdaptiveTier,
+    seen_at: Instant,
+}
+
+fn profiles() -> &'static DashMap<String, UserAdaptiveProfile> {
+    static USER_PROFILES: OnceLock<DashMap<String, UserAdaptiveProfile>> = OnceLock::new();
+    USER_PROFILES.get_or_init(DashMap::new)
+}
+
+pub fn seed_tier_for_user(user: &str) -> AdaptiveTier {
+    let now = Instant::now();
+    if let Some(entry) = profiles().get(user) {
+        let value = entry.value();
+        if now.duration_since(value.seen_at) <= PROFILE_TTL {
+            return value.tier;
+        }
+    }
+    AdaptiveTier::Base
+}
+
+pub fn record_user_tier(user: &str, tier: AdaptiveTier) {
+    let now = Instant::now();
+    if let Some(mut entry) = profiles().get_mut(user) {
+        let existing = *entry;
+        let effective = if now.duration_since(existing.seen_at) > PROFILE_TTL {
+            tier
+        } else {
+            max(existing.tier, tier)
+        };
+        *entry = UserAdaptiveProfile {
+            tier: effective,
+            seen_at: now,
+        };
+        return;
+    }
+    profiles().insert(
+        user.to_string(),
+        UserAdaptiveProfile { tier, seen_at: now },
+    );
+}
+
+pub fn direct_copy_buffers_for_tier(
+    tier: AdaptiveTier,
+    base_c2s: usize,
+    base_s2c: usize,
+) -> (usize, usize) {
+    let (num, den) = tier.ratio();
+    (
+        scale(base_c2s, num, den, DIRECT_C2S_CAP_BYTES),
+        scale(base_s2c, num, den, DIRECT_S2C_CAP_BYTES),
+    )
+}
+
+pub fn me_flush_policy_for_tier(
+    tier: AdaptiveTier,
+    base_frames: usize,
+    base_bytes: usize,
+    base_delay: Duration,
+) -> (usize, usize, Duration) {
+    let (num, den) = tier.ratio();
+    let frames = scale(base_frames, num, den, ME_FRAMES_CAP).max(1);
+    let bytes = scale(base_bytes, num, den, ME_BYTES_CAP).max(4096);
+    let delay_us = base_delay.as_micros() as u64;
+    let adjusted_delay_us = match tier {
+        AdaptiveTier::Base => delay_us,
+        AdaptiveTier::Tier1 => (delay_us.saturating_mul(7)).saturating_div(10),
+        AdaptiveTier::Tier2 => delay_us.saturating_div(2),
+        AdaptiveTier::Tier3 => (delay_us.saturating_mul(3)).saturating_div(10),
+    }
+    .max(ME_DELAY_MIN_US)
+    .min(delay_us.max(ME_DELAY_MIN_US));
+    (frames, bytes, Duration::from_micros(adjusted_delay_us))
+}
+
+fn ema(prev: f64, value: f64) -> f64 {
+    if prev <= f64::EPSILON {
+        value
+    } else {
+        (prev * (1.0 - EMA_ALPHA)) + (value * EMA_ALPHA)
+    }
+}
+
+fn scale(base: usize, numerator: usize, denominator: usize, cap: usize) -> usize {
+    let scaled = base
+        .saturating_mul(numerator)
+        .saturating_div(denominator.max(1));
+    scaled.min(cap).max(1)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn sample(
+        c2s_bytes: u64,
+        s2c_requested_bytes: u64,
+        s2c_written_bytes: u64,
+        s2c_write_ops: u64,
+        s2c_partial_writes: u64,
+        s2c_consecutive_pending_writes: u32,
+    ) -> RelaySignalSample {
+        RelaySignalSample {
+            c2s_bytes,
+            s2c_requested_bytes,
+            s2c_written_bytes,
+            s2c_write_ops,
+            s2c_partial_writes,
+            s2c_consecutive_pending_writes,
+        }
+    }
+
+    #[test]
+    fn test_soft_promotion_requires_tier1_and_tier2() {
+        let mut ctrl = SessionAdaptiveController::new(AdaptiveTier::Base);
+        let tick_secs = 0.25;
+        let mut promoted = None;
+        for _ in 0..8 {
+            promoted = ctrl.observe(
+                sample(
+                    300_000, // ~9.6 Mbps
+                    320_000, // incoming > outgoing to confirm tier2
+                    250_000,
+                    10,
+                    0,
+                    0,
+                ),
+                tick_secs,
+            );
+        }
+
+        let transition = promoted.expect("expected soft promotion");
+        assert_eq!(transition.from, AdaptiveTier::Base);
+        assert_eq!(transition.to, AdaptiveTier::Tier1);
+        assert_eq!(transition.reason, TierTransitionReason::SoftConfirmed);
+    }
+
+    #[test]
+    fn test_hard_promotion_on_pending_pressure() {
+        let mut ctrl = SessionAdaptiveController::new(AdaptiveTier::Base);
+        let transition = ctrl
+            .observe(
+                sample(10_000, 20_000, 10_000, 4, 1, 3),
+                0.25,
+            )
+            .expect("expected hard promotion");
+        assert_eq!(transition.reason, TierTransitionReason::HardPressure);
+        assert_eq!(transition.to, AdaptiveTier::Tier1);
+    }
+
+    #[test]
+    fn test_quiet_demotion_is_slow_and_stepwise() {
+        let mut ctrl = SessionAdaptiveController::new(AdaptiveTier::Tier2);
+        let mut demotion = None;
+        for _ in 0..QUIET_DEMOTE_TICKS {
+            demotion = ctrl.observe(sample(1, 1, 1, 1, 0, 0), 0.25);
+        }
+
+        let transition = demotion.expect("expected quiet demotion");
+        assert_eq!(transition.from, AdaptiveTier::Tier2);
+        assert_eq!(transition.to, AdaptiveTier::Tier1);
+        assert_eq!(transition.reason, TierTransitionReason::QuietDemotion);
+    }
+}

src/proxy/client.rs

@@ -40,6 +40,7 @@ use crate::proxy::handshake::{HandshakeSuccess, handle_mtproto_handshake, handle
 use crate::proxy::masking::handle_bad_client;
 use crate::proxy::middle_relay::handle_via_middle_proxy;
 use crate::proxy::route_mode::{RelayRouteMode, RouteRuntimeController};
+use crate::proxy::session_eviction::register_session;
 
 fn beobachten_ttl(config: &ProxyConfig) -> Duration {
     Duration::from_secs(config.general.beobachten_minutes.saturating_mul(60))
@@ -731,6 +732,17 @@ impl RunningClientHandler {
             return Err(e);
         }
 
+        let registration = register_session(&user, success.dc_idx);
+        if registration.replaced_existing {
+            stats.increment_reconnect_evict_total();
+            warn!(
+                user = %user,
+                dc = success.dc_idx,
+                "Reconnect detected: replacing active session for user+dc"
+            );
+        }
+        let session_lease = registration.lease;
+
         let route_snapshot = route_runtime.snapshot();
         let session_id = rng.u64();
         let relay_result = if config.general.use_middle_proxy
@@ -750,6 +762,7 @@ impl RunningClientHandler {
                     route_runtime.subscribe(),
                     route_snapshot,
                     session_id,
+                    session_lease.clone(),
                 )
                 .await
             } else {
@@ -766,6 +779,7 @@ impl RunningClientHandler {
                     route_runtime.subscribe(),
                     route_snapshot,
                     session_id,
+                    session_lease.clone(),
                 )
                 .await
             }
@@ -783,6 +797,7 @@ impl RunningClientHandler {
                 route_runtime.subscribe(),
                 route_snapshot,
                 session_id,
+                session_lease.clone(),
             )
             .await
         };

src/proxy/direct_relay.rs

@@ -17,6 +17,8 @@ use crate::proxy::route_mode::{
     ROUTE_SWITCH_ERROR_MSG, RelayRouteMode, RouteCutoverState, affected_cutover_state,
     cutover_stagger_delay,
 };
+use crate::proxy::adaptive_buffers;
+use crate::proxy::session_eviction::SessionLease;
 use crate::stats::Stats;
 use crate::stream::{BufferPool, CryptoReader, CryptoWriter};
 use crate::transport::UpstreamManager;
@@ -33,6 +35,7 @@ pub(crate) async fn handle_via_direct<R, W>(
     mut route_rx: watch::Receiver<RouteCutoverState>,
     route_snapshot: RouteCutoverState,
     session_id: u64,
+    session_lease: SessionLease,
 ) -> Result<()>
 where
     R: AsyncRead + Unpin + Send + 'static,
@@ -70,16 +73,26 @@ where
     stats.increment_user_curr_connects(user);
     stats.increment_current_connections_direct();
 
+    let seed_tier = adaptive_buffers::seed_tier_for_user(user);
+    let (c2s_copy_buf, s2c_copy_buf) = adaptive_buffers::direct_copy_buffers_for_tier(
+        seed_tier,
+        config.general.direct_relay_copy_buf_c2s_bytes,
+        config.general.direct_relay_copy_buf_s2c_bytes,
+    );
+
     let relay_result = relay_bidirectional(
         client_reader,
         client_writer,
         tg_reader,
         tg_writer,
-        config.general.direct_relay_copy_buf_c2s_bytes,
-        config.general.direct_relay_copy_buf_s2c_bytes,
+        c2s_copy_buf,
+        s2c_copy_buf,
         user,
+        success.dc_idx,
         Arc::clone(&stats),
         buffer_pool,
+        session_lease,
+        seed_tier,
     );
     tokio::pin!(relay_result);
     let relay_result = loop {

src/proxy/middle_relay.rs

@@ -20,6 +20,8 @@ use crate::proxy::route_mode::{
     RelayRouteMode, RouteCutoverState, ROUTE_SWITCH_ERROR_MSG, affected_cutover_state,
     cutover_stagger_delay,
 };
+use crate::proxy::adaptive_buffers::{self, AdaptiveTier};
+use crate::proxy::session_eviction::SessionLease;
 use crate::stats::Stats;
 use crate::stream::{BufferPool, CryptoReader, CryptoWriter};
 use crate::transport::middle_proxy::{MePool, MeResponse, proto_flags_for_tag};
@@ -59,8 +61,8 @@ struct MeD2cFlushPolicy {
 }
 
 impl MeD2cFlushPolicy {
-    fn from_config(config: &ProxyConfig) -> Self {
-        Self {
+    fn from_config(config: &ProxyConfig, tier: AdaptiveTier) -> Self {
+        let base = Self {
             max_frames: config
                 .general
                 .me_d2c_flush_batch_max_frames
@@ -71,6 +73,18 @@ impl MeD2cFlushPolicy {
                 .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,
+        };
+        let (max_frames, max_bytes, max_delay) = adaptive_buffers::me_flush_policy_for_tier(
+            tier,
+            base.max_frames,
+            base.max_bytes,
+            base.max_delay,
+        );
+        Self {
+            max_frames,
+            max_bytes,
+            max_delay,
+            ack_flush_immediate: base.ack_flush_immediate,
         }
     }
 }
@@ -235,6 +249,7 @@ pub(crate) async fn handle_via_middle_proxy<R, W>(
     mut route_rx: watch::Receiver<RouteCutoverState>,
     route_snapshot: RouteCutoverState,
     session_id: u64,
+    session_lease: SessionLease,
 ) -> Result<()>
 where
     R: AsyncRead + Unpin + Send + 'static,
@@ -244,6 +259,7 @@ where
     let peer = success.peer;
     let proto_tag = success.proto_tag;
     let pool_generation = me_pool.current_generation();
+    let seed_tier = adaptive_buffers::seed_tier_for_user(&user);
 
     debug!(
         user = %user,
@@ -295,6 +311,15 @@ where
         return Err(ProxyError::Proxy(ROUTE_SWITCH_ERROR_MSG.to_string()));
     }
 
+    if session_lease.is_stale() {
+        stats.increment_reconnect_stale_close_total();
+        let _ = me_pool.send_close(conn_id).await;
+        me_pool.registry().unregister(conn_id).await;
+        stats.decrement_current_connections_me();
+        stats.decrement_user_curr_connects(&user);
+        return Err(ProxyError::Proxy("Session evicted by reconnect".to_string()));
+    }
+
     // Per-user ad_tag from access.user_ad_tags; fallback to general.ad_tag (hot-reloadable)
     let user_tag: Option<Vec<u8>> = config
         .access
@@ -368,7 +393,7 @@ where
     let rng_clone = rng.clone();
     let user_clone = user.clone();
     let bytes_me2c_clone = bytes_me2c.clone();
-    let d2c_flush_policy = MeD2cFlushPolicy::from_config(&config);
+    let d2c_flush_policy = MeD2cFlushPolicy::from_config(&config, seed_tier);
     let me_writer = tokio::spawn(async move {
         let mut writer = crypto_writer;
         let mut frame_buf = Vec::with_capacity(16 * 1024);
@@ -528,6 +553,12 @@ where
     let mut frame_counter: u64 = 0;
     let mut route_watch_open = true;
     loop {
+        if session_lease.is_stale() {
+            stats.increment_reconnect_stale_close_total();
+            let _ = enqueue_c2me_command(&c2me_tx, C2MeCommand::Close).await;
+            main_result = Err(ProxyError::Proxy("Session evicted by reconnect".to_string()));
+            break;
+        }
         if let Some(cutover) = affected_cutover_state(
             &route_rx,
             RelayRouteMode::Middle,
@@ -636,6 +667,7 @@ where
         frames_ok = frame_counter,
         "ME relay cleanup"
     );
+    adaptive_buffers::record_user_tier(&user, seed_tier);
     me_pool.registry().unregister(conn_id).await;
     stats.decrement_current_connections_me();
     stats.decrement_user_curr_connects(&user);

src/proxy/mod.rs

@@ -1,5 +1,6 @@
 //! Proxy Defs
 
+pub mod adaptive_buffers;
 pub mod client;
 pub mod direct_relay;
 pub mod handshake;
@@ -7,6 +8,7 @@ pub mod masking;
 pub mod middle_relay;
 pub mod route_mode;
 pub mod relay;
+pub mod session_eviction;
 
 pub use client::ClientHandler;
 #[allow(unused_imports)]

src/proxy/relay.rs

@@ -63,6 +63,10 @@ use tokio::io::{
 use tokio::time::Instant;
 use tracing::{debug, trace, warn};
 use crate::error::Result;
+use crate::proxy::adaptive_buffers::{
+    self, AdaptiveTier, RelaySignalSample, SessionAdaptiveController, TierTransitionReason,
+};
+use crate::proxy::session_eviction::SessionLease;
 use crate::stats::Stats;
 use crate::stream::BufferPool;
 
@@ -79,6 +83,7 @@ const ACTIVITY_TIMEOUT: Duration = Duration::from_secs(1800);
 /// 10 seconds gives responsive timeout detection (±10s accuracy)
 /// without measurable overhead from atomic reads.
 const WATCHDOG_INTERVAL: Duration = Duration::from_secs(10);
+const ADAPTIVE_TICK: Duration = Duration::from_millis(250);
 
 // ============= CombinedStream =============
 
@@ -155,6 +160,16 @@ struct SharedCounters {
     s2c_ops: AtomicU64,
     /// Milliseconds since relay epoch of last I/O activity
     last_activity_ms: AtomicU64,
+    /// Bytes requested to write to client (S→C direction).
+    s2c_requested_bytes: AtomicU64,
+    /// Total write operations for S→C direction.
+    s2c_write_ops: AtomicU64,
+    /// Number of partial writes to client.
+    s2c_partial_writes: AtomicU64,
+    /// Number of times S→C poll_write returned Pending.
+    s2c_pending_writes: AtomicU64,
+    /// Consecutive pending writes in S→C direction.
+    s2c_consecutive_pending_writes: AtomicU64,
 }
 
 impl SharedCounters {
@@ -165,6 +180,11 @@ impl SharedCounters {
             c2s_ops: AtomicU64::new(0),
             s2c_ops: AtomicU64::new(0),
             last_activity_ms: AtomicU64::new(0),
+            s2c_requested_bytes: AtomicU64::new(0),
+            s2c_write_ops: AtomicU64::new(0),
+            s2c_partial_writes: AtomicU64::new(0),
+            s2c_pending_writes: AtomicU64::new(0),
+            s2c_consecutive_pending_writes: AtomicU64::new(0),
         }
     }
 
