Merge 25847c9a00 into f8e1e2f2ea

Add adversarial and security tests for client, handshake, and relay modules

.cargo/deny.toml

@@ -0,0 +1,15 @@
+[bans]
+multiple-versions = "deny"
+wildcards = "allow"
+highlight = "all"
+
+# Explicitly flag the weak cryptography so the agent is forced to justify its existence
+[[bans.skip]]
+name = "md-5"
+version = "*"
+reason = "MUST VERIFY: Only allowed for legacy checksums, never for security."
+
+[[bans.skip]]
+name = "sha1"
+version = "*"
+reason = "MUST VERIFY: Only allowed for backwards compatibility."

.gitignore

@@ -21,3 +21,4 @@ target
 #.idea/
 
 proxy-secret
+coverage-html/

AGENTS.md

@@ -5,6 +5,22 @@ Your responses are precise, minimal, and architecturally sound. You are working
 
 ---
 
+### Context: The Telemt Project
+
+You are working on **Telemt**, a high-performance, production-grade Telegram MTProxy implementation written in Rust. It is explicitly designed to operate in highly hostile network environments and evade advanced network censorship.
+
+**Adversarial Threat Model:**
+The proxy operates under constant surveillance by DPI (Deep Packet Inspection) systems and active scanners (state firewalls, mobile operator fraud controls). These entities actively probe IPs, analyze protocol handshakes, and look for known proxy signatures to block or throttle traffic.
+
+**Core Architectural Pillars:**
+1. **TLS-Fronting (TLS-F) & TCP-Splitting (TCP-S):** To the outside world, Telemt looks like a standard TLS server. If a client presents a valid MTProxy key, the connection is handled internally. If a censor's scanner, web browser, or unauthorized crawler connects, Telemt seamlessly splices the TCP connection (L4) to a real, legitimate HTTPS fallback server (e.g., Nginx) without modifying the `ClientHello` or terminating the TLS handshake.
+2. **Middle-End (ME) Orchestration:** A highly concurrent, generation-based pool managing upstream connections to Telegram Datacenters (DCs). It utilizes an **Adaptive Floor** (dynamically scaling writer connections based on traffic), **Hardswaps** (zero-downtime pool reconfiguration), and **STUN/NAT** reflection mechanisms. 
+3. **Strict KDF Routing:** Cryptographic Key Derivation Functions (KDF) in this protocol strictly rely on the exact pairing of Source IP/Port and Destination IP/Port. Deviations or missing port logic will silently break the MTProto handshake.
+4. **Data Plane vs. Control Plane Isolation:** The Data Plane (readers, writers, payload relay, TCP splicing) must remain strictly non-blocking, zero-allocation in hot paths, and highly resilient to network backpressure. The Control Plane (API, metrics, pool generation swaps, config reloads) orchestrates the state asynchronously without stalling the Data Plane.
+
+Any modification you make must preserve Telemt's invisibility to censors, its strict memory-safety invariants, and its hot-path throughput.
+
+
 ### 0. Priority Resolution — Scope Control
 
 This section resolves conflicts between code quality enforcement and scope limitation.
@@ -374,6 +390,12 @@ you MUST explain why existing invariants remain valid.
 - Do not modify existing tests unless the task explicitly requires it.
 - Do not weaken assertions.
 - Preserve determinism in testable components.
+- Bug-first forces the discipline of proving you understand a bug before you fix it. Tests written after a fix almost always pass trivially and catch nothing new.
+- Invariants over scenarios is the core shift. The route_mode table alone would have caught both BUG-1 and BUG-2 before they were written — "snapshot equals watch state after any transition burst" is a two-line property test that fails immediately on the current diverged-atomics code.
+- Differential/model catches logic drift over time.
+- Scheduler pressure is specifically aimed at the concurrent state bugs that keep reappearing. A single-threaded happy-path test of set_mode will never find subtle bugs; 10,000 concurrent calls will find it on the first run.
+- Mutation gate answers your original complaint directly. It measures test power. If you can remove a bounds check and nothing breaks, the suite isn't covering that branch yet — it just says so explicitly.
+- Dead parameter is a code smell rule. 
 
 ### 15. Security Constraints
 

Cargo.lock

@@ -425,6 +425,32 @@ dependencies = [
  "cipher",
 ]
 
+[[package]]
+name = "curve25519-dalek"
+version = "4.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "97fb8b7c4503de7d6ae7b42ab72a5a59857b4c937ec27a3d4539dba95b5ab2be"
+dependencies = [
+ "cfg-if",
+ "cpufeatures",
+ "curve25519-dalek-derive",
+ "fiat-crypto",
+ "rustc_version",
+ "subtle",
+ "zeroize",
+]
+
+[[package]]
+name = "curve25519-dalek-derive"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f46882e17999c6cc590af592290432be3bce0428cb0d5f8b6715e4dc7b383eb3"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.114",
+]
+
 [[package]]
 name = "dashmap"
 version = "5.5.3"
@@ -517,6 +543,12 @@ version = "2.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "37909eebbb50d72f9059c3b6d82c0463f2ff062c9e95845c43a6c9c0355411be"
 
+[[package]]
+name = "fiat-crypto"
+version = "0.2.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "28dea519a9695b9977216879a3ebfddf92f1c08c05d984f8996aecd6ecdc811d"
+
 [[package]]
 name = "filetime"
 version = "0.2.27"
@@ -1609,7 +1641,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6db2770f06117d490610c7488547d543617b21bfa07796d7a12f6f1bd53850d1"
 dependencies = [
  "rand_chacha",
- "rand_core",
+ "rand_core 0.9.5",
 ]
 
 [[package]]
@@ -1619,9 +1651,15 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
 dependencies = [
  "ppv-lite86",
- "rand_core",
+ "rand_core 0.9.5",
 ]
 
+[[package]]
+name = "rand_core"
+version = "0.6.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c"
+
 [[package]]
 name = "rand_core"
 version = "0.9.5"
@@ -1637,7 +1675,7 @@ version = "0.4.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "513962919efc330f829edb2535844d1b912b0fbe2ca165d613e4e8788bb05a5a"
 dependencies = [
- "rand_core",
+ "rand_core 0.9.5",
 ]
 
 [[package]]
@@ -2025,6 +2063,12 @@ version = "1.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6ce2be8dc25455e1f91df71bfa12ad37d7af1092ae736f3a6cd0e37bc7810596"
 
+[[package]]
+name = "static_assertions"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a2eb9349b6444b326872e140eb1cf5e7c522154d69e7a0ffb0fb81c06b37543f"
+
 [[package]]
 name = "subtle"
 version = "2.6.1"
@@ -2087,7 +2131,7 @@ dependencies = [
 