@@ -259,9 +279,21 @@ impl<S: AsyncWrite + Unpin> AsyncWrite for StatsIo<S> {
         buf: &[u8],
     ) -> Poll<io::Result<usize>> {
         let this = self.get_mut();
+        this.counters
+            .s2c_requested_bytes
+            .fetch_add(buf.len() as u64, Ordering::Relaxed);
 
         match Pin::new(&mut this.inner).poll_write(cx, buf) {
             Poll::Ready(Ok(n)) => {
+                this.counters.s2c_write_ops.fetch_add(1, Ordering::Relaxed);
+                this.counters
+                    .s2c_consecutive_pending_writes
+                    .store(0, Ordering::Relaxed);
+                if n < buf.len() {
+                    this.counters
+                        .s2c_partial_writes
+                        .fetch_add(1, Ordering::Relaxed);
+                }
                 if n > 0 {
                     // S→C: data written to client
                     this.counters.s2c_bytes.fetch_add(n as u64, Ordering::Relaxed);
@@ -275,6 +307,15 @@ impl<S: AsyncWrite + Unpin> AsyncWrite for StatsIo<S> {
                 }
                 Poll::Ready(Ok(n))
             }
+            Poll::Pending => {
+                this.counters
+                    .s2c_pending_writes
+                    .fetch_add(1, Ordering::Relaxed);
+                this.counters
+                    .s2c_consecutive_pending_writes
+                    .fetch_add(1, Ordering::Relaxed);
+                Poll::Pending
+            }
             other => other,
         }
     }
@@ -316,8 +357,11 @@ pub async fn relay_bidirectional<CR, CW, SR, SW>(
     c2s_buf_size: usize,
     s2c_buf_size: usize,
     user: &str,
+    dc_idx: i16,
     stats: Arc<Stats>,
     _buffer_pool: Arc<BufferPool>,
+    session_lease: SessionLease,
+    seed_tier: AdaptiveTier,
 ) -> Result<()>
 where
     CR: AsyncRead + Unpin + Send + 'static,
@@ -345,13 +389,33 @@ where
     // ── Watchdog: activity timeout + periodic rate logging ──────────
     let wd_counters = Arc::clone(&counters);
     let wd_user = user_owned.clone();
+    let wd_dc = dc_idx;
+    let wd_stats = Arc::clone(&stats);
+    let wd_session = session_lease.clone();
 
     let watchdog = async {
-        let mut prev_c2s: u64 = 0;
-        let mut prev_s2c: u64 = 0;
+        let mut prev_c2s_log: u64 = 0;
+        let mut prev_s2c_log: u64 = 0;
+        let mut prev_c2s_sample: u64 = 0;
+        let mut prev_s2c_requested_sample: u64 = 0;
+        let mut prev_s2c_written_sample: u64 = 0;
+        let mut prev_s2c_write_ops_sample: u64 = 0;
+        let mut prev_s2c_partial_sample: u64 = 0;
+        let mut accumulated_log = Duration::ZERO;
+        let mut adaptive = SessionAdaptiveController::new(seed_tier);
 
         loop {
-            tokio::time::sleep(WATCHDOG_INTERVAL).await;
+            tokio::time::sleep(ADAPTIVE_TICK).await;
+
+            if wd_session.is_stale() {
+                wd_stats.increment_reconnect_stale_close_total();
+                warn!(
+                    user = %wd_user,
+                    dc = wd_dc,
+                    "Session evicted by reconnect"
+                );
+                return;
+            }
 
             let now = Instant::now();
             let idle = wd_counters.idle_duration(now, epoch);
@@ -370,11 +434,80 @@ where
                 return; // Causes select! to cancel copy_bidirectional
             }
 
+            let c2s_total = wd_counters.c2s_bytes.load(Ordering::Relaxed);
+            let s2c_requested_total = wd_counters
+                .s2c_requested_bytes
+                .load(Ordering::Relaxed);
+            let s2c_written_total = wd_counters.s2c_bytes.load(Ordering::Relaxed);
+            let s2c_write_ops_total = wd_counters
+                .s2c_write_ops
+                .load(Ordering::Relaxed);
+            let s2c_partial_total = wd_counters
+                .s2c_partial_writes
+                .load(Ordering::Relaxed);
+            let consecutive_pending = wd_counters
+                .s2c_consecutive_pending_writes
+                .load(Ordering::Relaxed) as u32;
+
+            let sample = RelaySignalSample {
+                c2s_bytes: c2s_total.saturating_sub(prev_c2s_sample),
+                s2c_requested_bytes: s2c_requested_total
+                    .saturating_sub(prev_s2c_requested_sample),
+                s2c_written_bytes: s2c_written_total
+                    .saturating_sub(prev_s2c_written_sample),
+                s2c_write_ops: s2c_write_ops_total
+                    .saturating_sub(prev_s2c_write_ops_sample),
+                s2c_partial_writes: s2c_partial_total
+                    .saturating_sub(prev_s2c_partial_sample),
+                s2c_consecutive_pending_writes: consecutive_pending,
+            };
+
+            if let Some(transition) = adaptive.observe(sample, ADAPTIVE_TICK.as_secs_f64()) {
+                match transition.reason {
+                    TierTransitionReason::SoftConfirmed => {
+                        wd_stats.increment_relay_adaptive_promotions_total();
+                    }
+                    TierTransitionReason::HardPressure => {
+                        wd_stats.increment_relay_adaptive_promotions_total();
+                        wd_stats.increment_relay_adaptive_hard_promotions_total();
+                    }
+                    TierTransitionReason::QuietDemotion => {
+                        wd_stats.increment_relay_adaptive_demotions_total();
+                    }
+                }
+                adaptive_buffers::record_user_tier(&wd_user, adaptive.max_tier_seen());
+                debug!(
+                    user = %wd_user,
+                    dc = wd_dc,
+                    from_tier = transition.from.as_u8(),
+                    to_tier = transition.to.as_u8(),
+                    reason = ?transition.reason,
+                    throughput_ema_bps = sample
+                        .c2s_bytes
+                        .max(sample.s2c_written_bytes)
+                        .saturating_mul(8)
+                        .saturating_mul(4),
+                    "Adaptive relay tier transition"
+                );
+            }
+
+            prev_c2s_sample = c2s_total;
+            prev_s2c_requested_sample = s2c_requested_total;
+            prev_s2c_written_sample = s2c_written_total;
+            prev_s2c_write_ops_sample = s2c_write_ops_total;
+            prev_s2c_partial_sample = s2c_partial_total;
+
+            accumulated_log = accumulated_log.saturating_add(ADAPTIVE_TICK);
+            if accumulated_log < WATCHDOG_INTERVAL {
+                continue;
+            }
+            accumulated_log = Duration::ZERO;
+
             // ── Periodic rate logging ───────────────────────────────
             let c2s = wd_counters.c2s_bytes.load(Ordering::Relaxed);
             let s2c = wd_counters.s2c_bytes.load(Ordering::Relaxed);
-            let c2s_delta = c2s - prev_c2s;
-            let s2c_delta = s2c - prev_s2c;
+            let c2s_delta = c2s.saturating_sub(prev_c2s_log);
+            let s2c_delta = s2c.saturating_sub(prev_s2c_log);
 
             if c2s_delta > 0 || s2c_delta > 0 {
                 let secs = WATCHDOG_INTERVAL.as_secs_f64();
@@ -388,8 +521,8 @@ where
                 );
             }
 
-            prev_c2s = c2s;
-            prev_s2c = s2c;
+            prev_c2s_log = c2s;
+            prev_s2c_log = s2c;
         }
     };
 
@@ -424,6 +557,7 @@ where
     let c2s_ops = counters.c2s_ops.load(Ordering::Relaxed);
     let s2c_ops = counters.s2c_ops.load(Ordering::Relaxed);
     let duration = epoch.elapsed();
+    adaptive_buffers::record_user_tier(&user_owned, seed_tier);
 
     match copy_result {
         Some(Ok((c2s, s2c))) => {

src/proxy/session_eviction.rs

@@ -0,0 +1,46 @@
+/// Session eviction is intentionally disabled in runtime.
+///
+/// The initial `user+dc` single-lease model caused valid parallel client
+/// connections to evict each other. Keep the API shape for compatibility,
+/// but make it a no-op until a safer policy is introduced.
+
+#[derive(Debug, Clone, Default)]
+pub struct SessionLease;
+
+impl SessionLease {
+    pub fn is_stale(&self) -> bool {
+        false
+    }
+
+    #[allow(dead_code)]
+    pub fn release(&self) {}
+}
+
+pub struct RegistrationResult {
+    pub lease: SessionLease,
+    pub replaced_existing: bool,
+}
+
+pub fn register_session(_user: &str, _dc_idx: i16) -> RegistrationResult {
+    RegistrationResult {
+        lease: SessionLease,
+        replaced_existing: false,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_session_eviction_disabled_behavior() {
+        let first = register_session("alice", 2);
+        let second = register_session("alice", 2);
+        assert!(!first.replaced_existing);
+        assert!(!second.replaced_existing);
+        assert!(!first.lease.is_stale());
+        assert!(!second.lease.is_stale());
+        first.lease.release();
+        second.lease.release();
+    }
+}

src/stats/mod.rs

@@ -120,6 +120,8 @@ pub struct Stats {
     pool_swap_total: AtomicU64,
     pool_drain_active: AtomicU64,
     pool_force_close_total: AtomicU64,
+    pool_drain_soft_evict_total: AtomicU64,
+    pool_drain_soft_evict_writer_total: AtomicU64,
     pool_stale_pick_total: AtomicU64,
     me_writer_removed_total: AtomicU64,
     me_writer_removed_unexpected_total: AtomicU64,
@@ -133,6 +135,11 @@ pub struct Stats {
     me_inline_recovery_total: AtomicU64,
     ip_reservation_rollback_tcp_limit_total: AtomicU64,
     ip_reservation_rollback_quota_limit_total: AtomicU64,
+    relay_adaptive_promotions_total: AtomicU64,
+    relay_adaptive_demotions_total: AtomicU64,
+    relay_adaptive_hard_promotions_total: AtomicU64,
+    reconnect_evict_total: AtomicU64,
+    reconnect_stale_close_total: AtomicU64,
     telemetry_core_enabled: AtomicBool,
     telemetry_user_enabled: AtomicBool,
     telemetry_me_level: AtomicU8,
@@ -285,6 +292,36 @@ impl Stats {
     pub fn decrement_current_connections_me(&self) {
         Self::decrement_atomic_saturating(&self.current_connections_me);
     }
+    pub fn increment_relay_adaptive_promotions_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.relay_adaptive_promotions_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_relay_adaptive_demotions_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.relay_adaptive_demotions_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_relay_adaptive_hard_promotions_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.relay_adaptive_hard_promotions_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_reconnect_evict_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.reconnect_evict_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_reconnect_stale_close_total(&self) {
+        if self.telemetry_core_enabled() {
+            self.reconnect_stale_close_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
     pub fn increment_handshake_timeouts(&self) {
         if self.telemetry_core_enabled() {
             self.handshake_timeouts.fetch_add(1, Ordering::Relaxed);
@@ -680,6 +717,18 @@ impl Stats {
             self.pool_force_close_total.fetch_add(1, Ordering::Relaxed);
         }
     }
+    pub fn increment_pool_drain_soft_evict_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.pool_drain_soft_evict_total
+                .fetch_add(1, Ordering::Relaxed);
+        }
+    }
+    pub fn increment_pool_drain_soft_evict_writer_total(&self) {
+        if self.telemetry_me_allows_normal() {
+            self.pool_drain_soft_evict_writer_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);
@@ -933,6 +982,22 @@ impl Stats {
         self.get_current_connections_direct()
             .saturating_add(self.get_current_connections_me())
     }
+    pub fn get_relay_adaptive_promotions_total(&self) -> u64 {
+        self.relay_adaptive_promotions_total.load(Ordering::Relaxed)
+    }
+    pub fn get_relay_adaptive_demotions_total(&self) -> u64 {
+        self.relay_adaptive_demotions_total.load(Ordering::Relaxed)
+    }
+    pub fn get_relay_adaptive_hard_promotions_total(&self) -> u64 {
+        self.relay_adaptive_hard_promotions_total
+            .load(Ordering::Relaxed)
+    }
+    pub fn get_reconnect_evict_total(&self) -> u64 {
+        self.reconnect_evict_total.load(Ordering::Relaxed)
+    }
+    pub fn get_reconnect_stale_close_total(&self) -> u64 {
+        self.reconnect_stale_close_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) }
@@ -1185,6 +1250,12 @@ impl Stats {
     pub fn get_pool_force_close_total(&self) -> u64 {
         self.pool_force_close_total.load(Ordering::Relaxed)
     }
+    pub fn get_pool_drain_soft_evict_total(&self) -> u64 {
+        self.pool_drain_soft_evict_total.load(Ordering::Relaxed)
+    }
+    pub fn get_pool_drain_soft_evict_writer_total(&self) -> u64 {
+        self.pool_drain_soft_evict_writer_total.load(Ordering::Relaxed)
+    }
     pub fn get_pool_stale_pick_total(&self) -> u64 {
         self.pool_stale_pick_total.load(Ordering::Relaxed)
     }
@@ -1258,6 +1329,9 @@ impl Stats {
     }
     
     pub fn decrement_user_curr_connects(&self, user: &str) {
+        if !self.telemetry_user_enabled() {
+            return;
+        }
         self.maybe_cleanup_user_stats();
         if let Some(stats) = self.user_stats.get(user) {
             Self::touch_user_stats(stats.value());

src/stream/buffer_pool.rs

@@ -14,8 +14,7 @@ use std::sync::Arc;
 // ============= Configuration =============
 
 /// Default buffer size
-/// CHANGED: Reduced from 64KB to 16KB to match TLS record size and prevent bufferbloat.
-pub const DEFAULT_BUFFER_SIZE: usize = 16 * 1024;
+pub const DEFAULT_BUFFER_SIZE: usize = 64 * 1024;
 
 /// Default maximum number of pooled buffers
 pub const DEFAULT_MAX_BUFFERS: usize = 1024;

src/transport/middle_proxy/config_updater.rs

@@ -299,6 +299,11 @@ async fn run_update_cycle(
         cfg.general.hardswap,
         cfg.general.me_pool_drain_ttl_secs,
         cfg.general.me_pool_drain_threshold,
+        cfg.general.me_pool_drain_soft_evict_enabled,
+        cfg.general.me_pool_drain_soft_evict_grace_secs,
+        cfg.general.me_pool_drain_soft_evict_per_writer,
+        cfg.general.me_pool_drain_soft_evict_budget_per_core,
+        cfg.general.me_pool_drain_soft_evict_cooldown_ms,
         cfg.general.effective_me_pool_force_close_secs(),
         cfg.general.me_pool_min_fresh_ratio,
         cfg.general.me_hardswap_warmup_delay_min_ms,
@@ -526,6 +531,11 @@ pub async fn me_config_updater(
                     cfg.general.hardswap,
                     cfg.general.me_pool_drain_ttl_secs,
                     cfg.general.me_pool_drain_threshold,
+                    cfg.general.me_pool_drain_soft_evict_enabled,
+                    cfg.general.me_pool_drain_soft_evict_grace_secs,
+                    cfg.general.me_pool_drain_soft_evict_per_writer,
+                    cfg.general.me_pool_drain_soft_evict_budget_per_core,
+                    cfg.general.me_pool_drain_soft_evict_cooldown_ms,
                     cfg.general.effective_me_pool_force_close_secs(),
                     cfg.general.me_pool_min_fresh_ratio,
                     cfg.general.me_hardswap_warmup_delay_min_ms,

src/transport/middle_proxy/health.rs

@@ -25,6 +25,11 @@ const HEALTH_RECONNECT_BUDGET_PER_CORE: usize = 2;
 const HEALTH_RECONNECT_BUDGET_PER_DC: usize = 1;
 const HEALTH_RECONNECT_BUDGET_MIN: usize = 4;
 const HEALTH_RECONNECT_BUDGET_MAX: usize = 128;
+const HEALTH_DRAIN_CLOSE_BUDGET_PER_CORE: usize = 16;
+const HEALTH_DRAIN_CLOSE_BUDGET_MIN: usize = 16;
+const HEALTH_DRAIN_CLOSE_BUDGET_MAX: usize = 256;
+const HEALTH_DRAIN_SOFT_EVICT_BUDGET_MIN: usize = 8;
+const HEALTH_DRAIN_SOFT_EVICT_BUDGET_MAX: usize = 256;
 
 #[derive(Debug, Clone)]
 struct DcFloorPlanEntry {
@@ -63,6 +68,7 @@ pub async fn me_health_monitor(pool: Arc<MePool>, rng: Arc<SecureRandom>, _min_c
     let mut adaptive_recover_until: HashMap<(i32, IpFamily), Instant> = HashMap::new();
     let mut floor_warn_next_allowed: HashMap<(i32, IpFamily), Instant> = HashMap::new();
     let mut drain_warn_next_allowed: HashMap<u64, Instant> = HashMap::new();
+    let mut drain_soft_evict_next_allowed: HashMap<u64, Instant> = HashMap::new();
     let mut degraded_interval = true;
     loop {
         let interval = if degraded_interval {
@@ -72,7 +78,12 @@ pub async fn me_health_monitor(pool: Arc<MePool>, rng: Arc<SecureRandom>, _min_c
         };
         tokio::time::sleep(interval).await;
         pool.prune_closed_writers().await;
-        reap_draining_writers(&pool, &mut drain_warn_next_allowed).await;
+        reap_draining_writers(
+            &pool,
+            &mut drain_warn_next_allowed,
+            &mut drain_soft_evict_next_allowed,
+        )
+        .await;
         let v4_degraded = check_family(
             IpFamily::V4,
             &pool,
@@ -111,9 +122,10 @@ pub async fn me_health_monitor(pool: Arc<MePool>, rng: Arc<SecureRandom>, _min_c
     }
 }
 