 [[package]]
 name = "telemt"
-version = "3.3.19"
+version = "3.3.20"
 dependencies = [
  "aes",
  "anyhow",
@@ -2127,6 +2171,8 @@ dependencies = [
  "sha1",
  "sha2",
  "socket2 0.5.10",
+ "static_assertions",
+ "subtle",
  "thiserror 2.0.18",
  "tokio",
  "tokio-rustls",
@@ -2137,6 +2183,7 @@ dependencies = [
  "tracing-subscriber",
  "url",
  "webpki-roots 0.26.11",
+ "x25519-dalek",
  "x509-parser",
  "zeroize",
 ]
@@ -3136,6 +3183,18 @@ version = "0.6.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9edde0db4769d2dc68579893f2306b26c6ecfbe0ef499b013d731b7b9247e0b9"
 
+[[package]]
+name = "x25519-dalek"
+version = "2.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c7e468321c81fb07fa7f4c636c3972b9100f0346e5b6a9f2bd0603a52f7ed277"
+dependencies = [
+ "curve25519-dalek",
+ "rand_core 0.6.4",
+ "serde",
+ "zeroize",
+]
+
 [[package]]
 name = "x509-parser"
 version = "0.15.1"

Cargo.toml

@@ -22,6 +22,8 @@ hmac = "0.12"
 crc32fast = "1.4"
 crc32c = "0.6"
 zeroize = { version = "1.8", features = ["derive"] }
+subtle = "2.6"
+static_assertions = "1.1"
 
 # Network
 socket2 = { version = "0.5", features = ["all"] }
@@ -50,6 +52,7 @@ regex = "1.11"
 crossbeam-queue = "0.3"
 num-bigint = "0.4"
 num-traits = "0.2"
+x25519-dalek = "2"
 anyhow = "1.0"
 
 # HTTP

config.toml

@@ -38,7 +38,7 @@ port = 443
 [server.api]
 enabled = true
 listen = "0.0.0.0:9091"
-whitelist = ["127.0.0.0/8"]
+whitelist = ["127.0.0.0/8", "172.16.0.0/12"]
 minimal_runtime_enabled = false
 minimal_runtime_cache_ttl_ms = 1000
 

docker-compose.yml

@@ -6,7 +6,8 @@ services:
     restart: unless-stopped
     ports:
       - "443:443"
-      - "127.0.0.1:9090:9090"
+      - "127.0.0.1:9090:9090"   # Metrics
+      - "127.0.0.1:9091:9091"   # API
     # Allow caching 'proxy-secret' in read-only container
     working_dir: /run/telemt
     volumes:
@@ -28,3 +29,8 @@ services:
       nofile:
         soft: 65536
         hard: 65536
+    logging:
+      driver: "json-file"
+      options:
+        max-size: "10m"
+        max-file: "3"

src/cli.rs

@@ -239,7 +239,7 @@ tls_full_cert_ttl_secs = 90
 
 [access]
 replay_check_len = 65536
-replay_window_secs = 1800
+replay_window_secs = 120
 ignore_time_skew = false
 
 [access.users]

src/config/defaults.rs

@@ -73,7 +73,9 @@ pub(crate) fn default_replay_check_len() -> usize {
 }
 
 pub(crate) fn default_replay_window_secs() -> u64 {
-    1800
+    // Keep replay cache TTL tight by default to reduce replay surface.
+    // Deployments with higher RTT or longer reconnect jitter can override this in config.
+    120
 }
 
 pub(crate) fn default_handshake_timeout() -> u64 {
@@ -456,11 +458,11 @@ pub(crate) fn default_tls_full_cert_ttl_secs() -> u64 {
 }
 
 pub(crate) fn default_server_hello_delay_min_ms() -> u64 {
-    0
+    8
 }
 
 pub(crate) fn default_server_hello_delay_max_ms() -> u64 {
-    0
+    24
 }
 
 pub(crate) fn default_alpn_enforce() -> bool {

src/config/hot_reload.rs

@@ -37,7 +37,6 @@ use crate::config::{
 };
 use super::load::{LoadedConfig, ProxyConfig};
 
-const HOT_RELOAD_STABLE_SNAPSHOTS: u8 = 2;
 const HOT_RELOAD_DEBOUNCE: Duration = Duration::from_millis(50);
 
 // ── Hot fields ────────────────────────────────────────────────────────────────
@@ -329,41 +328,19 @@ impl WatchManifest {
 #[derive(Debug, Default)]
 struct ReloadState {
     applied_snapshot_hash: Option<u64>,
-    candidate_snapshot_hash: Option<u64>,
-    candidate_hits: u8,
 }
 
 impl ReloadState {
     fn new(applied_snapshot_hash: Option<u64>) -> Self {
-        Self {
-            applied_snapshot_hash,
-            candidate_snapshot_hash: None,
-            candidate_hits: 0,
-        }
+        Self { applied_snapshot_hash }
     }
 
     fn is_applied(&self, hash: u64) -> bool {
         self.applied_snapshot_hash == Some(hash)
     }
 
-    fn observe_candidate(&mut self, hash: u64) -> u8 {
-        if self.candidate_snapshot_hash == Some(hash) {
-            self.candidate_hits = self.candidate_hits.saturating_add(1);
-        } else {
-            self.candidate_snapshot_hash = Some(hash);
-            self.candidate_hits = 1;
-        }
-        self.candidate_hits
-    }
-
-    fn reset_candidate(&mut self) {
-        self.candidate_snapshot_hash = None;
-        self.candidate_hits = 0;
-    }
-
     fn mark_applied(&mut self, hash: u64) {
         self.applied_snapshot_hash = Some(hash);
-        self.reset_candidate();
     }
 }
 
@@ -1138,7 +1115,6 @@ fn reload_config(
     let loaded = match ProxyConfig::load_with_metadata(config_path) {
         Ok(loaded) => loaded,
         Err(e) => {
-            reload_state.reset_candidate();
             error!("config reload: failed to parse {:?}: {}", config_path, e);
             return None;
         }
@@ -1151,7 +1127,6 @@ fn reload_config(
     let next_manifest = WatchManifest::from_source_files(&source_files);
 
     if let Err(e) = new_cfg.validate() {
-        reload_state.reset_candidate();
         error!("config reload: validation failed: {}; keeping old config", e);
         return Some(next_manifest);
     }
@@ -1160,17 +1135,6 @@ fn reload_config(
         return Some(next_manifest);
     }
 