-async fn reap_draining_writers(
+pub(super) async fn reap_draining_writers(
     pool: &Arc<MePool>,
     warn_next_allowed: &mut HashMap<u64, Instant>,
+    soft_evict_next_allowed: &mut HashMap<u64, Instant>,
 ) {
     let now_epoch_secs = MePool::now_epoch_secs();
     let now = Instant::now();
@@ -122,14 +134,22 @@ async fn reap_draining_writers(
         .me_pool_drain_threshold
         .load(std::sync::atomic::Ordering::Relaxed);
     let writers = pool.writers.read().await.clone();
+    let activity = pool.registry.writer_activity_snapshot().await;
     let mut draining_writers = Vec::new();
+    let mut empty_writer_ids = Vec::<u64>::new();
+    let mut force_close_writer_ids = Vec::<u64>::new();
     for writer in writers {
         if !writer.draining.load(std::sync::atomic::Ordering::Relaxed) {
             continue;
         }
-        let is_empty = pool.registry.is_writer_empty(writer.id).await;
-        if is_empty {
-            pool.remove_writer_and_close_clients(writer.id).await;
+        if activity
+            .bound_clients_by_writer
+            .get(&writer.id)
+            .copied()
+            .unwrap_or(0)
+            == 0
+        {
+            empty_writer_ids.push(writer.id);
             continue;
         }
         draining_writers.push(writer);
@@ -156,12 +176,13 @@ async fn reap_draining_writers(
             "ME draining writer threshold exceeded, force-closing oldest draining writers"
         );
         for writer in draining_writers.drain(..overflow) {
-            pool.stats.increment_pool_force_close_total();
-            pool.remove_writer_and_close_clients(writer.id).await;
+            force_close_writer_ids.push(writer.id);
         }
     }
 
-    for writer in draining_writers {
+    let mut active_draining_writer_ids = HashSet::with_capacity(draining_writers.len());
+    for writer in &draining_writers {
+        active_draining_writer_ids.insert(writer.id);
         let drain_started_at_epoch_secs = writer
             .draining_started_at_epoch_secs
             .load(std::sync::atomic::Ordering::Relaxed);
@@ -191,10 +212,152 @@ async fn reap_draining_writers(
             .load(std::sync::atomic::Ordering::Relaxed);
         if deadline_epoch_secs != 0 && now_epoch_secs >= deadline_epoch_secs {
             warn!(writer_id = writer.id, "Drain timeout, force-closing");
-            pool.stats.increment_pool_force_close_total();
-            pool.remove_writer_and_close_clients(writer.id).await;
+            force_close_writer_ids.push(writer.id);
+            active_draining_writer_ids.remove(&writer.id);
         }
     }
+
+    warn_next_allowed.retain(|writer_id, _| active_draining_writer_ids.contains(writer_id));
+    soft_evict_next_allowed.retain(|writer_id, _| active_draining_writer_ids.contains(writer_id));
+
+    if pool.drain_soft_evict_enabled() && drain_ttl_secs > 0 && !draining_writers.is_empty() {
+        let mut force_close_ids = HashSet::<u64>::with_capacity(force_close_writer_ids.len());
+        for writer_id in &force_close_writer_ids {
+            force_close_ids.insert(*writer_id);
+        }
+        let soft_grace_secs = pool.drain_soft_evict_grace_secs();
+        let soft_trigger_age_secs = drain_ttl_secs.saturating_add(soft_grace_secs);
+        let per_writer_limit = pool.drain_soft_evict_per_writer();
+        let soft_budget = health_drain_soft_evict_budget(pool);
+        let soft_cooldown = pool.drain_soft_evict_cooldown();
+        let mut soft_evicted_total = 0usize;
+
+        for writer in &draining_writers {
+            if soft_evicted_total >= soft_budget {
+                break;
+            }
+            if force_close_ids.contains(&writer.id) {
+                continue;
+            }
+            if pool.writer_accepts_new_binding(writer) {
+                continue;
+            }
+            let started_epoch_secs = writer
+                .draining_started_at_epoch_secs
+                .load(std::sync::atomic::Ordering::Relaxed);
+            if started_epoch_secs == 0
+                || now_epoch_secs.saturating_sub(started_epoch_secs) < soft_trigger_age_secs
+            {
+                continue;
+            }
+            if !should_emit_writer_warn(
+                soft_evict_next_allowed,
+                writer.id,
+                now,
+                soft_cooldown,
+            ) {
+                continue;
+            }
+
+            let remaining_budget = soft_budget.saturating_sub(soft_evicted_total);
+            let limit = per_writer_limit.min(remaining_budget);
+            if limit == 0 {
+                break;
+            }
+            let conn_ids = pool
+                .registry
+                .bound_conn_ids_for_writer_limited(writer.id, limit)
+                .await;
+            if conn_ids.is_empty() {
+                continue;
+            }
+
+            let mut evicted_for_writer = 0usize;
+            for conn_id in conn_ids {
+                if pool.registry.evict_bound_conn_if_writer(conn_id, writer.id).await {
+                    evicted_for_writer = evicted_for_writer.saturating_add(1);
+                    soft_evicted_total = soft_evicted_total.saturating_add(1);
+                    pool.stats.increment_pool_drain_soft_evict_total();
+                    if soft_evicted_total >= soft_budget {
+                        break;
+                    }
+                }
+            }
+
+            if evicted_for_writer > 0 {
+                pool.stats.increment_pool_drain_soft_evict_writer_total();
+                info!(
+                    writer_id = writer.id,
+                    writer_dc = writer.writer_dc,
+                    endpoint = %writer.addr,
+                    drained_connections = evicted_for_writer,
+                    soft_budget,
+                    soft_trigger_age_secs,
+                    "ME draining writer soft-evicted bound clients"
+                );
+            }
+        }
+    }
+
+    let close_budget = health_drain_close_budget();
+    let requested_force_close = force_close_writer_ids.len();
+    let requested_empty_close = empty_writer_ids.len();
+    let requested_close_total = requested_force_close.saturating_add(requested_empty_close);
+    let mut closed_writer_ids = HashSet::<u64>::new();
+    let mut closed_total = 0usize;
+    for writer_id in force_close_writer_ids {
+        if closed_total >= close_budget {
+            break;
+        }
+        if !closed_writer_ids.insert(writer_id) {
+            continue;
+        }
+        pool.stats.increment_pool_force_close_total();
+        pool.remove_writer_and_close_clients(writer_id).await;
+        closed_total = closed_total.saturating_add(1);
+    }
+    for writer_id in empty_writer_ids {
+        if closed_total >= close_budget {
+            break;
+        }
+        if !closed_writer_ids.insert(writer_id) {
+            continue;
+        }
+        pool.remove_writer_and_close_clients(writer_id).await;
+        closed_total = closed_total.saturating_add(1);
+    }
+
+    let pending_close_total = requested_close_total.saturating_sub(closed_total);
+    if pending_close_total > 0 {
+        warn!(
+            close_budget,
+            closed_total,
+            pending_close_total,
+            "ME draining close backlog deferred to next health cycle"
+        );
+    }
+}
+
+pub(super) fn health_drain_close_budget() -> usize {
+    let cpu_cores = std::thread::available_parallelism()
+        .map(std::num::NonZeroUsize::get)
+        .unwrap_or(1);
+    cpu_cores
+        .saturating_mul(HEALTH_DRAIN_CLOSE_BUDGET_PER_CORE)
+        .clamp(HEALTH_DRAIN_CLOSE_BUDGET_MIN, HEALTH_DRAIN_CLOSE_BUDGET_MAX)
+}
+
+pub(super) fn health_drain_soft_evict_budget(pool: &MePool) -> usize {
+    let cpu_cores = std::thread::available_parallelism()
+        .map(std::num::NonZeroUsize::get)
+        .unwrap_or(1);
+    let per_core = pool.drain_soft_evict_budget_per_core();
+    cpu_cores
+        .saturating_mul(per_core)
+        .clamp(
+            HEALTH_DRAIN_SOFT_EVICT_BUDGET_MIN,
+            HEALTH_DRAIN_SOFT_EVICT_BUDGET_MAX,
+        )
 }
 
 fn should_emit_writer_warn(
@@ -1382,6 +1545,11 @@ mod tests {
             general.hardswap,
             general.me_pool_drain_ttl_secs,
             general.me_pool_drain_threshold,
+            general.me_pool_drain_soft_evict_enabled,
+            general.me_pool_drain_soft_evict_grace_secs,
+            general.me_pool_drain_soft_evict_per_writer,
+            general.me_pool_drain_soft_evict_budget_per_core,
+            general.me_pool_drain_soft_evict_cooldown_ms,
             general.effective_me_pool_force_close_secs(),
             general.me_pool_min_fresh_ratio,
             general.me_hardswap_warmup_delay_min_ms,
@@ -1463,8 +1631,9 @@ mod tests {
         let conn_b = insert_draining_writer(&pool, 20, now_epoch_secs.saturating_sub(20)).await;
         let conn_c = insert_draining_writer(&pool, 30, now_epoch_secs.saturating_sub(10)).await;
         let mut warn_next_allowed = HashMap::new();
+        let mut soft_evict_next_allowed = HashMap::new();
 
-        reap_draining_writers(&pool, &mut warn_next_allowed).await;
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
 
         let writer_ids: Vec<u64> = pool.writers.read().await.iter().map(|writer| writer.id).collect();
         assert_eq!(writer_ids, vec![20, 30]);
@@ -1481,8 +1650,9 @@ mod tests {
         let conn_b = insert_draining_writer(&pool, 20, now_epoch_secs.saturating_sub(20)).await;
         let conn_c = insert_draining_writer(&pool, 30, now_epoch_secs.saturating_sub(10)).await;
         let mut warn_next_allowed = HashMap::new();
+        let mut soft_evict_next_allowed = HashMap::new();
 
-        reap_draining_writers(&pool, &mut warn_next_allowed).await;
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
 
         let writer_ids: Vec<u64> = pool.writers.read().await.iter().map(|writer| writer.id).collect();
         assert_eq!(writer_ids, vec![10, 20, 30]);

src/transport/middle_proxy/health_adversarial_tests.rs

@@ -0,0 +1,450 @@
+use std::collections::HashMap;
+use std::net::{IpAddr, Ipv4Addr, SocketAddr};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, AtomicU8, AtomicU32, AtomicU64, Ordering};
+use std::time::{Duration, Instant};
+
+use tokio::sync::mpsc;
+use tokio_util::sync::CancellationToken;
+
+use super::codec::WriterCommand;
+use super::health::{health_drain_close_budget, reap_draining_writers};
+use super::pool::{MePool, MeWriter, WriterContour};
+use super::registry::ConnMeta;
+use super::me_health_monitor;
+use crate::config::{GeneralConfig, MeRouteNoWriterMode, MeSocksKdfPolicy, MeWriterPickMode};
+use crate::crypto::SecureRandom;
+use crate::network::probe::NetworkDecision;
+use crate::stats::Stats;
+
+async fn make_pool(
+    me_pool_drain_threshold: u64,
+    me_health_interval_ms_unhealthy: u64,
+    me_health_interval_ms_healthy: u64,
+) -> (Arc<MePool>, Arc<SecureRandom>) {
+    let general = GeneralConfig {
+        me_pool_drain_threshold,
+        me_health_interval_ms_unhealthy,
+        me_health_interval_ms_healthy,
+        ..GeneralConfig::default()
+    };
+
+    let rng = Arc::new(SecureRandom::new());
+    let pool = MePool::new(
+        None,
+        vec![1u8; 32],
+        None,
+        false,
+        None,
+        Vec::new(),
+        1,
+        None,
+        12,
+        1200,
+        HashMap::new(),
+        HashMap::new(),
+        None,
+        NetworkDecision::default(),
+        None,
+        rng.clone(),
+        Arc::new(Stats::default()),
+        general.me_keepalive_enabled,
+        general.me_keepalive_interval_secs,
+        general.me_keepalive_jitter_secs,
+        general.me_keepalive_payload_random,
+        general.rpc_proxy_req_every,
+        general.me_warmup_stagger_enabled,
+        general.me_warmup_step_delay_ms,
+        general.me_warmup_step_jitter_ms,
+        general.me_reconnect_max_concurrent_per_dc,
+        general.me_reconnect_backoff_base_ms,
+        general.me_reconnect_backoff_cap_ms,
+        general.me_reconnect_fast_retry_count,
+        general.me_single_endpoint_shadow_writers,
+        general.me_single_endpoint_outage_mode_enabled,
+        general.me_single_endpoint_outage_disable_quarantine,
+        general.me_single_endpoint_outage_backoff_min_ms,
+        general.me_single_endpoint_outage_backoff_max_ms,
+        general.me_single_endpoint_shadow_rotate_every_secs,
+        general.me_floor_mode,
+        general.me_adaptive_floor_idle_secs,
+        general.me_adaptive_floor_min_writers_single_endpoint,
+        general.me_adaptive_floor_min_writers_multi_endpoint,
+        general.me_adaptive_floor_recover_grace_secs,
+        general.me_adaptive_floor_writers_per_core_total,
+        general.me_adaptive_floor_cpu_cores_override,
+        general.me_adaptive_floor_max_extra_writers_single_per_core,
+        general.me_adaptive_floor_max_extra_writers_multi_per_core,
+        general.me_adaptive_floor_max_active_writers_per_core,
+        general.me_adaptive_floor_max_warm_writers_per_core,
+        general.me_adaptive_floor_max_active_writers_global,
+        general.me_adaptive_floor_max_warm_writers_global,
+        general.hardswap,
+        general.me_pool_drain_ttl_secs,
+        general.me_pool_drain_threshold,
+        general.me_pool_drain_soft_evict_enabled,
+        general.me_pool_drain_soft_evict_grace_secs,
+        general.me_pool_drain_soft_evict_per_writer,
+        general.me_pool_drain_soft_evict_budget_per_core,
+        general.me_pool_drain_soft_evict_cooldown_ms,
+        general.effective_me_pool_force_close_secs(),
+        general.me_pool_min_fresh_ratio,
+        general.me_hardswap_warmup_delay_min_ms,
+        general.me_hardswap_warmup_delay_max_ms,
+        general.me_hardswap_warmup_extra_passes,
+        general.me_hardswap_warmup_pass_backoff_base_ms,
+        general.me_bind_stale_mode,
+        general.me_bind_stale_ttl_secs,
+        general.me_secret_atomic_snapshot,
+        general.me_deterministic_writer_sort,
+        MeWriterPickMode::default(),
+        general.me_writer_pick_sample_size,
+        MeSocksKdfPolicy::default(),
+        general.me_writer_cmd_channel_capacity,
+        general.me_route_channel_capacity,
+        general.me_route_backpressure_base_timeout_ms,
+        general.me_route_backpressure_high_timeout_ms,
+        general.me_route_backpressure_high_watermark_pct,
+        general.me_reader_route_data_wait_ms,
+        general.me_health_interval_ms_unhealthy,
+        general.me_health_interval_ms_healthy,
+        general.me_warn_rate_limit_ms,
+        MeRouteNoWriterMode::default(),
+        general.me_route_no_writer_wait_ms,
+        general.me_route_inline_recovery_attempts,
+        general.me_route_inline_recovery_wait_ms,
+    );
+
+    (pool, rng)
+}
+
+async fn insert_draining_writer(
+    pool: &Arc<MePool>,
+    writer_id: u64,
+    drain_started_at_epoch_secs: u64,
+    bound_clients: usize,
+    drain_deadline_epoch_secs: u64,
+) {
+    let (tx, _writer_rx) = mpsc::channel::<WriterCommand>(8);
+    let writer = MeWriter {
+        id: writer_id,
+        addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 6000 + writer_id as u16),
+        source_ip: IpAddr::V4(Ipv4Addr::LOCALHOST),
+        writer_dc: 2,
+        generation: 1,
+        contour: Arc::new(AtomicU8::new(WriterContour::Draining.as_u8())),
+        created_at: Instant::now() - Duration::from_secs(writer_id),
+        tx: tx.clone(),
+        cancel: CancellationToken::new(),
+        degraded: Arc::new(AtomicBool::new(false)),
+        rtt_ema_ms_x10: Arc::new(AtomicU32::new(0)),
+        draining: Arc::new(AtomicBool::new(true)),
+        draining_started_at_epoch_secs: Arc::new(AtomicU64::new(drain_started_at_epoch_secs)),
+        drain_deadline_epoch_secs: Arc::new(AtomicU64::new(drain_deadline_epoch_secs)),
+        allow_drain_fallback: Arc::new(AtomicBool::new(false)),
+    };
+
+    pool.writers.write().await.push(writer);
+    pool.registry.register_writer(writer_id, tx).await;
+    pool.conn_count.fetch_add(1, Ordering::Relaxed);
+
+    for idx in 0..bound_clients {
+        let (conn_id, _rx) = pool.registry.register().await;
+        assert!(
+            pool.registry
+                .bind_writer(
+                    conn_id,
+                    writer_id,
+                    ConnMeta {
+                        target_dc: 2,
+                        client_addr: SocketAddr::new(
+                            IpAddr::V4(Ipv4Addr::LOCALHOST),
+                            8000 + idx as u16,
+                        ),
+                        our_addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443),
+                        proto_flags: 0,
+                    },
+                )
+                .await
+        );
+    }
+}
+
+async fn writer_count(pool: &Arc<MePool>) -> usize {
+    pool.writers.read().await.len()
+}
+
+async fn sorted_writer_ids(pool: &Arc<MePool>) -> Vec<u64> {
+    let mut ids = pool
+        .writers
+        .read()
+        .await
+        .iter()
+        .map(|writer| writer.id)
+        .collect::<Vec<_>>();
+    ids.sort_unstable();
+    ids
+}
+
+#[tokio::test]
+async fn reap_draining_writers_clears_warn_state_when_pool_empty() {
+    let (pool, _rng) = make_pool(128, 1, 1).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+    warn_next_allowed.insert(11, Instant::now() + Duration::from_secs(5));
+    warn_next_allowed.insert(22, Instant::now() + Duration::from_secs(5));
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert!(warn_next_allowed.is_empty());
+}
+
+#[tokio::test]
+async fn reap_draining_writers_respects_threshold_across_multiple_overflow_cycles() {
+    let threshold = 3u64;
+    let (pool, _rng) = make_pool(threshold, 1, 1).await;
+    pool.me_pool_drain_soft_evict_enabled
+        .store(false, Ordering::Relaxed);
+    let now_epoch_secs = MePool::now_epoch_secs();
+
+    for writer_id in 1..=60u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(600).saturating_add(writer_id),
+            1,
+            0,
+        )
+        .await;
+    }
+
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+    for _ in 0..64 {
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+        if writer_count(&pool).await <= threshold as usize {
+            break;
+        }
+    }
+
+    assert_eq!(writer_count(&pool).await, threshold as usize);
+    assert_eq!(sorted_writer_ids(&pool).await, vec![58, 59, 60]);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_handles_large_empty_writer_population() {
+    let (pool, _rng) = make_pool(128, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let total = health_drain_close_budget().saturating_mul(3).saturating_add(27);
+
+    for writer_id in 1..=total as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(120),
+            0,
+            0,
+        )
+        .await;
+    }
+
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+    for _ in 0..24 {
+        if writer_count(&pool).await == 0 {
+            break;
+        }
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+    }
+
+    assert_eq!(writer_count(&pool).await, 0);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_processes_mass_deadline_expiry_without_unbounded_growth() {
+    let (pool, _rng) = make_pool(128, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let total = health_drain_close_budget().saturating_mul(4).saturating_add(31);
+
+    for writer_id in 1..=total as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(180),
+            1,
+            now_epoch_secs.saturating_sub(1),
+        )
+        .await;
+    }
+
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+    for _ in 0..40 {
+        if writer_count(&pool).await == 0 {
+            break;
+        }
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+    }
+
+    assert_eq!(writer_count(&pool).await, 0);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_maintains_warn_state_subset_property_under_bulk_churn() {
+    let (pool, _rng) = make_pool(128, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    for wave in 0..40u64 {
+        for offset in 0..8u64 {
+            insert_draining_writer(
+                &pool,
+                wave * 100 + offset,
+                now_epoch_secs.saturating_sub(400 + offset),
+                1,
+                0,
+            )
+            .await;
+        }
+
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+        assert!(warn_next_allowed.len() <= writer_count(&pool).await);
+
+        let ids = sorted_writer_ids(&pool).await;
+        for writer_id in ids.into_iter().take(3) {
+            let _ = pool.remove_writer_and_close_clients(writer_id).await;
+        }
+
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+        assert!(warn_next_allowed.len() <= writer_count(&pool).await);
+    }
+}
+
+#[tokio::test]
+async fn reap_draining_writers_budgeted_cleanup_never_increases_pool_size() {
+    let (pool, _rng) = make_pool(5, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+
+    for writer_id in 1..=200u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(240).saturating_add(writer_id),
+            1,
+            0,
+        )
+        .await;
+    }
+
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+    let mut previous = writer_count(&pool).await;
+    for _ in 0..32 {
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+        let current = writer_count(&pool).await;
+        assert!(current <= previous);
+        previous = current;
+    }
+}
+
+#[tokio::test]
+async fn me_health_monitor_converges_to_threshold_under_live_injection_churn() {
+    let threshold = 7u64;
+    let (pool, rng) = make_pool(threshold, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+
+    for writer_id in 1..=40u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(300).saturating_add(writer_id),
+            1,
+            0,
+        )
+        .await;
+    }
+
+    let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
+
+    for wave in 0..8u64 {
+        for offset in 0..10u64 {
+            insert_draining_writer(
+                &pool,
+                1000 + wave * 100 + offset,
+                now_epoch_secs.saturating_sub(120).saturating_add(offset),
+                1,
+                0,
+            )
+            .await;
+        }
+        tokio::time::sleep(Duration::from_millis(5)).await;
+    }
+
+    tokio::time::sleep(Duration::from_millis(120)).await;
+    monitor.abort();
+    let _ = monitor.await;
+
+    assert!(writer_count(&pool).await <= threshold as usize);
+}
+
+#[tokio::test]
+async fn me_health_monitor_drains_deadline_storm_with_budgeted_progress() {
+    let (pool, rng) = make_pool(128, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+
+    for writer_id in 1..=220u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(120),
+            1,
+            now_epoch_secs.saturating_sub(1),
+        )
+        .await;
+    }
+
+    let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
+    tokio::time::sleep(Duration::from_millis(120)).await;
+    monitor.abort();
+    let _ = monitor.await;
+
+    assert_eq!(writer_count(&pool).await, 0);
+}
+
+#[tokio::test]
+async fn me_health_monitor_eliminates_mixed_empty_and_deadline_backlog() {
+    let threshold = 12u64;
+    let (pool, rng) = make_pool(threshold, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+
+    for writer_id in 1..=180u64 {
+        let bound_clients = if writer_id % 3 == 0 { 0 } else { 1 };
+        let deadline = if writer_id % 2 == 0 {
+            now_epoch_secs.saturating_sub(1)
+        } else {
+            0
+        };
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(250).saturating_add(writer_id),
+            bound_clients,
+            deadline,
+        )
+        .await;
+    }
+
+    let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
+    tokio::time::sleep(Duration::from_millis(140)).await;
+    monitor.abort();
+    let _ = monitor.await;
+
+    assert!(writer_count(&pool).await <= threshold as usize);
+}
+
+#[test]
+fn health_drain_close_budget_is_within_expected_bounds() {
+    let budget = health_drain_close_budget();
+    assert!((16..=256).contains(&budget));
+}

src/transport/middle_proxy/health_integration_tests.rs

@@ -0,0 +1,232 @@
+use std::collections::HashMap;
+use std::net::{IpAddr, Ipv4Addr, SocketAddr};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, AtomicU8, AtomicU32, AtomicU64, Ordering};
+use std::time::{Duration, Instant};
+
+use tokio::sync::mpsc;
+use tokio_util::sync::CancellationToken;
+
+use super::codec::WriterCommand;
+use super::health::health_drain_close_budget;
+use super::pool::{MePool, MeWriter, WriterContour};
+use super::registry::ConnMeta;
+use super::me_health_monitor;
+use crate::config::{GeneralConfig, MeRouteNoWriterMode, MeSocksKdfPolicy, MeWriterPickMode};
+use crate::crypto::SecureRandom;
+use crate::network::probe::NetworkDecision;
+use crate::stats::Stats;
+
+async fn make_pool(
+    me_pool_drain_threshold: u64,
+    me_health_interval_ms_unhealthy: u64,
+    me_health_interval_ms_healthy: u64,
+) -> (Arc<MePool>, Arc<SecureRandom>) {
+    let general = GeneralConfig {
+        me_pool_drain_threshold,
+        me_health_interval_ms_unhealthy,
+        me_health_interval_ms_healthy,
+        ..GeneralConfig::default()
+    };
+    let rng = Arc::new(SecureRandom::new());
+    let pool = MePool::new(
+        None,
+        vec![1u8; 32],
+        None,
+        false,
+        None,
+        Vec::new(),
+        1,
+        None,
+        12,
+        1200,
+        HashMap::new(),
+        HashMap::new(),
+        None,
+        NetworkDecision::default(),
+        None,
+        rng.clone(),
+        Arc::new(Stats::default()),
+        general.me_keepalive_enabled,
+        general.me_keepalive_interval_secs,
+        general.me_keepalive_jitter_secs,
+        general.me_keepalive_payload_random,
+        general.rpc_proxy_req_every,
+        general.me_warmup_stagger_enabled,
+        general.me_warmup_step_delay_ms,
+        general.me_warmup_step_jitter_ms,
+        general.me_reconnect_max_concurrent_per_dc,
+        general.me_reconnect_backoff_base_ms,
+        general.me_reconnect_backoff_cap_ms,
+        general.me_reconnect_fast_retry_count,
+        general.me_single_endpoint_shadow_writers,
+        general.me_single_endpoint_outage_mode_enabled,
+        general.me_single_endpoint_outage_disable_quarantine,
+        general.me_single_endpoint_outage_backoff_min_ms,
+        general.me_single_endpoint_outage_backoff_max_ms,
+        general.me_single_endpoint_shadow_rotate_every_secs,
+        general.me_floor_mode,
+        general.me_adaptive_floor_idle_secs,
+        general.me_adaptive_floor_min_writers_single_endpoint,
+        general.me_adaptive_floor_min_writers_multi_endpoint,
+        general.me_adaptive_floor_recover_grace_secs,
+        general.me_adaptive_floor_writers_per_core_total,
+        general.me_adaptive_floor_cpu_cores_override,
+        general.me_adaptive_floor_max_extra_writers_single_per_core,
+        general.me_adaptive_floor_max_extra_writers_multi_per_core,
+        general.me_adaptive_floor_max_active_writers_per_core,
+        general.me_adaptive_floor_max_warm_writers_per_core,
+        general.me_adaptive_floor_max_active_writers_global,
+        general.me_adaptive_floor_max_warm_writers_global,
+        general.hardswap,
+        general.me_pool_drain_ttl_secs,
+        general.me_pool_drain_threshold,
+        general.me_pool_drain_soft_evict_enabled,
+        general.me_pool_drain_soft_evict_grace_secs,
+        general.me_pool_drain_soft_evict_per_writer,
+        general.me_pool_drain_soft_evict_budget_per_core,
+        general.me_pool_drain_soft_evict_cooldown_ms,
+        general.effective_me_pool_force_close_secs(),
+        general.me_pool_min_fresh_ratio,
+        general.me_hardswap_warmup_delay_min_ms,
+        general.me_hardswap_warmup_delay_max_ms,
+        general.me_hardswap_warmup_extra_passes,
+        general.me_hardswap_warmup_pass_backoff_base_ms,
+        general.me_bind_stale_mode,
+        general.me_bind_stale_ttl_secs,
+        general.me_secret_atomic_snapshot,
+        general.me_deterministic_writer_sort,
+        MeWriterPickMode::default(),
+        general.me_writer_pick_sample_size,
+        MeSocksKdfPolicy::default(),
+        general.me_writer_cmd_channel_capacity,
+        general.me_route_channel_capacity,
+        general.me_route_backpressure_base_timeout_ms,
+        general.me_route_backpressure_high_timeout_ms,
+        general.me_route_backpressure_high_watermark_pct,
+        general.me_reader_route_data_wait_ms,
+        general.me_health_interval_ms_unhealthy,
+        general.me_health_interval_ms_healthy,
+        general.me_warn_rate_limit_ms,
+        MeRouteNoWriterMode::default(),
+        general.me_route_no_writer_wait_ms,
+        general.me_route_inline_recovery_attempts,
+        general.me_route_inline_recovery_wait_ms,
+    );
+    (pool, rng)
+}
+
+async fn insert_draining_writer(
+    pool: &Arc<MePool>,
+    writer_id: u64,
+    drain_started_at_epoch_secs: u64,
+    bound_clients: usize,
+    drain_deadline_epoch_secs: u64,
+) {
+    let (tx, _writer_rx) = mpsc::channel::<WriterCommand>(8);
+    let writer = MeWriter {
+        id: writer_id,
+        addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 5500 + writer_id as u16),
+        source_ip: IpAddr::V4(Ipv4Addr::LOCALHOST),
+        writer_dc: 2,
+        generation: 1,
+        contour: Arc::new(AtomicU8::new(WriterContour::Draining.as_u8())),
+        created_at: Instant::now() - Duration::from_secs(writer_id),
+        tx: tx.clone(),
+        cancel: CancellationToken::new(),
+        degraded: Arc::new(AtomicBool::new(false)),
+        rtt_ema_ms_x10: Arc::new(AtomicU32::new(0)),
+        draining: Arc::new(AtomicBool::new(true)),
+        draining_started_at_epoch_secs: Arc::new(AtomicU64::new(drain_started_at_epoch_secs)),
+        drain_deadline_epoch_secs: Arc::new(AtomicU64::new(drain_deadline_epoch_secs)),
+        allow_drain_fallback: Arc::new(AtomicBool::new(false)),
+    };
+    pool.writers.write().await.push(writer);
+    pool.registry.register_writer(writer_id, tx).await;
+    pool.conn_count.fetch_add(1, Ordering::Relaxed);
+    for idx in 0..bound_clients {
+        let (conn_id, _rx) = pool.registry.register().await;
+        assert!(
+            pool.registry
+                .bind_writer(
+                    conn_id,
+                    writer_id,
+                    ConnMeta {
+                        target_dc: 2,
+                        client_addr: SocketAddr::new(
+                            IpAddr::V4(Ipv4Addr::LOCALHOST),
+                            7200 + idx as u16,
+                        ),
+                        our_addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443),
+                        proto_flags: 0,
+                    },
+                )
+                .await
+        );
+    }
+}
+
+#[tokio::test]
+async fn me_health_monitor_drains_expired_backlog_over_multiple_cycles() {
+    let (pool, rng) = make_pool(128, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let writer_total = health_drain_close_budget().saturating_mul(2).saturating_add(9);
+    for writer_id in 1..=writer_total as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(120),
+            1,
+            now_epoch_secs.saturating_sub(1),
+        )
+        .await;
+    }
+
+    let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
+    tokio::time::sleep(Duration::from_millis(60)).await;
+    monitor.abort();
+    let _ = monitor.await;
+
+    assert!(pool.writers.read().await.is_empty());
+}
+
+#[tokio::test]
+async fn me_health_monitor_cleans_empty_draining_writers_without_force_close() {
+    let (pool, rng) = make_pool(128, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    for writer_id in 1..=24u64 {
+        insert_draining_writer(&pool, writer_id, now_epoch_secs.saturating_sub(60), 0, 0).await;
+    }
+
+    let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
+    tokio::time::sleep(Duration::from_millis(30)).await;
+    monitor.abort();
+    let _ = monitor.await;
+
+    assert!(pool.writers.read().await.is_empty());
+}
+
+#[tokio::test]
+async fn me_health_monitor_converges_retry_like_threshold_backlog_to_empty() {
+    let threshold = 4u64;
+    let (pool, rng) = make_pool(threshold, 1, 1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let writer_total = threshold as usize + health_drain_close_budget().saturating_add(11);
+    for writer_id in 1..=writer_total as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(300).saturating_add(writer_id),
+            1,
+            0,
+        )
+        .await;
+    }
+
+    let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
+    tokio::time::sleep(Duration::from_millis(60)).await;
+    monitor.abort();
+    let _ = monitor.await;
+
+    assert!(pool.writers.read().await.is_empty());
+}