-    let candidate_hits = reload_state.observe_candidate(rendered_hash);
-    if candidate_hits < HOT_RELOAD_STABLE_SNAPSHOTS {
-        info!(
-            snapshot_hash = rendered_hash,
-            candidate_hits,
-            required_hits = HOT_RELOAD_STABLE_SNAPSHOTS,
-            "config reload: candidate snapshot observed but not stable yet"
-        );
-        return Some(next_manifest);
-    }
-
     let old_cfg = config_tx.borrow().clone();
     let applied_cfg = overlay_hot_fields(&old_cfg, &new_cfg);
     let old_hot = HotFields::from_config(&old_cfg);
@@ -1190,7 +1154,6 @@ fn reload_config(
     if old_hot.dns_overrides != applied_hot.dns_overrides
         && let Err(e) = crate::network::dns_overrides::install_entries(&applied_hot.dns_overrides)
     {
-        reload_state.reset_candidate();
         error!(
             "config reload: invalid network.dns_overrides: {}; keeping old config",
             e
@@ -1334,14 +1297,28 @@ pub fn spawn_config_watcher(
             tokio::time::sleep(HOT_RELOAD_DEBOUNCE).await;
             while notify_rx.try_recv().is_ok() {}
 
-            if let Some(next_manifest) = reload_config(
+            let mut next_manifest = reload_config(
                 &config_path,
                 &config_tx,
                 &log_tx,
                 detected_ip_v4,
                 detected_ip_v6,
                 &mut reload_state,
-            ) {
+            );
+            if next_manifest.is_none() {
+                tokio::time::sleep(HOT_RELOAD_DEBOUNCE).await;
+                while notify_rx.try_recv().is_ok() {}
+                next_manifest = reload_config(
+                    &config_path,
+                    &config_tx,
+                    &log_tx,
+                    detected_ip_v4,
+                    detected_ip_v6,
+                    &mut reload_state,
+                );
+            }
+
+            if let Some(next_manifest) = next_manifest {
                 apply_watch_manifest(
                     inotify_watcher.as_mut(),
                     poll_watcher.as_mut(),
@@ -1466,7 +1443,7 @@ mod tests {
     }
 
     #[test]
-    fn reload_requires_stable_snapshot_before_hot_apply() {
+    fn reload_applies_hot_change_on_first_observed_snapshot() {
         let initial_tag = "11111111111111111111111111111111";
         let final_tag = "22222222222222222222222222222222";
         let path = temp_config_path("telemt_hot_reload_stable");
@@ -1478,20 +1455,7 @@ mod tests {
         let (log_tx, _log_rx) = watch::channel(initial_cfg.general.log_level.clone());
         let mut reload_state = ReloadState::new(Some(initial_hash));
 
-        write_reload_config(&path, None, None);
-        reload_config(&path, &config_tx, &log_tx, None, None, &mut reload_state).unwrap();
-        assert_eq!(
-            config_tx.borrow().general.ad_tag.as_deref(),
-            Some(initial_tag)
-        );
-
         write_reload_config(&path, Some(final_tag), None);
-        reload_config(&path, &config_tx, &log_tx, None, None, &mut reload_state).unwrap();
-        assert_eq!(
-            config_tx.borrow().general.ad_tag.as_deref(),
-            Some(initial_tag)
-        );
-
         reload_config(&path, &config_tx, &log_tx, None, None, &mut reload_state).unwrap();
         assert_eq!(config_tx.borrow().general.ad_tag.as_deref(), Some(final_tag));
 
@@ -1513,7 +1477,6 @@ mod tests {
 
         write_reload_config(&path, Some(final_tag), Some(initial_cfg.server.port + 1));
         reload_config(&path, &config_tx, &log_tx, None, None, &mut reload_state).unwrap();
-        reload_config(&path, &config_tx, &log_tx, None, None, &mut reload_state).unwrap();
 
         let applied = config_tx.borrow().clone();
         assert_eq!(applied.general.ad_tag.as_deref(), Some(final_tag));
@@ -1521,4 +1484,31 @@ mod tests {
 
         let _ = std::fs::remove_file(path);
     }
+
+    #[test]
+    fn reload_recovers_after_parse_error_on_next_attempt() {
+        let initial_tag = "cccccccccccccccccccccccccccccccc";
+        let final_tag = "dddddddddddddddddddddddddddddddd";
+        let path = temp_config_path("telemt_hot_reload_parse_recovery");
+
+        write_reload_config(&path, Some(initial_tag), None);
+        let initial_cfg = Arc::new(ProxyConfig::load(&path).unwrap());
+        let initial_hash = ProxyConfig::load_with_metadata(&path).unwrap().rendered_hash;
+        let (config_tx, _config_rx) = watch::channel(initial_cfg.clone());
+        let (log_tx, _log_rx) = watch::channel(initial_cfg.general.log_level.clone());
+        let mut reload_state = ReloadState::new(Some(initial_hash));
+
+        std::fs::write(&path, "[access.users\nuser = \"broken\"\n").unwrap();
+        assert!(reload_config(&path, &config_tx, &log_tx, None, None, &mut reload_state).is_none());
+        assert_eq!(
+            config_tx.borrow().general.ad_tag.as_deref(),
+            Some(initial_tag)
+        );
+
+        write_reload_config(&path, Some(final_tag), None);
+        reload_config(&path, &config_tx, &log_tx, None, None, &mut reload_state).unwrap();
+        assert_eq!(config_tx.borrow().general.ad_tag.as_deref(), Some(final_tag));
+
+        let _ = std::fs::remove_file(path);
+    }
 }

src/config/types.rs

@@ -1156,6 +1156,13 @@ pub struct ServerConfig {
     #[serde(default = "default_proxy_protocol_header_timeout_ms")]
     pub proxy_protocol_header_timeout_ms: u64,
 
+    /// Trusted source CIDRs allowed to send incoming PROXY protocol headers.
+    ///
+    /// When non-empty, connections from addresses outside this allowlist are
+    /// rejected before `src_addr` is applied.
+    #[serde(default)]
+    pub proxy_protocol_trusted_cidrs: Vec<IpNetwork>,
+
     /// Port for the Prometheus-compatible metrics endpoint.
     /// Enables metrics when set; binds on all interfaces (dual-stack) by default.
     #[serde(default)]
@@ -1193,6 +1200,7 @@ impl Default for ServerConfig {
             listen_tcp: None,
             proxy_protocol: false,
             proxy_protocol_header_timeout_ms: default_proxy_protocol_header_timeout_ms(),
+            proxy_protocol_trusted_cidrs: Vec::new(),
             metrics_port: None,
             metrics_listen: None,
             metrics_whitelist: default_metrics_whitelist(),

src/ip_tracker.rs

@@ -7,8 +7,9 @@ use std::net::IpAddr;
 use std::sync::Arc;
 use std::sync::atomic::{AtomicU64, Ordering};
 use std::time::{Duration, Instant};
+use std::sync::Mutex;
 
-use tokio::sync::RwLock;
+use tokio::sync::{Mutex as AsyncMutex, RwLock};
 
 use crate::config::UserMaxUniqueIpsMode;
 
@@ -21,6 +22,8 @@ pub struct UserIpTracker {
     limit_mode: Arc<RwLock<UserMaxUniqueIpsMode>>,
     limit_window: Arc<RwLock<Duration>>,
     last_compact_epoch_secs: Arc<AtomicU64>,
+    pub(crate) cleanup_queue: Arc<Mutex<Vec<(String, IpAddr)>>>,
+    cleanup_drain_lock: Arc<AsyncMutex<()>>,
 }
 
 impl UserIpTracker {
@@ -33,6 +36,67 @@ impl UserIpTracker {
             limit_mode: Arc::new(RwLock::new(UserMaxUniqueIpsMode::ActiveWindow)),
             limit_window: Arc::new(RwLock::new(Duration::from_secs(30))),
             last_compact_epoch_secs: Arc::new(AtomicU64::new(0)),
+            cleanup_queue: Arc::new(Mutex::new(Vec::new())),
+            cleanup_drain_lock: Arc::new(AsyncMutex::new(())),
+        }
+    }
+
+
+    pub fn enqueue_cleanup(&self, user: String, ip: IpAddr) {
+        match self.cleanup_queue.lock() {
+            Ok(mut queue) => queue.push((user, ip)),
+            Err(poisoned) => {
+                let mut queue = poisoned.into_inner();
+                queue.push((user.clone(), ip));
+                self.cleanup_queue.clear_poison();
+                tracing::warn!(
+                    "UserIpTracker cleanup_queue lock poisoned; recovered and enqueued IP cleanup for {} ({})",
+                    user,
+                    ip
+                );
+            }
+        }
+    }
+
+    pub(crate) async fn drain_cleanup_queue(&self) {
+        // Serialize queue draining and active-IP mutation so check-and-add cannot
+        // observe stale active entries that are already queued for removal.
+        let _drain_guard = self.cleanup_drain_lock.lock().await;
+        let to_remove = {
+            match self.cleanup_queue.lock() {
+                Ok(mut queue) => {
+                    if queue.is_empty() {
+                        return;
+                    }
+                    std::mem::take(&mut *queue)
+                }
+                Err(poisoned) => {
+                    let mut queue = poisoned.into_inner();
+                    if queue.is_empty() {
+                        self.cleanup_queue.clear_poison();
+                        return;
+                    }
+                    let drained = std::mem::take(&mut *queue);
+                    self.cleanup_queue.clear_poison();
+                    drained
+                }
+            }
+        };
+
+        let mut active_ips = self.active_ips.write().await;
+        for (user, ip) in to_remove {
+            if let Some(user_ips) = active_ips.get_mut(&user) {
+                if let Some(count) = user_ips.get_mut(&ip) {
+                    if *count > 1 {
+                        *count -= 1;
+                    } else {
+                        user_ips.remove(&ip);
+                    }
+                }
+                if user_ips.is_empty() {
+                    active_ips.remove(&user);
+                }
+            }
         }
     }
 
@@ -118,6 +182,7 @@ impl UserIpTracker {
     }
 
     pub async fn check_and_add(&self, username: &str, ip: IpAddr) -> Result<(), String> {
+        self.drain_cleanup_queue().await;
         self.maybe_compact_empty_users().await;
         let default_max_ips = *self.default_max_ips.read().await;
         let limit = {
@@ -194,6 +259,7 @@ impl UserIpTracker {
     }
 
     pub async fn get_recent_counts_for_users(&self, users: &[String]) -> HashMap<String, usize> {
+        self.drain_cleanup_queue().await;
         let window = *self.limit_window.read().await;
         let now = Instant::now();
         let recent_ips = self.recent_ips.read().await;
@@ -214,6 +280,7 @@ impl UserIpTracker {
     }
 
     pub async fn get_active_ips_for_users(&self, users: &[String]) -> HashMap<String, Vec<IpAddr>> {
+        self.drain_cleanup_queue().await;
         let active_ips = self.active_ips.read().await;
         let mut out = HashMap::with_capacity(users.len());
         for user in users {
@@ -228,6 +295,7 @@ impl UserIpTracker {
     }
 
     pub async fn get_recent_ips_for_users(&self, users: &[String]) -> HashMap<String, Vec<IpAddr>> {
+        self.drain_cleanup_queue().await;
         let window = *self.limit_window.read().await;
         let now = Instant::now();
         let recent_ips = self.recent_ips.read().await;
@@ -250,11 +318,13 @@ impl UserIpTracker {
     }
 
     pub async fn get_active_ip_count(&self, username: &str) -> usize {
+        self.drain_cleanup_queue().await;
         let active_ips = self.active_ips.read().await;
         active_ips.get(username).map(|ips| ips.len()).unwrap_or(0)
     }
 
     pub async fn get_active_ips(&self, username: &str) -> Vec<IpAddr> {
+        self.drain_cleanup_queue().await;
         let active_ips = self.active_ips.read().await;
         active_ips
             .get(username)
@@ -263,6 +333,7 @@ impl UserIpTracker {
     }
 
     pub async fn get_stats(&self) -> Vec<(String, usize, usize)> {
+        self.drain_cleanup_queue().await;
         let active_ips = self.active_ips.read().await;
         let max_ips = self.max_ips.read().await;
         let default_max_ips = *self.default_max_ips.read().await;
@@ -301,6 +372,7 @@ impl UserIpTracker {
     }
 
     pub async fn is_ip_active(&self, username: &str, ip: IpAddr) -> bool {
+        self.drain_cleanup_queue().await;
         let active_ips = self.active_ips.read().await;
         active_ips
             .get(username)

src/ip_tracker_regression_tests.rs

@@ -448,3 +448,172 @@ async fn concurrent_reconnect_and_disconnect_preserves_non_negative_counts() {
 