src/transport/middle_proxy/health_regression_tests.rs

@@ -0,0 +1,533 @@
+use std::collections::HashMap;
+use std::net::{IpAddr, Ipv4Addr, SocketAddr};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, AtomicU8, AtomicU32, AtomicU64, Ordering};
+use std::time::{Duration, Instant};
+
+use tokio::sync::mpsc;
+use tokio_util::sync::CancellationToken;
+
+use super::codec::WriterCommand;
+use super::health::{health_drain_close_budget, reap_draining_writers};
+use super::pool::{MePool, MeWriter, WriterContour};
+use super::registry::ConnMeta;
+use crate::config::{GeneralConfig, MeRouteNoWriterMode, MeSocksKdfPolicy, MeWriterPickMode};
+use crate::crypto::SecureRandom;
+use crate::network::probe::NetworkDecision;
+use crate::stats::Stats;
+
+async fn make_pool(me_pool_drain_threshold: u64) -> Arc<MePool> {
+    let general = GeneralConfig {
+        me_pool_drain_threshold,
+        ..GeneralConfig::default()
+    };
+
+    MePool::new(
+        None,
+        vec![1u8; 32],
+        None,
+        false,
+        None,
+        Vec::new(),
+        1,
+        None,
+        12,
+        1200,
+        HashMap::new(),
+        HashMap::new(),
+        None,
+        NetworkDecision::default(),
+        None,
+        Arc::new(SecureRandom::new()),
+        Arc::new(Stats::new()),
+        general.me_keepalive_enabled,
+        general.me_keepalive_interval_secs,
+        general.me_keepalive_jitter_secs,
+        general.me_keepalive_payload_random,
+        general.rpc_proxy_req_every,
+        general.me_warmup_stagger_enabled,
+        general.me_warmup_step_delay_ms,
+        general.me_warmup_step_jitter_ms,
+        general.me_reconnect_max_concurrent_per_dc,
+        general.me_reconnect_backoff_base_ms,
+        general.me_reconnect_backoff_cap_ms,
+        general.me_reconnect_fast_retry_count,
+        general.me_single_endpoint_shadow_writers,
+        general.me_single_endpoint_outage_mode_enabled,
+        general.me_single_endpoint_outage_disable_quarantine,
+        general.me_single_endpoint_outage_backoff_min_ms,
+        general.me_single_endpoint_outage_backoff_max_ms,
+        general.me_single_endpoint_shadow_rotate_every_secs,
+        general.me_floor_mode,
+        general.me_adaptive_floor_idle_secs,
+        general.me_adaptive_floor_min_writers_single_endpoint,
+        general.me_adaptive_floor_min_writers_multi_endpoint,
+        general.me_adaptive_floor_recover_grace_secs,
+        general.me_adaptive_floor_writers_per_core_total,
+        general.me_adaptive_floor_cpu_cores_override,
+        general.me_adaptive_floor_max_extra_writers_single_per_core,
+        general.me_adaptive_floor_max_extra_writers_multi_per_core,
+        general.me_adaptive_floor_max_active_writers_per_core,
+        general.me_adaptive_floor_max_warm_writers_per_core,
+        general.me_adaptive_floor_max_active_writers_global,
+        general.me_adaptive_floor_max_warm_writers_global,
+        general.hardswap,
+        general.me_pool_drain_ttl_secs,
+        general.me_pool_drain_threshold,
+        general.me_pool_drain_soft_evict_enabled,
+        general.me_pool_drain_soft_evict_grace_secs,
+        general.me_pool_drain_soft_evict_per_writer,
+        general.me_pool_drain_soft_evict_budget_per_core,
+        general.me_pool_drain_soft_evict_cooldown_ms,
+        general.effective_me_pool_force_close_secs(),
+        general.me_pool_min_fresh_ratio,
+        general.me_hardswap_warmup_delay_min_ms,
+        general.me_hardswap_warmup_delay_max_ms,
+        general.me_hardswap_warmup_extra_passes,
+        general.me_hardswap_warmup_pass_backoff_base_ms,
+        general.me_bind_stale_mode,
+        general.me_bind_stale_ttl_secs,
+        general.me_secret_atomic_snapshot,
+        general.me_deterministic_writer_sort,
+        MeWriterPickMode::default(),
+        general.me_writer_pick_sample_size,
+        MeSocksKdfPolicy::default(),
+        general.me_writer_cmd_channel_capacity,
+        general.me_route_channel_capacity,
+        general.me_route_backpressure_base_timeout_ms,
+        general.me_route_backpressure_high_timeout_ms,
+        general.me_route_backpressure_high_watermark_pct,
+        general.me_reader_route_data_wait_ms,
+        general.me_health_interval_ms_unhealthy,
+        general.me_health_interval_ms_healthy,
+        general.me_warn_rate_limit_ms,
+        MeRouteNoWriterMode::default(),
+        general.me_route_no_writer_wait_ms,
+        general.me_route_inline_recovery_attempts,
+        general.me_route_inline_recovery_wait_ms,
+    )
+}
+
+async fn insert_draining_writer(
+    pool: &Arc<MePool>,
+    writer_id: u64,
+    drain_started_at_epoch_secs: u64,
+    bound_clients: usize,
+    drain_deadline_epoch_secs: u64,
+) -> Vec<u64> {
+    let mut conn_ids = Vec::with_capacity(bound_clients);
+    let (tx, _writer_rx) = mpsc::channel::<WriterCommand>(8);
+    let writer = MeWriter {
+        id: writer_id,
+        addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 4500 + writer_id as u16),
+        source_ip: IpAddr::V4(Ipv4Addr::LOCALHOST),
+        writer_dc: 2,
+        generation: 1,
+        contour: Arc::new(AtomicU8::new(WriterContour::Draining.as_u8())),
+        created_at: Instant::now() - Duration::from_secs(writer_id),
+        tx: tx.clone(),
+        cancel: CancellationToken::new(),
+        degraded: Arc::new(AtomicBool::new(false)),
+        rtt_ema_ms_x10: Arc::new(AtomicU32::new(0)),
+        draining: Arc::new(AtomicBool::new(true)),
+        draining_started_at_epoch_secs: Arc::new(AtomicU64::new(drain_started_at_epoch_secs)),
+        drain_deadline_epoch_secs: Arc::new(AtomicU64::new(drain_deadline_epoch_secs)),
+        allow_drain_fallback: Arc::new(AtomicBool::new(false)),
+    };
+    pool.writers.write().await.push(writer);
+    pool.registry.register_writer(writer_id, tx).await;
+    pool.conn_count.fetch_add(1, Ordering::Relaxed);
+    for idx in 0..bound_clients {
+        let (conn_id, _rx) = pool.registry.register().await;
+        assert!(
+            pool.registry
+                .bind_writer(
+                    conn_id,
+                    writer_id,
+                    ConnMeta {
+                        target_dc: 2,
+                        client_addr: SocketAddr::new(
+                            IpAddr::V4(Ipv4Addr::LOCALHOST),
+                            6200 + idx as u16,
+                        ),
+                        our_addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443),
+                        proto_flags: 0,
+                    },
+                )
+                .await
+        );
+        conn_ids.push(conn_id);
+    }
+    conn_ids
+}
+
+async fn current_writer_ids(pool: &Arc<MePool>) -> Vec<u64> {
+    let mut writer_ids = pool
+        .writers
+        .read()
+        .await
+        .iter()
+        .map(|writer| writer.id)
+        .collect::<Vec<_>>();
+    writer_ids.sort_unstable();
+    writer_ids
+}
+
+#[tokio::test]
+async fn reap_draining_writers_drops_warn_state_for_removed_writer() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let conn_ids =
+        insert_draining_writer(&pool, 7, now_epoch_secs.saturating_sub(180), 1, 0).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+    assert!(warn_next_allowed.contains_key(&7));
+
+    let _ = pool.remove_writer_and_close_clients(7).await;
+    assert!(pool.registry.get_writer(conn_ids[0]).await.is_none());
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+    assert!(!warn_next_allowed.contains_key(&7));
+}
+
+#[tokio::test]
+async fn reap_draining_writers_removes_empty_draining_writers() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(&pool, 1, now_epoch_secs.saturating_sub(40), 0, 0).await;
+    insert_draining_writer(&pool, 2, now_epoch_secs.saturating_sub(30), 0, 0).await;
+    insert_draining_writer(&pool, 3, now_epoch_secs.saturating_sub(20), 1, 0).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert_eq!(current_writer_ids(&pool).await, vec![3]);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_overflow_closes_oldest_non_empty_writers() {
+    let pool = make_pool(2).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(&pool, 11, now_epoch_secs.saturating_sub(40), 1, 0).await;
+    insert_draining_writer(&pool, 22, now_epoch_secs.saturating_sub(30), 1, 0).await;
+    insert_draining_writer(&pool, 33, now_epoch_secs.saturating_sub(20), 1, 0).await;
+    insert_draining_writer(&pool, 44, now_epoch_secs.saturating_sub(10), 1, 0).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert_eq!(current_writer_ids(&pool).await, vec![33, 44]);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_deadline_force_close_applies_under_threshold() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(
+        &pool,
+        50,
+        now_epoch_secs.saturating_sub(15),
+        1,
+        now_epoch_secs.saturating_sub(1),
+    )
+    .await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert!(current_writer_ids(&pool).await.is_empty());
+}
+
+#[tokio::test]
+async fn reap_draining_writers_limits_closes_per_health_tick() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let close_budget = health_drain_close_budget();
+    let writer_total = close_budget.saturating_add(19);
+    for writer_id in 1..=writer_total as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(20),
+            1,
+            now_epoch_secs.saturating_sub(1),
+        )
+        .await;
+    }
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert_eq!(pool.writers.read().await.len(), writer_total - close_budget);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_backlog_drains_across_ticks() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let close_budget = health_drain_close_budget();
+    let writer_total = close_budget.saturating_mul(2).saturating_add(7);
+    for writer_id in 1..=writer_total as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(20),
+            1,
+            now_epoch_secs.saturating_sub(1),
+        )
+        .await;
+    }
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    for _ in 0..8 {
+        if pool.writers.read().await.is_empty() {
+            break;
+        }
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+    }
+
+    assert!(pool.writers.read().await.is_empty());
+}
+
+#[tokio::test]
+async fn reap_draining_writers_threshold_backlog_converges_to_threshold() {
+    let threshold = 5u64;
+    let pool = make_pool(threshold).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let close_budget = health_drain_close_budget();
+    let writer_total = threshold as usize + close_budget.saturating_add(12);
+    for writer_id in 1..=writer_total as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(200).saturating_add(writer_id),
+            1,
+            0,
+        )
+        .await;
+    }
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    for _ in 0..16 {
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+        if pool.writers.read().await.len() <= threshold as usize {
+            break;
+        }
+    }
+
+    assert_eq!(pool.writers.read().await.len(), threshold as usize);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_threshold_zero_preserves_non_expired_non_empty_writers() {
+    let pool = make_pool(0).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(&pool, 10, now_epoch_secs.saturating_sub(40), 1, 0).await;
+    insert_draining_writer(&pool, 20, now_epoch_secs.saturating_sub(30), 1, 0).await;
+    insert_draining_writer(&pool, 30, now_epoch_secs.saturating_sub(20), 1, 0).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert_eq!(current_writer_ids(&pool).await, vec![10, 20, 30]);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_prioritizes_force_close_before_empty_cleanup() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let close_budget = health_drain_close_budget();
+    for writer_id in 1..=close_budget as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(20),
+            1,
+            now_epoch_secs.saturating_sub(1),
+        )
+        .await;
+    }
+    let empty_writer_id = close_budget as u64 + 1;
+    insert_draining_writer(&pool, empty_writer_id, now_epoch_secs.saturating_sub(20), 0, 0).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert_eq!(current_writer_ids(&pool).await, vec![empty_writer_id]);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_empty_cleanup_does_not_increment_force_close_metric() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(&pool, 1, now_epoch_secs.saturating_sub(60), 0, 0).await;
+    insert_draining_writer(&pool, 2, now_epoch_secs.saturating_sub(50), 0, 0).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert!(current_writer_ids(&pool).await.is_empty());
+    assert_eq!(pool.stats.get_pool_force_close_total(), 0);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_handles_duplicate_force_close_requests_for_same_writer() {
+    let pool = make_pool(1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(
+        &pool,
+        10,
+        now_epoch_secs.saturating_sub(30),
+        1,
+        now_epoch_secs.saturating_sub(1),
+    )
+    .await;
+    insert_draining_writer(
+        &pool,
+        20,
+        now_epoch_secs.saturating_sub(20),
+        1,
+        now_epoch_secs.saturating_sub(1),
+    )
+    .await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    assert!(current_writer_ids(&pool).await.is_empty());
+}
+
+#[tokio::test]
+async fn reap_draining_writers_warn_state_never_exceeds_live_draining_population_under_churn() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    for wave in 0..12u64 {
+        for offset in 0..9u64 {
+            insert_draining_writer(
+                &pool,
+                wave * 100 + offset,
+                now_epoch_secs.saturating_sub(120 + offset),
+                1,
+                0,
+            )
+            .await;
+        }
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+        assert!(warn_next_allowed.len() <= pool.writers.read().await.len());
+
+        let existing_writer_ids = current_writer_ids(&pool).await;
+        for writer_id in existing_writer_ids.into_iter().take(4) {
+            let _ = pool.remove_writer_and_close_clients(writer_id).await;
+        }
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+        assert!(warn_next_allowed.len() <= pool.writers.read().await.len());
+    }
+}
+
+#[tokio::test]
+async fn reap_draining_writers_mixed_backlog_converges_without_leaking_warn_state() {
+    let pool = make_pool(6).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    for writer_id in 1..=18u64 {
+        let bound_clients = if writer_id % 3 == 0 { 0 } else { 1 };
+        let deadline = if writer_id % 2 == 0 {
+            now_epoch_secs.saturating_sub(1)
+        } else {
+            0
+        };
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(300).saturating_add(writer_id),
+            bound_clients,
+            deadline,
+        )
+        .await;
+    }
+
+    for _ in 0..16 {
+        reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+        if pool.writers.read().await.len() <= 6 {
+            break;
+        }
+    }
+
+    assert!(pool.writers.read().await.len() <= 6);
+    assert!(warn_next_allowed.len() <= pool.writers.read().await.len());
+}
+
+#[tokio::test]
+async fn reap_draining_writers_soft_evicts_stuck_writer_with_per_writer_cap() {
+    let pool = make_pool(128).await;
+    pool.me_pool_drain_soft_evict_enabled.store(true, Ordering::Relaxed);
+    pool.me_pool_drain_soft_evict_grace_secs.store(0, Ordering::Relaxed);
+    pool.me_pool_drain_soft_evict_per_writer.store(1, Ordering::Relaxed);
+    pool.me_pool_drain_soft_evict_budget_per_core.store(8, Ordering::Relaxed);
+    pool.me_pool_drain_soft_evict_cooldown_ms
+        .store(1, Ordering::Relaxed);
+
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(&pool, 77, now_epoch_secs.saturating_sub(240), 3, 0).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    let activity = pool.registry.writer_activity_snapshot().await;
+    assert_eq!(activity.bound_clients_by_writer.get(&77), Some(&2));
+    assert_eq!(pool.stats.get_pool_drain_soft_evict_total(), 1);
+    assert_eq!(pool.stats.get_pool_drain_soft_evict_writer_total(), 1);
+    assert_eq!(current_writer_ids(&pool).await, vec![77]);
+}
+
+#[tokio::test]
+async fn reap_draining_writers_soft_evict_respects_cooldown_per_writer() {
+    let pool = make_pool(128).await;
+    pool.me_pool_drain_soft_evict_enabled.store(true, Ordering::Relaxed);
+    pool.me_pool_drain_soft_evict_grace_secs.store(0, Ordering::Relaxed);
+    pool.me_pool_drain_soft_evict_per_writer.store(1, Ordering::Relaxed);
+    pool.me_pool_drain_soft_evict_budget_per_core.store(8, Ordering::Relaxed);
+    pool.me_pool_drain_soft_evict_cooldown_ms
+        .store(60_000, Ordering::Relaxed);
+
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(&pool, 88, now_epoch_secs.saturating_sub(240), 3, 0).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut soft_evict_next_allowed = HashMap::new();
+
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+    reap_draining_writers(&pool, &mut warn_next_allowed, &mut soft_evict_next_allowed).await;
+
+    let activity = pool.registry.writer_activity_snapshot().await;
+    assert_eq!(activity.bound_clients_by_writer.get(&88), Some(&2));
+    assert_eq!(pool.stats.get_pool_drain_soft_evict_total(), 1);
+    assert_eq!(pool.stats.get_pool_drain_soft_evict_writer_total(), 1);
+}
+
+#[test]
+fn general_config_default_drain_threshold_remains_enabled() {
+    assert_eq!(GeneralConfig::default().me_pool_drain_threshold, 128);
+    assert!(GeneralConfig::default().me_pool_drain_soft_evict_enabled);
+    assert_eq!(
+        GeneralConfig::default().me_pool_drain_soft_evict_per_writer,
+        1
+    );
+}

src/transport/middle_proxy/mod.rs

@@ -21,6 +21,12 @@ mod secret;
 mod selftest;
 mod wire;
 mod pool_status;
+#[cfg(test)]
+mod health_regression_tests;
+#[cfg(test)]
+mod health_integration_tests;
+#[cfg(test)]
+mod health_adversarial_tests;
 
 use bytes::Bytes;
 

src/transport/middle_proxy/pool.rs

@@ -172,6 +172,11 @@ pub struct MePool {
     pub(super) kdf_material_fingerprint: Arc<RwLock<HashMap<SocketAddr, (u64, u16)>>>,
     pub(super) me_pool_drain_ttl_secs: AtomicU64,
     pub(super) me_pool_drain_threshold: AtomicU64,
+    pub(super) me_pool_drain_soft_evict_enabled: AtomicBool,
+    pub(super) me_pool_drain_soft_evict_grace_secs: AtomicU64,
+    pub(super) me_pool_drain_soft_evict_per_writer: AtomicU8,
+    pub(super) me_pool_drain_soft_evict_budget_per_core: AtomicU32,
+    pub(super) me_pool_drain_soft_evict_cooldown_ms: AtomicU64,
     pub(super) me_pool_force_close_secs: AtomicU64,
     pub(super) me_pool_min_fresh_ratio_permille: AtomicU32,
     pub(super) me_hardswap_warmup_delay_min_ms: AtomicU64,
@@ -273,6 +278,11 @@ impl MePool {
         hardswap: bool,
         me_pool_drain_ttl_secs: u64,
         me_pool_drain_threshold: u64,
+        me_pool_drain_soft_evict_enabled: bool,
+        me_pool_drain_soft_evict_grace_secs: u64,
+        me_pool_drain_soft_evict_per_writer: u8,
+        me_pool_drain_soft_evict_budget_per_core: u16,
+        me_pool_drain_soft_evict_cooldown_ms: u64,
         me_pool_force_close_secs: u64,
         me_pool_min_fresh_ratio: f32,
         me_hardswap_warmup_delay_min_ms: u64,
@@ -449,6 +459,17 @@ impl MePool {
             kdf_material_fingerprint: Arc::new(RwLock::new(HashMap::new())),
             me_pool_drain_ttl_secs: AtomicU64::new(me_pool_drain_ttl_secs),
             me_pool_drain_threshold: AtomicU64::new(me_pool_drain_threshold),
+            me_pool_drain_soft_evict_enabled: AtomicBool::new(me_pool_drain_soft_evict_enabled),
+            me_pool_drain_soft_evict_grace_secs: AtomicU64::new(me_pool_drain_soft_evict_grace_secs),
+            me_pool_drain_soft_evict_per_writer: AtomicU8::new(
+                me_pool_drain_soft_evict_per_writer.max(1),
+            ),
+            me_pool_drain_soft_evict_budget_per_core: AtomicU32::new(
+                me_pool_drain_soft_evict_budget_per_core.max(1) as u32,
+            ),
+            me_pool_drain_soft_evict_cooldown_ms: AtomicU64::new(
+                me_pool_drain_soft_evict_cooldown_ms.max(1),
+            ),
             me_pool_force_close_secs: AtomicU64::new(me_pool_force_close_secs),
             me_pool_min_fresh_ratio_permille: AtomicU32::new(Self::ratio_to_permille(
                 me_pool_min_fresh_ratio,
@@ -496,6 +517,11 @@ impl MePool {
         hardswap: bool,
         drain_ttl_secs: u64,
         pool_drain_threshold: u64,
+        pool_drain_soft_evict_enabled: bool,
+        pool_drain_soft_evict_grace_secs: u64,
+        pool_drain_soft_evict_per_writer: u8,
+        pool_drain_soft_evict_budget_per_core: u16,
+        pool_drain_soft_evict_cooldown_ms: u64,
         force_close_secs: u64,
         min_fresh_ratio: f32,
         hardswap_warmup_delay_min_ms: u64,
@@ -536,6 +562,18 @@ impl MePool {
             .store(drain_ttl_secs, Ordering::Relaxed);
         self.me_pool_drain_threshold
             .store(pool_drain_threshold, Ordering::Relaxed);
+        self.me_pool_drain_soft_evict_enabled
+            .store(pool_drain_soft_evict_enabled, Ordering::Relaxed);
+        self.me_pool_drain_soft_evict_grace_secs
+            .store(pool_drain_soft_evict_grace_secs, Ordering::Relaxed);
+        self.me_pool_drain_soft_evict_per_writer
+            .store(pool_drain_soft_evict_per_writer.max(1), Ordering::Relaxed);
+        self.me_pool_drain_soft_evict_budget_per_core.store(
+            pool_drain_soft_evict_budget_per_core.max(1) as u32,
+            Ordering::Relaxed,
+        );
+        self.me_pool_drain_soft_evict_cooldown_ms
+            .store(pool_drain_soft_evict_cooldown_ms.max(1), Ordering::Relaxed);
         self.me_pool_force_close_secs
             .store(force_close_secs, Ordering::Relaxed);
         self.me_pool_min_fresh_ratio_permille
@@ -690,6 +728,36 @@ impl MePool {
         }
     }
 
+    pub(super) fn drain_soft_evict_enabled(&self) -> bool {
+        self.me_pool_drain_soft_evict_enabled
+            .load(Ordering::Relaxed)
+    }
+
+    pub(super) fn drain_soft_evict_grace_secs(&self) -> u64 {
+        self.me_pool_drain_soft_evict_grace_secs
+            .load(Ordering::Relaxed)
+    }
+
+    pub(super) fn drain_soft_evict_per_writer(&self) -> usize {
+        self.me_pool_drain_soft_evict_per_writer
+            .load(Ordering::Relaxed)
+            .max(1) as usize
+    }
+
+    pub(super) fn drain_soft_evict_budget_per_core(&self) -> usize {
+        self.me_pool_drain_soft_evict_budget_per_core
+            .load(Ordering::Relaxed)
+            .max(1) as usize
+    }
+
+    pub(super) fn drain_soft_evict_cooldown(&self) -> Duration {
+        Duration::from_millis(
+            self.me_pool_drain_soft_evict_cooldown_ms
+                .load(Ordering::Relaxed)
+                .max(1),
+        )
+    }
+
     pub(super) async fn key_selector(&self) -> u32 {
         self.proxy_secret.read().await.key_selector
     }

src/transport/middle_proxy/pool_reinit.rs

@@ -70,10 +70,12 @@ impl MePool {
 
         let mut missing_dc = Vec::<i32>::new();
         let mut covered = 0usize;
+        let mut total = 0usize;
         for (dc, endpoints) in desired_by_dc {
             if endpoints.is_empty() {
                 continue;
             }
+            total += 1;
             if endpoints
                 .iter()
                 .any(|addr| active_writer_addrs.contains(&(*dc, *addr)))
@@ -85,7 +87,9 @@ impl MePool {
         }
 
         missing_dc.sort_unstable();
-        let total = desired_by_dc.len().max(1);
+        if total == 0 {
+            return (1.0, missing_dc);
+        }
         let ratio = (covered as f32) / (total as f32);
         (ratio, missing_dc)
     }
@@ -399,29 +403,21 @@ impl MePool {
         }
 
         if hardswap {
-            let mut fresh_missing_dc = Vec::<(i32, usize, usize)>::new();
-            for (dc, endpoints) in &desired_by_dc {
-                if endpoints.is_empty() {
-                    continue;
-                }
-                let required = self.required_writers_for_dc(endpoints.len());
-                let fresh_count = writers
-                    .iter()
-                    .filter(|w| !w.draining.load(Ordering::Relaxed))
-                    .filter(|w| w.generation == generation)
-                    .filter(|w| w.writer_dc == *dc)
-                    .filter(|w| endpoints.contains(&w.addr))
-                    .count();
-                if fresh_count < required {
-                    fresh_missing_dc.push((*dc, fresh_count, required));
-                }
-            }
+            let fresh_writer_addrs: HashSet<(i32, SocketAddr)> = writers
+                .iter()
+                .filter(|w| !w.draining.load(Ordering::Relaxed))
+                .filter(|w| w.generation == generation)
+                .map(|w| (w.writer_dc, w.addr))
+                .collect();
+            let (fresh_coverage_ratio, fresh_missing_dc) =
+                Self::coverage_ratio(&desired_by_dc, &fresh_writer_addrs);
             if !fresh_missing_dc.is_empty() {
                 warn!(
                     previous_generation,
                     generation,
+                    fresh_coverage_ratio = format_args!("{fresh_coverage_ratio:.3}"),
                     missing_dc = ?fresh_missing_dc,
-                    "ME hardswap pending: fresh generation coverage incomplete"
+                    "ME hardswap pending: fresh generation DC coverage incomplete"
                 );
                 return;
             }
@@ -491,3 +487,61 @@ impl MePool {
         self.zero_downtime_reinit_after_map_change(rng).await;
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::collections::{HashMap, HashSet};
+    use std::net::{IpAddr, Ipv4Addr, SocketAddr};
+
+    use super::MePool;
+
+    fn addr(octet: u8, port: u16) -> SocketAddr {
+        SocketAddr::new(IpAddr::V4(Ipv4Addr::new(127, 0, 0, octet)), port)
+    }
+
+    #[test]
+    fn coverage_ratio_counts_dc_coverage_not_floor() {
+        let dc1 = addr(1, 2001);
+        let dc2 = addr(2, 2002);
+
+        let mut desired_by_dc = HashMap::<i32, HashSet<SocketAddr>>::new();
+        desired_by_dc.insert(1, HashSet::from([dc1]));
+        desired_by_dc.insert(2, HashSet::from([dc2]));
+
+        let active_writer_addrs = HashSet::from([(1, dc1)]);
+        let (ratio, missing_dc) = MePool::coverage_ratio(&desired_by_dc, &active_writer_addrs);
+
+        assert_eq!(ratio, 0.5);
+        assert_eq!(missing_dc, vec![2]);
+    }
+
+    #[test]
+    fn coverage_ratio_ignores_empty_dc_groups() {
+        let dc1 = addr(1, 2001);
+
+        let mut desired_by_dc = HashMap::<i32, HashSet<SocketAddr>>::new();
+        desired_by_dc.insert(1, HashSet::from([dc1]));
+        desired_by_dc.insert(2, HashSet::new());
+
+        let active_writer_addrs = HashSet::from([(1, dc1)]);
+        let (ratio, missing_dc) = MePool::coverage_ratio(&desired_by_dc, &active_writer_addrs);
+
+        assert_eq!(ratio, 1.0);
+        assert!(missing_dc.is_empty());
+    }
+
+    #[test]
+    fn coverage_ratio_reports_missing_dcs_sorted() {
+        let dc1 = addr(1, 2001);
+        let dc2 = addr(2, 2002);
+
+        let mut desired_by_dc = HashMap::<i32, HashSet<SocketAddr>>::new();
+        desired_by_dc.insert(2, HashSet::from([dc2]));
+        desired_by_dc.insert(1, HashSet::from([dc1]));
+
+        let (ratio, missing_dc) = MePool::coverage_ratio(&desired_by_dc, &HashSet::new());
+
+        assert_eq!(ratio, 0.0);
+        assert_eq!(missing_dc, vec![1, 2]);
+    }
+}

src/transport/middle_proxy/pool_status.rs

@@ -40,6 +40,7 @@ pub(crate) struct MeApiDcStatusSnapshot {
     pub floor_max: usize,
     pub floor_capped: bool,
     pub alive_writers: usize,
+    pub coverage_ratio: f64,
     pub coverage_pct: f64,
     pub fresh_alive_writers: usize,
     pub fresh_coverage_pct: f64,
@@ -62,6 +63,7 @@ pub(crate) struct MeApiStatusSnapshot {
     pub available_pct: f64,
     pub required_writers: usize,
     pub alive_writers: usize,
+    pub coverage_ratio: f64,
     pub coverage_pct: f64,
     pub fresh_alive_writers: usize,
     pub fresh_coverage_pct: f64,
@@ -124,6 +126,11 @@ pub(crate) struct MeApiRuntimeSnapshot {
     pub me_reconnect_backoff_cap_ms: u64,
     pub me_reconnect_fast_retry_count: u32,
     pub me_pool_drain_ttl_secs: u64,
+    pub me_pool_drain_soft_evict_enabled: bool,
+    pub me_pool_drain_soft_evict_grace_secs: u64,
+    pub me_pool_drain_soft_evict_per_writer: u8,
+    pub me_pool_drain_soft_evict_budget_per_core: u16,
+    pub me_pool_drain_soft_evict_cooldown_ms: u64,
     pub me_pool_force_close_secs: u64,
     pub me_pool_min_fresh_ratio: f32,
     pub me_bind_stale_mode: &'static str,
@@ -337,6 +344,8 @@ impl MePool {
         let mut available_endpoints = 0usize;
         let mut alive_writers = 0usize;
         let mut fresh_alive_writers = 0usize;
+        let mut coverage_ratio_dcs_total = 0usize;
+        let mut coverage_ratio_dcs_covered = 0usize;
         let floor_mode = self.floor_mode();
         let adaptive_cpu_cores = (self
             .me_adaptive_floor_cpu_cores_effective
@@ -388,6 +397,12 @@ impl MePool {
             available_endpoints += dc_available_endpoints;
             alive_writers += dc_alive_writers;
             fresh_alive_writers += dc_fresh_alive_writers;
+            if endpoint_count > 0 {
+                coverage_ratio_dcs_total += 1;
+                if dc_alive_writers > 0 {
+                    coverage_ratio_dcs_covered += 1;
+                }
+            }
 
             dcs.push(MeApiDcStatusSnapshot {
                 dc,
@@ -410,6 +425,11 @@ impl MePool {
                 floor_max,
                 floor_capped,
                 alive_writers: dc_alive_writers,
+                coverage_ratio: if endpoint_count > 0 && dc_alive_writers > 0 {
+                    100.0
+                } else {
+                    0.0
+                },
                 coverage_pct: ratio_pct(dc_alive_writers, dc_required_writers),
                 fresh_alive_writers: dc_fresh_alive_writers,
                 fresh_coverage_pct: ratio_pct(dc_fresh_alive_writers, dc_required_writers),
@@ -426,6 +446,7 @@ impl MePool {
             available_pct: ratio_pct(available_endpoints, configured_endpoints),
             required_writers,
             alive_writers,
+            coverage_ratio: ratio_pct(coverage_ratio_dcs_covered, coverage_ratio_dcs_total),
             coverage_pct: ratio_pct(alive_writers, required_writers),
             fresh_alive_writers,
             fresh_coverage_pct: ratio_pct(fresh_alive_writers, required_writers),
@@ -562,6 +583,22 @@ impl MePool {
             me_reconnect_backoff_cap_ms: self.me_reconnect_backoff_cap.as_millis() as u64,
             me_reconnect_fast_retry_count: self.me_reconnect_fast_retry_count,
             me_pool_drain_ttl_secs: self.me_pool_drain_ttl_secs.load(Ordering::Relaxed),
+            me_pool_drain_soft_evict_enabled: self
+                .me_pool_drain_soft_evict_enabled
+                .load(Ordering::Relaxed),
+            me_pool_drain_soft_evict_grace_secs: self
+                .me_pool_drain_soft_evict_grace_secs
+                .load(Ordering::Relaxed),
+            me_pool_drain_soft_evict_per_writer: self
+                .me_pool_drain_soft_evict_per_writer
+                .load(Ordering::Relaxed),
+            me_pool_drain_soft_evict_budget_per_core: self
+                .me_pool_drain_soft_evict_budget_per_core
+                .load(Ordering::Relaxed)
+                .min(u16::MAX as u32) as u16,
+            me_pool_drain_soft_evict_cooldown_ms: self
+                .me_pool_drain_soft_evict_cooldown_ms
+                .load(Ordering::Relaxed),
             me_pool_force_close_secs: self.me_pool_force_close_secs.load(Ordering::Relaxed),
             me_pool_min_fresh_ratio: Self::permille_to_ratio(
                 self.me_pool_min_fresh_ratio_permille.load(Ordering::Relaxed),

src/transport/middle_proxy/registry.rs

@@ -394,6 +394,56 @@ impl ConnRegistry {
         inner.writer_for_conn.keys().copied().collect()
     }
 
+    pub(super) async fn bound_conn_ids_for_writer_limited(
+        &self,
+        writer_id: u64,
+        limit: usize,
+    ) -> Vec<u64> {
+        if limit == 0 {
+            return Vec::new();
+        }
+        let inner = self.inner.read().await;
+        let Some(conn_ids) = inner.conns_for_writer.get(&writer_id) else {
+            return Vec::new();
+        };
+        let mut out = conn_ids.iter().copied().collect::<Vec<_>>();
+        out.sort_unstable();
+        out.truncate(limit);
+        out
+    }
+
+    pub(super) async fn evict_bound_conn_if_writer(&self, conn_id: u64, writer_id: u64) -> bool {
+        let maybe_client_tx = {
+            let mut inner = self.inner.write().await;
+            if inner.writer_for_conn.get(&conn_id).copied() != Some(writer_id) {
+                return false;
+            }
+
+            let client_tx = inner.map.get(&conn_id).cloned();
+            inner.map.remove(&conn_id);
+            inner.meta.remove(&conn_id);
+            inner.writer_for_conn.remove(&conn_id);
+
+            let became_empty = if let Some(set) = inner.conns_for_writer.get_mut(&writer_id) {
+                set.remove(&conn_id);
+                set.is_empty()
+            } else {
+                false
+            };
+            if became_empty {
+                inner
+                    .writer_idle_since_epoch_secs
+                    .insert(writer_id, Self::now_epoch_secs());
+            }
+            client_tx
+        };
+
+        if let Some(client_tx) = maybe_client_tx {
+            let _ = client_tx.try_send(MeResponse::Close);
+        }
+        true
+    }
+
     pub async fn writer_lost(&self, writer_id: u64) -> Vec<BoundConn> {
         let mut inner = self.inner.write().await;
         inner.writers.remove(&writer_id);
@@ -444,6 +494,7 @@ mod tests {
 
     use super::ConnMeta;
     use super::ConnRegistry;
+    use super::MeResponse;
 
     #[tokio::test]
     async fn writer_activity_snapshot_tracks_writer_and_dc_load() {
@@ -634,4 +685,86 @@ mod tests {
         );
         assert!(registry.get_writer(conn_id).await.is_none());
     }
+
+    #[tokio::test]
+    async fn bound_conn_ids_for_writer_limited_is_sorted_and_bounded() {
+        let registry = ConnRegistry::new();
+        let (writer_tx, _writer_rx) = tokio::sync::mpsc::channel(8);
+        registry.register_writer(10, writer_tx).await;
+        let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
+        let mut conn_ids = Vec::new();
+        for _ in 0..5 {
+            let (conn_id, _rx) = registry.register().await;
+            assert!(
+                registry
+                    .bind_writer(
+                        conn_id,
+                        10,
+                        ConnMeta {
+                            target_dc: 2,
+                            client_addr: addr,
+                            our_addr: addr,
+                            proto_flags: 0,
+                        },
+                    )
+                    .await
+            );
+            conn_ids.push(conn_id);
+        }
+        conn_ids.sort_unstable();
+
+        let limited = registry.bound_conn_ids_for_writer_limited(10, 3).await;
+        assert_eq!(limited.len(), 3);
+        assert_eq!(limited, conn_ids.into_iter().take(3).collect::<Vec<_>>());
+    }
+
+    #[tokio::test]
+    async fn evict_bound_conn_if_writer_does_not_touch_rebound_conn() {
+        let registry = ConnRegistry::new();
+        let (conn_id, mut 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 evicted = registry.evict_bound_conn_if_writer(conn_id, 10).await;
+        assert!(!evicted);
+        assert_eq!(registry.get_writer(conn_id).await.expect("writer").writer_id, 20);
+        assert!(rx.try_recv().is_err());
+
+        let evicted = registry.evict_bound_conn_if_writer(conn_id, 20).await;
+        assert!(evicted);
+        assert!(registry.get_writer(conn_id).await.is_none());
+        assert!(matches!(rx.try_recv(), Ok(MeResponse::Close)));
+    }
 }

src/transport/socket.rs

@@ -11,6 +11,8 @@ use tokio::net::TcpStream;
 use socket2::{Socket, TcpKeepalive, Domain, Type, Protocol};
 use tracing::debug;
 