     assert!(tracker.get_active_ip_count("cc").await <= 8);
 }
+
+#[tokio::test]
+async fn enqueue_cleanup_recovers_from_poisoned_mutex() {
+    let tracker = UserIpTracker::new();
+    let ip = ip_from_idx(99);
+    
+    // Poison the lock by panicking while holding it
+    let result = std::panic::catch_unwind(|| {
+        let _guard = tracker.cleanup_queue.lock().unwrap();
+        panic!("Intentional poison panic");
+    });
+    assert!(result.is_err(), "Expected panic to poison mutex");
+    
+    // Attempt to enqueue anyway; should hit the poison catch arm and still insert
+    tracker.enqueue_cleanup("poison-user".to_string(), ip);
+    
+    tracker.drain_cleanup_queue().await;
+    
+    assert_eq!(tracker.get_active_ip_count("poison-user").await, 0);
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn mass_reconnect_sync_cleanup_prevents_temporary_reservation_bloat() {
+    // Tests that synchronous M-01 drop mechanism protects against starvation
+    let tracker = Arc::new(UserIpTracker::new());
+    tracker.set_user_limit("mass", 5).await;
+    
+    let ip = ip_from_idx(42);
+    let mut join_handles = Vec::new();
+
+    // 10,000 rapid concurrent requests hitting the same IP limit
+    for _ in 0..10_000 {
+        let tracker_clone = tracker.clone();
+        join_handles.push(tokio::spawn(async move {
+            if tracker_clone.check_and_add("mass", ip).await.is_ok() {
+                // Instantly enqueue cleanup, simulating synchronous reservation drop
+                tracker_clone.enqueue_cleanup("mass".to_string(), ip);
+                // The next caller will drain it before acquiring again
+            }
+        }));
+    }
+
+    for handle in join_handles {
+        let _ = handle.await;
+    }
+
+    // Force flush
+    tracker.drain_cleanup_queue().await;
+    assert_eq!(tracker.get_active_ip_count("mass").await, 0, "No leaked footprints");
+}
+
+#[tokio::test]
+async fn adversarial_drain_cleanup_queue_race_does_not_cause_false_rejections() {
+    // Regression guard: concurrent cleanup draining must not produce false
+    // limit denials for a new IP when the previous IP is already queued.
+    let tracker = Arc::new(UserIpTracker::new());
+    tracker.set_user_limit("racer", 1).await;
+    let ip1 = ip_from_idx(1);
+    let ip2 = ip_from_idx(2);
+
+    // Initial state: add ip1
+    tracker.check_and_add("racer", ip1).await.unwrap();
+
+    // User disconnects from ip1, queuing it
+    tracker.enqueue_cleanup("racer".to_string(), ip1);
+
+    let mut saw_false_rejection = false;
+    for _ in 0..100 {
+        // Queue cleanup then race explicit drain and check-and-add on the alternative IP.
+        tracker.enqueue_cleanup("racer".to_string(), ip1);
+        let tracker_a = tracker.clone();
+        let tracker_b = tracker.clone();
+
+        let drain_handle = tokio::spawn(async move {
+            tracker_a.drain_cleanup_queue().await;
+        });
+        let handle = tokio::spawn(async move {
+            tracker_b.check_and_add("racer", ip2).await
+        });
+
+        drain_handle.await.unwrap();
+        let res = handle.await.unwrap();
+        if res.is_err() {
+            saw_false_rejection = true;
+            break;
+        }
+
+        // Restore baseline for next iteration.
+        tracker.remove_ip("racer", ip2).await;
+        tracker.check_and_add("racer", ip1).await.unwrap();
+    }
+
+    assert!(
+        !saw_false_rejection,
+        "Concurrent cleanup draining must not cause false-positive IP denials"
+    );
+}
+
+#[tokio::test]
+async fn poisoned_cleanup_queue_still_releases_slot_for_next_ip() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("poison-slot", 1).await;
+    let ip1 = ip_from_idx(7001);
+    let ip2 = ip_from_idx(7002);
+
+    tracker.check_and_add("poison-slot", ip1).await.unwrap();
+
+    // Poison the queue lock as an adversarial condition.
+    let _ = std::panic::catch_unwind(|| {
+        let _guard = tracker.cleanup_queue.lock().unwrap();
+        panic!("intentional queue poison");
+    });
+
+    // Disconnect path must still queue cleanup so the next IP can be admitted.
+    tracker.enqueue_cleanup("poison-slot".to_string(), ip1);
+    let admitted = tracker.check_and_add("poison-slot", ip2).await;
+    assert!(
+        admitted.is_ok(),
+        "cleanup queue poison must not permanently block slot release for the next IP"
+    );
+}
+
+#[tokio::test]
+async fn duplicate_cleanup_entries_do_not_break_future_admission() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("dup-cleanup", 1).await;
+    let ip1 = ip_from_idx(7101);
+    let ip2 = ip_from_idx(7102);
+
+    tracker.check_and_add("dup-cleanup", ip1).await.unwrap();
+    tracker.enqueue_cleanup("dup-cleanup".to_string(), ip1);
+    tracker.enqueue_cleanup("dup-cleanup".to_string(), ip1);
+    tracker.enqueue_cleanup("dup-cleanup".to_string(), ip1);
+
+    tracker.drain_cleanup_queue().await;
+
+    assert_eq!(tracker.get_active_ip_count("dup-cleanup").await, 0);
+    assert!(
+        tracker.check_and_add("dup-cleanup", ip2).await.is_ok(),
+        "extra queued cleanup entries must not leave user stuck in denied state"
+    );
+}
+
+#[tokio::test]
+async fn stress_repeated_queue_poison_recovery_preserves_admission_progress() {
+    let tracker = UserIpTracker::new();
+    tracker.set_user_limit("poison-stress", 1).await;
+    let ip_primary = ip_from_idx(7201);
+    let ip_alt = ip_from_idx(7202);
+
+    tracker.check_and_add("poison-stress", ip_primary).await.unwrap();
+
+    for _ in 0..64 {
+        let _ = std::panic::catch_unwind(|| {
+            let _guard = tracker.cleanup_queue.lock().unwrap();
+            panic!("intentional queue poison in stress loop");
+        });
+
+        tracker.enqueue_cleanup("poison-stress".to_string(), ip_primary);
+
+        assert!(
+            tracker.check_and_add("poison-stress", ip_alt).await.is_ok(),
+            "poison recovery must preserve admission progress under repeated queue poisoning"
+        );
+
+        tracker.remove_ip("poison-stress", ip_alt).await;
+        tracker.check_and_add("poison-stress", ip_primary).await.unwrap();
+    }
+}

src/maestro/helpers.rs

@@ -10,6 +10,16 @@ use crate::transport::middle_proxy::{
     ProxyConfigData, fetch_proxy_config_with_raw, load_proxy_config_cache, save_proxy_config_cache,
 };
 
+pub(crate) fn resolve_runtime_config_path(config_path_cli: &str, startup_cwd: &std::path::Path) -> PathBuf {
+    let raw = PathBuf::from(config_path_cli);
+    let absolute = if raw.is_absolute() {
+        raw
+    } else {
+        startup_cwd.join(raw)
+    };
+    absolute.canonicalize().unwrap_or(absolute)
+}
+
 pub(crate) fn parse_cli() -> (String, Option<PathBuf>, bool, Option<String>) {
     let mut config_path = "config.toml".to_string();
     let mut data_path: Option<PathBuf> = None;
@@ -96,6 +106,44 @@ pub(crate) fn parse_cli() -> (String, Option<PathBuf>, bool, Option<String>) {
     (config_path, data_path, silent, log_level)
 }
 