+const DEFAULT_SOCKET_BUFFER_BYTES: usize = 256 * 1024;
+
 /// Configure TCP socket with recommended settings for proxy use
 #[allow(dead_code)]
 pub fn configure_tcp_socket(
@@ -34,10 +36,10 @@ pub fn configure_tcp_socket(
         
         socket.set_tcp_keepalive(&keepalive)?;
     }
-    
-    // CHANGED: Removed manual buffer size setting (was 256KB).
-    // Allowing the OS kernel to handle TCP window scaling (Autotuning) is critical
-    // for mobile clients to avoid bufferbloat and stalled connections during uploads.
+
+    // Use explicit baseline buffers to reduce slow-start stalls on high RTT links.
+    socket.set_recv_buffer_size(DEFAULT_SOCKET_BUFFER_BYTES)?;
+    socket.set_send_buffer_size(DEFAULT_SOCKET_BUFFER_BYTES)?;
     
     Ok(())
 }
@@ -62,6 +64,10 @@ pub fn configure_client_socket(
     let keepalive = keepalive.with_interval(Duration::from_secs(keepalive_secs));
     
     socket.set_tcp_keepalive(&keepalive)?;
+
+    // Keep explicit baseline buffers for predictable throughput across busy hosts.
+    socket.set_recv_buffer_size(DEFAULT_SOCKET_BUFFER_BYTES)?;
+    socket.set_send_buffer_size(DEFAULT_SOCKET_BUFFER_BYTES)?;
     
     // Set TCP user timeout (Linux only)
     // NOTE: iOS does not support TCP_USER_TIMEOUT - application-level timeout 
@@ -124,6 +130,8 @@ pub fn create_outgoing_socket_bound(addr: SocketAddr, bind_addr: Option<IpAddr>)
     
     // Disable Nagle
     socket.set_nodelay(true)?;
+    socket.set_recv_buffer_size(DEFAULT_SOCKET_BUFFER_BYTES)?;
+    socket.set_send_buffer_size(DEFAULT_SOCKET_BUFFER_BYTES)?;
 
     if let Some(bind_ip) = bind_addr {
         let bind_sock_addr = SocketAddr::new(bind_ip, 0);

tools/telemt_api.py

@@ -0,0 +1,728 @@
+"""
+Telemt Control API Python Client
+Full-coverage client for https://github.com/telemt/telemt
+
+Usage:
+    client = TelemtAPI("http://127.0.0.1:9091", auth_header="your-secret")
+    client.health()
+    client.create_user("alice", max_tcp_conns=10)
+    client.patch_user("alice", data_quota_bytes=1_000_000_000)
+    client.delete_user("alice")
+"""
+
+from __future__ import annotations
+
+import json
+import secrets
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Union
+from urllib.error import HTTPError, URLError
+from urllib.request import Request, urlopen
+
+
+# ---------------------------------------------------------------------------
+# Exceptions
+# ---------------------------------------------------------------------------
+
+class TememtAPIError(Exception):
+    """Raised when the API returns an error envelope or a transport error."""
+
+    def __init__(self, message: str, code: str | None = None,
+                 http_status: int | None = None, request_id: int | None = None):
+        super().__init__(message)
+        self.code = code
+        self.http_status = http_status
+        self.request_id = request_id
+
+    def __repr__(self) -> str:
+        return (f"TememtAPIError(message={str(self)!r}, code={self.code!r}, "
+                f"http_status={self.http_status}, request_id={self.request_id})")
+
+
+# ---------------------------------------------------------------------------
+# Response wrapper
+# ---------------------------------------------------------------------------
+
+@dataclass
+class APIResponse:
+    """Wraps a successful API response envelope."""
+    ok: bool
+    data: Any
+    revision: str | None = None
+
+    def __repr__(self) -> str:  # pragma: no cover
+        return f"APIResponse(ok={self.ok}, revision={self.revision!r}, data={self.data!r})"
+
+
+# ---------------------------------------------------------------------------
+# Main client
+# ---------------------------------------------------------------------------
+
+class TememtAPI:
+    """
+    HTTP client for the Telemt Control API.
+
+    Parameters
+    ----------
+    base_url:
+        Scheme + host + port, e.g. ``"http://127.0.0.1:9091"``.
+        Trailing slash is stripped automatically.
+    auth_header:
+        Exact value for the ``Authorization`` header.
+        Leave *None* when ``auth_header`` is not configured server-side.
+    timeout:
+        Socket timeout in seconds for every request (default 10).
+    """
+
+    def __init__(
+        self,
+        base_url: str = "http://127.0.0.1:9091",
+        auth_header: str | None = None,
+        timeout: int = 10,
+    ) -> None:
+        self.base_url = base_url.rstrip("/")
+        self.auth_header = auth_header
+        self.timeout = timeout
+
+    # ------------------------------------------------------------------
+    # Low-level HTTP helpers
+    # ------------------------------------------------------------------
+
+    def _headers(self, extra: dict | None = None) -> dict:
+        h = {"Content-Type": "application/json; charset=utf-8",
+             "Accept": "application/json"}
+        if self.auth_header:
+            h["Authorization"] = self.auth_header
+        if extra:
+            h.update(extra)
+        return h
+
+    def _request(
+        self,
+        method: str,
+        path: str,
+        body: dict | None = None,
+        if_match: str | None = None,
+        query: dict | None = None,
+    ) -> APIResponse:
+        url = self.base_url + path
+        if query:
+            qs = "&".join(f"{k}={v}" for k, v in query.items())
+            url = f"{url}?{qs}"
+
+        raw_body: bytes | None = None
+        if body is not None:
+            raw_body = json.dumps(body).encode()
+
+        extra_headers: dict = {}
+        if if_match is not None:
+            extra_headers["If-Match"] = if_match
+
+        req = Request(
+            url,
+            data=raw_body,
+            headers=self._headers(extra_headers),
+            method=method,
+        )
+
+        try:
+            with urlopen(req, timeout=self.timeout) as resp:
+                payload = json.loads(resp.read())
+        except HTTPError as exc:
+            raw = exc.read()
+            try:
+                payload = json.loads(raw)
+            except Exception:
+                raise TememtAPIError(
+                    str(exc), http_status=exc.code
+                ) from exc
+            err = payload.get("error", {})
+            raise TememtAPIError(
+                err.get("message", str(exc)),
+                code=err.get("code"),
+                http_status=exc.code,
+                request_id=payload.get("request_id"),
+            ) from exc
+        except URLError as exc:
+            raise TememtAPIError(str(exc)) from exc
+
+        if not payload.get("ok"):
+            err = payload.get("error", {})
+            raise TememtAPIError(
+                err.get("message", "unknown error"),
+                code=err.get("code"),
+                request_id=payload.get("request_id"),
+            )
+
+        return APIResponse(
+            ok=True,
+            data=payload.get("data"),
+            revision=payload.get("revision"),
+        )
+
+    def _get(self, path: str, query: dict | None = None) -> APIResponse:
+        return self._request("GET", path, query=query)
+
+    def _post(self, path: str, body: dict | None = None,
+              if_match: str | None = None) -> APIResponse:
+        return self._request("POST", path, body=body, if_match=if_match)
+
+    def _patch(self, path: str, body: dict,
+               if_match: str | None = None) -> APIResponse:
+        return self._request("PATCH", path, body=body, if_match=if_match)
+
+    def _delete(self, path: str, if_match: str | None = None) -> APIResponse:
+        return self._request("DELETE", path, if_match=if_match)
+
+    # ------------------------------------------------------------------
+    # Health & system
+    # ------------------------------------------------------------------
+
+    def health(self) -> APIResponse:
+        """GET /v1/health — liveness probe."""
+        return self._get("/v1/health")
+
+    def system_info(self) -> APIResponse:
+        """GET /v1/system/info — binary version, uptime, config hash."""
+        return self._get("/v1/system/info")
+
+    # ------------------------------------------------------------------
+    # Runtime gates & initialization
+    # ------------------------------------------------------------------
+
+    def runtime_gates(self) -> APIResponse:
+        """GET /v1/runtime/gates — admission gates and startup progress."""
+        return self._get("/v1/runtime/gates")
+
+    def runtime_initialization(self) -> APIResponse:
+        """GET /v1/runtime/initialization — detailed startup timeline."""
+        return self._get("/v1/runtime/initialization")
+
+    # ------------------------------------------------------------------
+    # Limits & security
+    # ------------------------------------------------------------------
+
+    def limits_effective(self) -> APIResponse:
+        """GET /v1/limits/effective — effective timeout/upstream/ME limits."""
+        return self._get("/v1/limits/effective")
+
+    def security_posture(self) -> APIResponse:
+        """GET /v1/security/posture — API auth, telemetry, log-level summary."""
+        return self._get("/v1/security/posture")
+
+    def security_whitelist(self) -> APIResponse:
+        """GET /v1/security/whitelist — current IP whitelist CIDRs."""
+        return self._get("/v1/security/whitelist")
+
+    # ------------------------------------------------------------------
+    # Stats
+    # ------------------------------------------------------------------
+
+    def stats_summary(self) -> APIResponse:
+        """GET /v1/stats/summary — uptime, connection totals, user count."""
+        return self._get("/v1/stats/summary")
+
+    def stats_zero_all(self) -> APIResponse:
+        """GET /v1/stats/zero/all — zero-cost counters (core, upstream, ME, pool, desync)."""
+        return self._get("/v1/stats/zero/all")
+
+    def stats_upstreams(self) -> APIResponse:
+        """GET /v1/stats/upstreams — upstream health + zero counters."""
+        return self._get("/v1/stats/upstreams")
+
+    def stats_minimal_all(self) -> APIResponse:
+        """GET /v1/stats/minimal/all — ME writers + DC snapshot (requires minimal_runtime_enabled)."""
+        return self._get("/v1/stats/minimal/all")
+
+    def stats_me_writers(self) -> APIResponse:
+        """GET /v1/stats/me-writers — per-writer ME status (requires minimal_runtime_enabled)."""
+        return self._get("/v1/stats/me-writers")
+
+    def stats_dcs(self) -> APIResponse:
+        """GET /v1/stats/dcs — per-DC coverage and writer counts (requires minimal_runtime_enabled)."""
+        return self._get("/v1/stats/dcs")
+
+    # ------------------------------------------------------------------
+    # Runtime deep-dive
+    # ------------------------------------------------------------------
+
+    def runtime_me_pool_state(self) -> APIResponse:
+        """GET /v1/runtime/me_pool_state — ME pool generation/writer/refill snapshot."""
+        return self._get("/v1/runtime/me_pool_state")
+
+    def runtime_me_quality(self) -> APIResponse:
+        """GET /v1/runtime/me_quality — ME KDF, route-drop, and per-DC RTT counters."""
+        return self._get("/v1/runtime/me_quality")
+
+    def runtime_upstream_quality(self) -> APIResponse:
+        """GET /v1/runtime/upstream_quality — per-upstream health, latency, DC preferences."""
+        return self._get("/v1/runtime/upstream_quality")
+
+    def runtime_nat_stun(self) -> APIResponse:
+        """GET /v1/runtime/nat_stun — NAT probe state, STUN servers, reflected IPs."""
+        return self._get("/v1/runtime/nat_stun")
+
+    def runtime_me_selftest(self) -> APIResponse:
+        """GET /v1/runtime/me-selftest — KDF/timeskew/IP/PID/BND health state."""
+        return self._get("/v1/runtime/me-selftest")
+
+    def runtime_connections_summary(self) -> APIResponse:
+        """GET /v1/runtime/connections/summary — live connection totals + top-N users (requires runtime_edge_enabled)."""
+        return self._get("/v1/runtime/connections/summary")
+
+    def runtime_events_recent(self, limit: int | None = None) -> APIResponse:
+        """GET /v1/runtime/events/recent — recent ring-buffer events (requires runtime_edge_enabled).
+
+        Parameters
+        ----------
+        limit:
+            Optional cap on returned events (1–1000, server default 50).
+        """
+        query = {"limit": str(limit)} if limit is not None else None
+        return self._get("/v1/runtime/events/recent", query=query)
+
+    # ------------------------------------------------------------------
+    # Users (read)
+    # ------------------------------------------------------------------
+
+    def list_users(self) -> APIResponse:
+        """GET /v1/users — list all users with connection/traffic info."""
+        return self._get("/v1/users")
+
+    def get_user(self, username: str) -> APIResponse:
+        """GET /v1/users/{username} — single user info."""
+        return self._get(f"/v1/users/{_safe(username)}")
+
+    # ------------------------------------------------------------------
+    # Users (write)
+    # ------------------------------------------------------------------
+
+    def create_user(
+        self,
+        username: str,
+        *,
+        secret: str | None = None,
+        user_ad_tag: str | None = None,
+        max_tcp_conns: int | None = None,
+        expiration_rfc3339: str | None = None,
+        data_quota_bytes: int | None = None,
+        max_unique_ips: int | None = None,
+        if_match: str | None = None,
+    ) -> APIResponse:
+        """POST /v1/users — create a new user.
+
+        Parameters
+        ----------
+        username:
+            ``[A-Za-z0-9_.-]``, length 1–64.
+        secret:
+            Exactly 32 hex chars. Auto-generated if omitted.
+        user_ad_tag:
+            Exactly 32 hex chars.
+        max_tcp_conns:
+            Per-user concurrent TCP limit.
+        expiration_rfc3339:
+            RFC3339 expiration timestamp, e.g. ``"2025-12-31T23:59:59Z"``.
+        data_quota_bytes:
+            Per-user traffic quota in bytes.
+        max_unique_ips:
+            Per-user unique source IP limit.
+        if_match:
+            Optional ``If-Match`` revision for optimistic concurrency.
+        """
+        body: Dict[str, Any] = {"username": username}
+        _opt(body, "secret", secret)
+        _opt(body, "user_ad_tag", user_ad_tag)
+        _opt(body, "max_tcp_conns", max_tcp_conns)
+        _opt(body, "expiration_rfc3339", expiration_rfc3339)
+        _opt(body, "data_quota_bytes", data_quota_bytes)
+        _opt(body, "max_unique_ips", max_unique_ips)
+        return self._post("/v1/users", body=body, if_match=if_match)
+
+    def patch_user(
+        self,
+        username: str,
+        *,
+        secret: str | None = None,
+        user_ad_tag: str | None = None,
+        max_tcp_conns: int | None = None,
+        expiration_rfc3339: str | None = None,
+        data_quota_bytes: int | None = None,
+        max_unique_ips: int | None = None,
+        if_match: str | None = None,
+    ) -> APIResponse:
+        """PATCH /v1/users/{username} — partial update; only provided fields change.
+
+        Parameters
+        ----------
+        username:
+            Existing username to update.
+        secret:
+            New secret (32 hex chars).
+        user_ad_tag:
+            New ad tag (32 hex chars).
+        max_tcp_conns:
+            New TCP concurrency limit.
+        expiration_rfc3339:
+            New expiration timestamp.
+        data_quota_bytes:
+            New quota in bytes.
+        max_unique_ips:
+            New unique IP limit.
+        if_match:
+            Optional ``If-Match`` revision.
+        """
+        body: Dict[str, Any] = {}
+        _opt(body, "secret", secret)
+        _opt(body, "user_ad_tag", user_ad_tag)
+        _opt(body, "max_tcp_conns", max_tcp_conns)
+        _opt(body, "expiration_rfc3339", expiration_rfc3339)
+        _opt(body, "data_quota_bytes", data_quota_bytes)
+        _opt(body, "max_unique_ips", max_unique_ips)
+        if not body:
+            raise ValueError("patch_user: at least one field must be provided")
+        return self._patch(f"/v1/users/{_safe(username)}", body=body,
+                           if_match=if_match)
+
+    def delete_user(
+        self,
+        username: str,
+        *,
+        if_match: str | None = None,
+    ) -> APIResponse:
+        """DELETE /v1/users/{username} — remove user; blocks deletion of last user.
+
+        Parameters
+        ----------
+        if_match:
+            Optional ``If-Match`` revision for optimistic concurrency.
+        """
+        return self._delete(f"/v1/users/{_safe(username)}", if_match=if_match)
+
+    # NOTE: POST /v1/users/{username}/rotate-secret currently returns 404
+    # in the route matcher (documented limitation). The method is provided
+    # for completeness and future compatibility.
+    def rotate_secret(
+        self,
+        username: str,
+        *,
+        secret: str | None = None,
+        if_match: str | None = None,
+    ) -> APIResponse:
+        """POST /v1/users/{username}/rotate-secret — rotate user secret.
+
+        .. warning::
+            This endpoint currently returns ``404 not_found`` in all released
+            versions (documented route matcher limitation). The method is
+            included for future compatibility.
+
+        Parameters
+        ----------
+        secret:
+            New secret (32 hex chars). Auto-generated if omitted.
+        """
+        body: Dict[str, Any] = {}
+        _opt(body, "secret", secret)
+        return self._post(f"/v1/users/{_safe(username)}/rotate-secret",
+                          body=body or None, if_match=if_match)
+
+    # ------------------------------------------------------------------
+    # Convenience helpers
+    # ------------------------------------------------------------------
+
+    @staticmethod
+    def generate_secret() -> str:
+        """Generate a random 32-character hex secret suitable for user creation."""
+        return secrets.token_hex(16)  # 16 bytes → 32 hex chars
+
+
+# ---------------------------------------------------------------------------
+# Internal helpers
+# ---------------------------------------------------------------------------
+
+def _safe(username: str) -> str:
+    """Minimal guard: reject obvious path-injection attempts."""
+    if "/" in username or "\\" in username:
+        raise ValueError(f"Invalid username: {username!r}")
+    return username
+
+
+def _opt(d: dict, key: str, value: Any) -> None:
+    """Add key to dict only when value is not None."""
+    if value is not None:
+        d[key] = value
+
+
+# ---------------------------------------------------------------------------
+# CLI
+# ---------------------------------------------------------------------------
+
+def _print(resp: APIResponse) -> None:
+    print(json.dumps(resp.data, indent=2))
+    if resp.revision:
+        print(f"# revision: {resp.revision}", flush=True)
+
+
+def _build_parser():
+    import argparse
+
+    p = argparse.ArgumentParser(
+        prog="telemt_api.py",
+        description="Telemt Control API CLI",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+COMMANDS (read)
+  health                          Liveness check
+  info                            System info (version, uptime, config hash)
+  status                          Runtime gates + startup progress
+  init                            Runtime initialization timeline
+  limits                          Effective limits (timeouts, upstream, ME)
+  posture                         Security posture summary
+  whitelist                       IP whitelist entries
+  summary                         Stats summary (conns, uptime, users)
+  zero                            Zero-cost counters (core/upstream/ME/pool/desync)
+  upstreams                       Upstream health + zero counters
+  minimal                         ME writers + DC snapshot  [minimal_runtime_enabled]
+  me-writers                      Per-writer ME status      [minimal_runtime_enabled]
+  dcs                             Per-DC coverage           [minimal_runtime_enabled]
+  me-pool                         ME pool generation/writer/refill snapshot
+  me-quality                      ME KDF, route-drops, per-DC RTT
+  upstream-quality                Per-upstream health + latency
+  nat-stun                        NAT probe state + STUN servers
+  me-selftest                     KDF/timeskew/IP/PID/BND health
+  connections                     Live connection totals + top-N  [runtime_edge_enabled]
+  events [--limit N]              Recent ring-buffer events       [runtime_edge_enabled]
+
+COMMANDS (users)
+  users                           List all users
+  user <username>                 Get single user
+  create <username> [OPTIONS]     Create user
+  patch  <username> [OPTIONS]     Partial update user
+  delete <username>               Delete user
+  secret <username> [--secret S]  Rotate secret (reserved; returns 404 in current release)
+  gen-secret                      Print a random 32-hex secret and exit
+
+USER OPTIONS (for create / patch)
+  --secret S          32 hex chars
+  --ad-tag S          32 hex chars (ad tag)
+  --max-conns N       Max concurrent TCP connections
+  --expires DATETIME  RFC3339 expiration (e.g. 2026-12-31T23:59:59Z)
+  --quota N           Data quota in bytes
+  --max-ips N         Max unique source IPs
+
+EXAMPLES
+  telemt_api.py health
+  telemt_api.py -u http://10.0.0.1:9091 -a mysecret users
+  telemt_api.py create alice --max-conns 5 --quota 10000000000
+  telemt_api.py patch  alice --expires 2027-01-01T00:00:00Z
+  telemt_api.py delete alice
+  telemt_api.py events --limit 20
+        """,
+    )
+
+    p.add_argument("-u", "--url", default="http://127.0.0.1:9091",
+                   metavar="URL", help="API base URL (default: http://127.0.0.1:9091)")
+    p.add_argument("-a", "--auth", default=None, metavar="TOKEN",
+                   help="Authorization header value")
+    p.add_argument("-t", "--timeout", type=int, default=10, metavar="SEC",
+                   help="Request timeout in seconds (default: 10)")
+
+    p.add_argument("command", nargs="?", default="help",
+                   help="Command to run (see COMMANDS below)")
+    p.add_argument("arg", nargs="?", default=None, metavar="USERNAME",
+                   help="Username for user commands")
+
+    # user create/patch fields
+    p.add_argument("--secret",    default=None)
+    p.add_argument("--ad-tag",    dest="ad_tag", default=None)
+    p.add_argument("--max-conns", dest="max_conns", type=int, default=None)
+    p.add_argument("--expires",   default=None)
+    p.add_argument("--quota",     type=int, default=None)
+    p.add_argument("--max-ips",   dest="max_ips", type=int, default=None)
+
+    # events
+    p.add_argument("--limit", type=int, default=None,
+                   help="Max events for `events` command")
+
+    # optimistic concurrency
+    p.add_argument("--if-match", dest="if_match", default=None,
+                   metavar="REVISION", help="If-Match revision header")
+
+    return p
+
+
+if __name__ == "__main__":
+    import sys
+
+    parser = _build_parser()
+    args = parser.parse_args()
+
+    cmd = (args.command or "help").lower()
+
+    if cmd in ("help", "--help"):
+        parser.print_help()
+        sys.exit(0)
+
+    if cmd == "gen-secret":
+        print(TememtAPI.generate_secret())
+        sys.exit(0)
+
+    api = TememtAPI(args.url, auth_header=args.auth, timeout=args.timeout)
+
+    try:
+        # -- read endpoints --------------------------------------------------
+        if cmd == "health":
+            _print(api.health())
+
+        elif cmd == "info":
+            _print(api.system_info())
+
+        elif cmd == "status":
+            _print(api.runtime_gates())
+
+        elif cmd == "init":
+            _print(api.runtime_initialization())
+
+        elif cmd == "limits":
+            _print(api.limits_effective())
+
+        elif cmd == "posture":
+            _print(api.security_posture())
+
+        elif cmd == "whitelist":
+            _print(api.security_whitelist())
+
+        elif cmd == "summary":
+            _print(api.stats_summary())
+
+        elif cmd == "zero":
+            _print(api.stats_zero_all())
+
+        elif cmd == "upstreams":
+            _print(api.stats_upstreams())
+
+        elif cmd == "minimal":
+            _print(api.stats_minimal_all())
+
+        elif cmd == "me-writers":
+            _print(api.stats_me_writers())
+
+        elif cmd == "dcs":
+            _print(api.stats_dcs())
+
+        elif cmd == "me-pool":
+            _print(api.runtime_me_pool_state())
+
+        elif cmd == "me-quality":
+            _print(api.runtime_me_quality())
+
+        elif cmd == "upstream-quality":
+            _print(api.runtime_upstream_quality())
+
+        elif cmd == "nat-stun":
+            _print(api.runtime_nat_stun())
+
+        elif cmd == "me-selftest":
+            _print(api.runtime_me_selftest())
+
+        elif cmd == "connections":
+            _print(api.runtime_connections_summary())
+
+        elif cmd == "events":
+            _print(api.runtime_events_recent(limit=args.limit))
+
+        # -- user read -------------------------------------------------------
+        elif cmd == "users":
+            resp = api.list_users()
+            users = resp.data or []
+            if not users:
+                print("No users configured.")
+            else:
+                fmt = "{:<24} {:>7}  {:>14}  {}"
+                print(fmt.format("USERNAME", "CONNS", "OCTETS", "LINKS"))
+                print("-" * 72)
+                for u in users:
+                    links = (u.get("links") or {})
+                    all_links = (links.get("classic") or []) + \
+                                (links.get("secure") or []) + \
+                                (links.get("tls") or [])
+                    link_str = all_links[0] if all_links else "-"
+                    print(fmt.format(
+                        u["username"],
+                        u.get("current_connections", 0),
+                        u.get("total_octets", 0),
+                        link_str,
+                    ))
+            if resp.revision:
+                print(f"# revision: {resp.revision}")
+
+        elif cmd == "user":
+            if not args.arg:
+                parser.error("user command requires <username>")
+            _print(api.get_user(args.arg))
+
+        # -- user write ------------------------------------------------------
+        elif cmd == "create":
+            if not args.arg:
+                parser.error("create command requires <username>")
+            resp = api.create_user(
+                args.arg,
+                secret=args.secret,
+                user_ad_tag=args.ad_tag,
+                max_tcp_conns=args.max_conns,
+                expiration_rfc3339=args.expires,
+                data_quota_bytes=args.quota,
+                max_unique_ips=args.max_ips,
+                if_match=args.if_match,
+            )
+            d = resp.data or {}
+            print(f"Created: {d.get('user', {}).get('username')}")
+            print(f"Secret:  {d.get('secret')}")
+            links = (d.get("user") or {}).get("links") or {}
+            for kind, lst in links.items():
+                for link in (lst or []):
+                    print(f"Link ({kind}): {link}")
+            if resp.revision:
+                print(f"# revision: {resp.revision}")
+
+        elif cmd == "patch":
+            if not args.arg:
+                parser.error("patch command requires <username>")
+            if not any([args.secret, args.ad_tag, args.max_conns,
+                        args.expires, args.quota, args.max_ips]):
+                parser.error("patch requires at least one field (--secret, --max-conns, --expires, --quota, --max-ips, --ad-tag)")
+            _print(api.patch_user(
+                args.arg,
+                secret=args.secret,
+                user_ad_tag=args.ad_tag,
+                max_tcp_conns=args.max_conns,
+                expiration_rfc3339=args.expires,
+                data_quota_bytes=args.quota,
+                max_unique_ips=args.max_ips,
+                if_match=args.if_match,
+            ))
+
+        elif cmd == "delete":
+            if not args.arg:
+                parser.error("delete command requires <username>")
+            resp = api.delete_user(args.arg, if_match=args.if_match)
+            print(f"Deleted: {resp.data}")
+            if resp.revision:
+                print(f"# revision: {resp.revision}")
+
+        elif cmd == "secret":
+            if not args.arg:
+                parser.error("secret command requires <username>")
+            _print(api.rotate_secret(args.arg, secret=args.secret,
+                                     if_match=args.if_match))
+
+        else:
+            print(f"Unknown command: {cmd!r}\nRun with 'help' to see available commands.",
+                  file=sys.stderr)
+            sys.exit(1)
+
+    except TememtAPIError as exc:
+        print(f"API error [{exc.http_status}] {exc.code}: {exc}", file=sys.stderr)
+        sys.exit(1)
+    except KeyboardInterrupt:
+        sys.exit(130)

tools/zbx_telemt_template.yaml

@@ -1165,6 +1165,60 @@ zabbix_export:
               tags:
                 - tag: Application
                   value: 'Users connections'
+          graph_prototypes:
+            - uuid: 4199de3dcea943d8a1ec62dc297b2e9f
+              name: 'User {#TELEMT_USER}: Connections'
+              graph_items:
+                - color: 1A7C11
+                  item:
+                    host: Telemt
+                    key: 'telemt.active_conn_[{#TELEMT_USER}]'
+                - color: F63100
+                  sortorder: '1'
+                  item:
+                    host: Telemt
+                    key: 'telemt.total_conn_[{#TELEMT_USER}]'
+            - uuid: 84b8f22d891e49768891f497cac12fb3
+              name: 'User {#TELEMT_USER}: IPs'
+              graph_items:
+                - color: 0080FF
+                  item:
+                    host: Telemt
+                    key: 'telemt.ips_current_[{#TELEMT_USER}]'
+                - color: FF8000
+                  sortorder: '1'
+                  item:
+                    host: Telemt
+                    key: 'telemt.ips_limit_[{#TELEMT_USER}]'
+                - color: AA00FF
+                  sortorder: '2'
+                  item:
+                    host: Telemt
+                    key: 'telemt.ips_utilization_[{#TELEMT_USER}]'
+            - uuid: 09dabe7125114e36a6ce40788a7cb888
+              name: 'User {#TELEMT_USER}: Traffic'
+              graph_items:
+                - color: 00AA00
+                  item:
+                    host: Telemt
+                    key: 'telemt.octets_from_[{#TELEMT_USER}]'
+                - color: AA0000
+                  sortorder: '1'
+                  item:
+                    host: Telemt
+                    key: 'telemt.octets_to_[{#TELEMT_USER}]'
+            - uuid: 367f458962574b0ab3c02278a4cd7ecb
+              name: 'User {#TELEMT_USER}: Messages'
+              graph_items:
+                - color: 00AAFF
+                  item:
+                    host: Telemt
+                    key: 'telemt.msgs_from_[{#TELEMT_USER}]'
+                - color: FF5500
+                  sortorder: '1'
+                  item:
+                    host: Telemt
+                    key: 'telemt.msgs_to_[{#TELEMT_USER}]'
           master_item:
             key: telemt.prom_metrics
           lld_macro_paths:
@@ -1177,3 +1231,206 @@ zabbix_export:
       tags:
         - tag: target
           value: Telemt
+  graphs:
+    - uuid: f162658049ca4f50893c5cc02515ff10
+      name: 'Telemt: Server Connections Overview'
+      graph_items:
+        - color: 1A7C11
+          item:
+            host: Telemt
+            key: telemt.conn_total
+        - color: F63100
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.conn_bad_total
+        - color: FC6EA3
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.handshake_timeouts_total
+    - uuid: 759eca5e687142f19248f9d9343e1adf
+      name: 'Telemt: Uptime'
+      graph_items:
+        - color: 0080FF
+          item:
+            host: Telemt
+            key: telemt.uptime
+    - uuid: 0a27dbd0490d4a508c03ed39fa18545d
+      name: 'Telemt: ME Keepalive'
+      graph_items:
+        - color: 1A7C11
+          item:
+            host: Telemt
+            key: telemt.me_keepalive_sent_total
+        - color: 00AA00
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.me_keepalive_pong_total
+        - color: F63100
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.me_keepalive_failed_total
+        - color: FF8000
+          sortorder: '3'
+          item:
+            host: Telemt
+            key: telemt.me_keepalive_timeout_total
+    - uuid: 4015e24ff70b49f484e884d1dde687c0
+      name: 'Telemt: ME Reconnects'
+      graph_items:
+        - color: 0080FF
+          item:
+            host: Telemt
+            key: telemt.me_reconnect_attempts_total
+        - color: 1A7C11
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.me_reconnect_success_total
+    - uuid: f3e3eeb0663c471aa26cf4b6872b0c50
+      name: 'Telemt: ME Route Drops'
+      graph_items:
+        - color: F63100
+          item:
+            host: Telemt
+            key: telemt.me_route_drop_channel_closed_total
+        - color: FF8000
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.me_route_drop_no_conn_total
+        - color: AA00FF
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.me_route_drop_queue_full_total
+    - uuid: 49b51ed78a5943bdbd6d1d34fe28bf61
+      name: 'Telemt: ME Writer Pool'
+      graph_items:
+        - color: 0080FF
+          item:
+            host: Telemt
+            key: telemt.pool_drain_active
+        - color: F63100
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.pool_force_close_total
+        - color: FF8000
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.pool_stale_pick_total
+        - color: 1A7C11
+          sortorder: '3'
+          item:
+            host: Telemt
+            key: telemt.pool_swap_total
+    - uuid: a0779e6c979f4c1ab7ac4da7123a5ecb
+      name: 'Telemt: ME Writer Removals and Restores'
+      graph_items:
+        - color: F63100
+          item:
+            host: Telemt
+            key: telemt.me_writer_removed_total
+        - color: FF8000
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.me_writer_removed_unexpected_total
+        - color: FFAA00
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.me_writer_removed_unexpected_minus_restored_total
+        - color: 1A7C11
+          sortorder: '3'
+          item:
+            host: Telemt
+            key: telemt.me_writer_restored_same_endpoint_total
+        - color: 00AA00
+          sortorder: '4'
+          item:
+            host: Telemt
+            key: telemt.me_writer_restored_fallback_total
+    - uuid: 4fead70290664953b026a228108bee0e
+      name: 'Telemt: Desync Detections'
+      graph_items:
+        - color: F63100
+          item:
+            host: Telemt
+            key: telemt.desync_total
+        - color: 1A7C11
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.desync_full_logged_total
+        - color: FF8000
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.desync_suppressed_total
+    - uuid: 9f8c9f48cb534a66ac21b1bba1acb602
+      name: 'Telemt: Upstream Connect Cycles'
+      graph_items:
+        - color: 0080FF
+          item:
+            host: Telemt
+            key: telemt.upstream_connect_attempt_total
+        - color: 1A7C11
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.upstream_connect_success_total
+        - color: F63100
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.upstream_connect_fail_total
+        - color: FF8000
+          sortorder: '3'
+          item:
+            host: Telemt
+            key: telemt.upstream_connect_failfast_hard_error_total
+    - uuid: 05182057727547f8b8884b7e71e34f19
+      name: 'Telemt: ME Single-Endpoint Outages'
+      graph_items:
+        - color: F63100
+          item:
+            host: Telemt
+            key: telemt.me_single_endpoint_outage_enter_total
+        - color: 1A7C11
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.me_single_endpoint_outage_exit_total
+        - color: 0080FF
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.me_single_endpoint_outage_reconnect_attempt_total
+        - color: 00AA00
+          sortorder: '3'
+          item:
+            host: Telemt
+            key: telemt.me_single_endpoint_outage_reconnect_success_total
+    - uuid: 6892e8b7fbd2445d9ccc0574af58a354
+      name: 'Telemt: ME Refill Activity'
+      graph_items:
+        - color: 0080FF
+          item:
+            host: Telemt
+            key: telemt.me_refill_triggered_total
+        - color: F63100
+          sortorder: '1'
+          item:
+            host: Telemt
+            key: telemt.me_refill_failed_total
+        - color: FF8000
+          sortorder: '2'
+          item:
+            host: Telemt
+            key: telemt.me_refill_skipped_inflight_total