+#[cfg(test)]
+mod tests {
+    use super::resolve_runtime_config_path;
+
+    #[test]
+    fn resolve_runtime_config_path_anchors_relative_to_startup_cwd() {
+        let nonce = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap()
+            .as_nanos();
+        let startup_cwd = std::env::temp_dir().join(format!("telemt_cfg_path_{nonce}"));
+        std::fs::create_dir_all(&startup_cwd).unwrap();
+        let target = startup_cwd.join("config.toml");
+        std::fs::write(&target, " ").unwrap();
+
+        let resolved = resolve_runtime_config_path("config.toml", &startup_cwd);
+        assert_eq!(resolved, target.canonicalize().unwrap());
+
+        let _ = std::fs::remove_file(&target);
+        let _ = std::fs::remove_dir(&startup_cwd);
+    }
+
+    #[test]
+    fn resolve_runtime_config_path_keeps_absolute_for_missing_file() {
+        let nonce = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap()
+            .as_nanos();
+        let startup_cwd = std::env::temp_dir().join(format!("telemt_cfg_path_missing_{nonce}"));
+        std::fs::create_dir_all(&startup_cwd).unwrap();
+
+        let resolved = resolve_runtime_config_path("missing.toml", &startup_cwd);
+        assert_eq!(resolved, startup_cwd.join("missing.toml"));
+
+        let _ = std::fs::remove_dir(&startup_cwd);
+    }
+}
+
 pub(crate) fn print_proxy_links(host: &str, port: u16, config: &ProxyConfig) {
     info!(target: "telemt::links", "--- Proxy Links ({}) ---", host);
     for user_name in config.general.links.show.resolve_users(&config.access.users) {

src/maestro/mod.rs

@@ -45,7 +45,7 @@ use crate::startup::{
 use crate::stream::BufferPool;
 use crate::transport::middle_proxy::MePool;
 use crate::transport::UpstreamManager;
-use helpers::parse_cli;
+use helpers::{parse_cli, resolve_runtime_config_path};
 
 /// Runs the full telemt runtime startup pipeline and blocks until shutdown.
 pub async fn run() -> std::result::Result<(), Box<dyn std::error::Error>> {
@@ -58,18 +58,26 @@ pub async fn run() -> std::result::Result<(), Box<dyn std::error::Error>> {
     startup_tracker
         .start_component(COMPONENT_CONFIG_LOAD, Some("load and validate config".to_string()))
         .await;
-    let (config_path, data_path, cli_silent, cli_log_level) = parse_cli();
+    let (config_path_cli, data_path, cli_silent, cli_log_level) = parse_cli();
+    let startup_cwd = match std::env::current_dir() {
+        Ok(cwd) => cwd,
+        Err(e) => {
+            eprintln!("[telemt] Can't read current_dir: {}", e);
+            std::process::exit(1);
+        }
+    };
+    let config_path = resolve_runtime_config_path(&config_path_cli, &startup_cwd);
 
     let mut config = match ProxyConfig::load(&config_path) {
         Ok(c) => c,
         Err(e) => {
-            if std::path::Path::new(&config_path).exists() {
+            if config_path.exists() {
                 eprintln!("[telemt] Error: {}", e);
                 std::process::exit(1);
             } else {
                 let default = ProxyConfig::default();
                 std::fs::write(&config_path, toml::to_string_pretty(&default).unwrap()).unwrap();
-                eprintln!("[telemt] Created default config at {}", config_path);
+                eprintln!("[telemt] Created default config at {}", config_path.display());
                 default
             }
         }
@@ -258,7 +266,7 @@ pub async fn run() -> std::result::Result<(), Box<dyn std::error::Error>> {
             let route_runtime_api = route_runtime.clone();
             let config_rx_api = api_config_rx.clone();
             let admission_rx_api = admission_rx.clone();
-            let config_path_api = std::path::PathBuf::from(&config_path);
+            let config_path_api = config_path.clone();
             let startup_tracker_api = startup_tracker.clone();
             let detected_ips_rx_api = detected_ips_rx.clone();
             tokio::spawn(async move {

src/maestro/runtime_tasks.rs

@@ -1,5 +1,5 @@
 use std::net::IpAddr;
-use std::path::PathBuf;
+use std::path::Path;
 use std::sync::Arc;
 
 use tokio::sync::{mpsc, watch};
@@ -32,7 +32,7 @@ pub(crate) struct RuntimeWatches {
 #[allow(clippy::too_many_arguments)]
 pub(crate) async fn spawn_runtime_tasks(
     config: &Arc<ProxyConfig>,
-    config_path: &str,
+    config_path: &Path,
     probe: &NetworkProbe,
     prefer_ipv6: bool,
     decision_ipv4_dc: bool,
@@ -83,7 +83,7 @@ pub(crate) async fn spawn_runtime_tasks(
         watch::Receiver<Arc<ProxyConfig>>,
         watch::Receiver<LogLevel>,
     ) = spawn_config_watcher(
-        PathBuf::from(config_path),
+        config_path.to_path_buf(),
         config.clone(),
         detected_ip_v4,
         detected_ip_v6,

src/protocol/tls.rs

@@ -11,8 +11,8 @@ use crate::crypto::{sha256_hmac, SecureRandom};
 use crate::error::ProxyError;
 use super::constants::*;
 use std::time::{SystemTime, UNIX_EPOCH};
-use num_bigint::BigUint;
-use num_traits::One;
+use subtle::ConstantTimeEq;
+use x25519_dalek::{X25519_BASEPOINT_BYTES, x25519};
 
 // ============= Public Constants =============
 
@@ -26,8 +26,17 @@ pub const TLS_DIGEST_POS: usize = 11;
 pub const TLS_DIGEST_HALF_LEN: usize = 16;
 
 /// Time skew limits for anti-replay (in seconds)
-pub const TIME_SKEW_MIN: i64 = -20 * 60; // 20 minutes before
-pub const TIME_SKEW_MAX: i64 = 10 * 60;  // 10 minutes after
+///
+/// The default window is intentionally narrow to reduce replay acceptance.
+/// Operators with known clock-drifted clients should tune deployment config
+/// (for example replay-window policy) to match their environment.
+pub const TIME_SKEW_MIN: i64 = -2 * 60; // 2 minutes before
+pub const TIME_SKEW_MAX: i64 = 2 * 60;  // 2 minutes after
+/// Maximum accepted boot-time timestamp (seconds) before skew checks are enforced.
+pub const BOOT_TIME_MAX_SECS: u32 = 7 * 24 * 60 * 60;
+/// Hard cap for boot-time compatibility bypass to avoid oversized acceptance
+/// windows when replay TTL is configured very large.
+pub const BOOT_TIME_COMPAT_MAX_SECS: u32 = 2 * 60;
 
 // ============= Private Constants =============
 
@@ -60,6 +69,7 @@ pub struct TlsValidation {
     /// Client digest for response generation
     pub digest: [u8; TLS_DIGEST_LEN],
     /// Timestamp extracted from digest
+    
     pub timestamp: u32,
 }
 
@@ -114,28 +124,8 @@ impl TlsExtensionBuilder {
         self
     }
 
-    /// Add ALPN extension with a single selected protocol.
-    fn add_alpn(&mut self, proto: &[u8]) -> &mut Self {
-        // Extension type: ALPN (0x0010)
-        self.extensions.extend_from_slice(&extension_type::ALPN.to_be_bytes());
-
-        // ALPN extension format:
-        // extension_data length (2 bytes)
-        //   protocols length (2 bytes)
-        //     protocol name length (1 byte)
-        //     protocol name bytes
-        let proto_len = proto.len() as u8;
-        let list_len: u16 = 1 + proto_len as u16;
-        let ext_len: u16 = 2 + list_len;
-
-        self.extensions.extend_from_slice(&ext_len.to_be_bytes());
-        self.extensions.extend_from_slice(&list_len.to_be_bytes());
-        self.extensions.push(proto_len);
-        self.extensions.extend_from_slice(proto);
-        self
-    }
-    
     /// Build final extensions with length prefix
+    
     fn build(self) -> Vec<u8> {
         let mut result = Vec::with_capacity(2 + self.extensions.len());
         
@@ -150,7 +140,7 @@ impl TlsExtensionBuilder {
     }
     
     /// Get current extensions without length prefix (for calculation)
-    #[allow(dead_code)]
+    
     fn as_bytes(&self) -> &[u8] {
         &self.extensions
     }
@@ -170,8 +160,6 @@ struct ServerHelloBuilder {
     compression: u8,
     /// Extensions
     extensions: TlsExtensionBuilder,
-    /// Selected ALPN protocol (if any)
-    alpn: Option<Vec<u8>>,
 }
 
 impl ServerHelloBuilder {
@@ -182,7 +170,6 @@ impl ServerHelloBuilder {
             cipher_suite: cipher_suite::TLS_AES_128_GCM_SHA256,
             compression: 0x00,
             extensions: TlsExtensionBuilder::new(),
-            alpn: None,
         }
     }
     
@@ -197,18 +184,9 @@ impl ServerHelloBuilder {
         self
     }
 
-    fn with_alpn(mut self, proto: Option<Vec<u8>>) -> Self {
-        self.alpn = proto;
-        self
-    }
-    
     /// Build ServerHello message (without record header)
     fn build_message(&self) -> Vec<u8> {
-        let mut ext_builder = self.extensions.clone();
-        if let Some(ref alpn) = self.alpn {
-            ext_builder.add_alpn(alpn);
-        }
-        let extensions = ext_builder.extensions.clone();
+        let extensions = self.extensions.extensions.clone();
         let extensions_len = extensions.len() as u16;
         
         // Calculate total length
@@ -273,13 +251,97 @@ impl ServerHelloBuilder {
 
 // ============= Public Functions =============
 
-/// Validate TLS ClientHello against user secrets
+/// Validate TLS ClientHello against user secrets.
 ///
 /// Returns validation result if a matching user is found.
+/// The result **must** be used — ignoring it silently bypasses authentication.
+#[must_use]
+
 pub fn validate_tls_handshake(
     handshake: &[u8],
     secrets: &[(String, Vec<u8>)],
     ignore_time_skew: bool,
+) -> Option<TlsValidation> {
+    validate_tls_handshake_with_replay_window(
+        handshake,
+        secrets,
+        ignore_time_skew,
+        u64::from(BOOT_TIME_MAX_SECS),
+    )
+}
+
+/// Validate TLS ClientHello and cap the boot-time bypass by replay-cache TTL.
+///
+/// A boot-time timestamp is only accepted when it falls below all three
+/// bounds: `BOOT_TIME_MAX_SECS`, configured replay window, and
+/// `BOOT_TIME_COMPAT_MAX_SECS`, preventing oversized compatibility windows.
+#[must_use]
+pub fn validate_tls_handshake_with_replay_window(
+    handshake: &[u8],
+    secrets: &[(String, Vec<u8>)],
+    ignore_time_skew: bool,
+    replay_window_secs: u64,
+) -> Option<TlsValidation> {
+    // Only pay the clock syscall when we will actually compare against it.
+    // If `ignore_time_skew` is set, a broken or unavailable system clock
+    // must not block legitimate clients — that would be a DoS via clock failure.
+    let now = if !ignore_time_skew {
+        system_time_to_unix_secs(SystemTime::now())?
+    } else {
+        0_i64
+    };
+
+    let replay_window_u32 = u32::try_from(replay_window_secs).unwrap_or(u32::MAX);
+    // Boot-time bypass and ignore_time_skew serve different compatibility paths.
+    // When skew checks are disabled, force boot-time cap to zero to prevent
+    // accidental future coupling of boot-time logic into the ignore-skew path.
+    let boot_time_cap_secs = if ignore_time_skew {
+        0
+    } else {
+        BOOT_TIME_MAX_SECS
+            .min(replay_window_u32)
+            .min(BOOT_TIME_COMPAT_MAX_SECS)
+    };
+
+    validate_tls_handshake_at_time_with_boot_cap(
+        handshake,
+        secrets,
+        ignore_time_skew,
+        now,
+        boot_time_cap_secs,
+    )
+}
+
+fn system_time_to_unix_secs(now: SystemTime) -> Option<i64> {
+    // `try_from` rejects values that overflow i64 (> ~292 billion years CE),
+    // whereas `as i64` would silently wrap to a negative timestamp and corrupt
+    // every subsequent time-skew comparison.
+    let d = now.duration_since(UNIX_EPOCH).ok()?;
+    i64::try_from(d.as_secs()).ok()
+}
+
+
+fn validate_tls_handshake_at_time(
+    handshake: &[u8],
+    secrets: &[(String, Vec<u8>)],
+    ignore_time_skew: bool,
+    now: i64,
+) -> Option<TlsValidation> {
+    validate_tls_handshake_at_time_with_boot_cap(
+        handshake,
+        secrets,
+        ignore_time_skew,
+        now,
+        BOOT_TIME_MAX_SECS,
+    )
+}
+
+fn validate_tls_handshake_at_time_with_boot_cap(
+    handshake: &[u8],
+    secrets: &[(String, Vec<u8>)],
+    ignore_time_skew: bool,
+    now: i64,
+    boot_time_cap_secs: u32,
 ) -> Option<TlsValidation> {
     if handshake.len() < TLS_DIGEST_POS + TLS_DIGEST_LEN + 1 {
         return None;
@@ -293,6 +355,9 @@ pub fn validate_tls_handshake(
     // Extract session ID
     let session_id_len_pos = TLS_DIGEST_POS + TLS_DIGEST_LEN;
     let session_id_len = handshake.get(session_id_len_pos).copied()? as usize;
+    if session_id_len > 32 {
+        return None;
+    }
     let session_id_start = session_id_len_pos + 1;
     
     if handshake.len() < session_id_start + session_id_len {
@@ -305,73 +370,66 @@ pub fn validate_tls_handshake(
     let mut msg = handshake.to_vec();
     msg[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN].fill(0);
     
-    // Get current time
-    let now = SystemTime::now()
-        .duration_since(UNIX_EPOCH)
-        .unwrap()
-        .as_secs() as i64;
-    
+    let mut first_match: Option<(&String, u32)> = None;
+
     for (user, secret) in secrets {
         let computed = sha256_hmac(secret, &msg);
-        
-        // XOR digests
-        let xored: Vec<u8> = digest.iter()
-            .zip(computed.iter())
-            .map(|(a, b)| a ^ b)
-            .collect();
-        
-        // Check that first 28 bytes are zeros (timestamp in last 4)
-        if !xored[..28].iter().all(|&b| b == 0) {
+
+        // Constant-time equality check on the 28-byte HMAC window.
+        // A variable-time short-circuit here lets an active censor measure how many
+        // bytes matched, enabling secret brute-force via timing side-channels.
+        // Direct comparison on the original arrays avoids a heap allocation and
+        // removes the `try_into().unwrap()` that the intermediate Vec would require.
+        if !bool::from(digest[..28].ct_eq(&computed[..28])) {
             continue;
         }
-        
-        // Extract timestamp
-        let timestamp = u32::from_le_bytes(xored[28..32].try_into().unwrap());
-        let time_diff = now - timestamp as i64;
-        
-        // Check time skew
+
+        // The last 4 bytes encode the timestamp as XOR(digest[28..32], computed[28..32]).
+        // Inline array construction is infallible: both slices are [u8; 32] by construction.
+        let timestamp = u32::from_le_bytes([
+            digest[28] ^ computed[28],
+            digest[29] ^ computed[29],
+            digest[30] ^ computed[30],
+            digest[31] ^ computed[31],
+        ]);
+
+        // time_diff is only meaningful (and `now` is only valid) when we are
+        // actually checking the window.  Keep both inside the guard to make
+        // the dead-code path explicit and prevent accidental future use of
+        // a sentinel `now` value outside its intended scope.
         if !ignore_time_skew {
             // Allow very small timestamps (boot time instead of unix time)
             // This is a quirk in some clients that use uptime instead of real time
-            let is_boot_time = timestamp < 60 * 60 * 24 * 1000; // < ~2.7 years in seconds
-            
-            if !is_boot_time && !(TIME_SKEW_MIN..=TIME_SKEW_MAX).contains(&time_diff) {
-                continue;
+            let is_boot_time = boot_time_cap_secs > 0 && timestamp < boot_time_cap_secs;
+            if !is_boot_time {
+                let time_diff = now - i64::from(timestamp);
+                if !(TIME_SKEW_MIN..=TIME_SKEW_MAX).contains(&time_diff) {
+                    continue;
+                }
             }
         }
         
-        return Some(TlsValidation {
-            user: user.clone(),
-            session_id,
-            digest,
-            timestamp,
-        });
+        if first_match.is_none() {
+            first_match = Some((user, timestamp));
+        }
     }
-    
-    None
-}
 
-fn curve25519_prime() -> BigUint {
-    (BigUint::one() << 255) - BigUint::from(19u32)
+    first_match.map(|(user, timestamp)| TlsValidation {
+        user: user.clone(),
+        session_id,
+        digest,
+        timestamp,
+    })
 }
 
 /// Generate a fake X25519 public key for TLS
 ///
-/// Produces a quadratic residue mod p = 2^255 - 19 by computing n² mod p,
-/// which matches Python/C behavior and avoids DPI fingerprinting.
+/// Uses RFC 7748 X25519 scalar multiplication over the canonical basepoint,
+/// yielding distribution-consistent public keys for anti-fingerprinting.
 pub fn gen_fake_x25519_key(rng: &SecureRandom) -> [u8; 32] {
-    let mut n_bytes = [0u8; 32];
-    n_bytes.copy_from_slice(&rng.bytes(32));
-
-    let n = BigUint::from_bytes_le(&n_bytes);
-    let p = curve25519_prime();
-    let pk = (&n * &n) % &p;
-
-    let mut out = pk.to_bytes_le();
-    out.resize(32, 0);
-    let mut result = [0u8; 32];
-    result.copy_from_slice(&out[..32]);
-    result
+    let mut scalar = [0u8; 32];
+    scalar.copy_from_slice(&rng.bytes(32));
+    x25519(scalar, X25519_BASEPOINT_BYTES)
 }
 
 /// Build TLS ServerHello response
@@ -400,7 +458,6 @@ pub fn build_server_hello(
     let server_hello = ServerHelloBuilder::new(session_id.to_vec())
         .with_x25519_key(&x25519_key)
         .with_tls13_version()
-        .with_alpn(alpn)
         .build_record();
     
     // Build Change Cipher Spec record
@@ -411,8 +468,27 @@ pub fn build_server_hello(
         0x01,       // CCS byte
     ];
     
-    // Build fake certificate (Application Data record)
-    let fake_cert = rng.bytes(fake_cert_len);
+    // Build first encrypted flight mimic as opaque ApplicationData bytes.
+    // Embed a compact EncryptedExtensions-like ALPN block when selected.
+    let mut fake_cert = Vec::with_capacity(fake_cert_len);
+    if let Some(proto) = alpn.as_ref().filter(|p| !p.is_empty() && p.len() <= u8::MAX as usize) {
+        let proto_list_len = 1usize + proto.len();
+        let ext_data_len = 2usize + proto_list_len;
+        let marker_len = 4usize + ext_data_len;
+        if marker_len <= fake_cert_len {
+            fake_cert.extend_from_slice(&0x0010u16.to_be_bytes());
+            fake_cert.extend_from_slice(&(ext_data_len as u16).to_be_bytes());
+            fake_cert.extend_from_slice(&(proto_list_len as u16).to_be_bytes());
+            fake_cert.push(proto.len() as u8);
+            fake_cert.extend_from_slice(proto);
+        }
+    }
+    if fake_cert.len() < fake_cert_len {
+        fake_cert.extend_from_slice(&rng.bytes(fake_cert_len - fake_cert.len()));
+    } else if fake_cert.len() > fake_cert_len {
+        fake_cert.truncate(fake_cert_len);
+    }
+
     let mut app_data_record = Vec::with_capacity(5 + fake_cert_len);
     app_data_record.push(TLS_RECORD_APPLICATION);
     app_data_record.extend_from_slice(&TLS_VERSION);
@@ -424,8 +500,9 @@ pub fn build_server_hello(
     // Build optional NewSessionTicket records (TLS 1.3 handshake messages are encrypted;
     // here we mimic with opaque ApplicationData records of plausible size).
     let mut tickets = Vec::new();
-    if new_session_tickets > 0 {
-        for _ in 0..new_session_tickets {
+    let ticket_count = new_session_tickets.min(4);
+    if ticket_count > 0 {
+        for _ in 0..ticket_count {
             let ticket_len: usize = rng.range(48) + 48; // 48-95 bytes
             let mut record = Vec::with_capacity(5 + ticket_len);
             record.push(TLS_RECORD_APPLICATION);
@@ -467,6 +544,11 @@ pub fn extract_sni_from_client_hello(handshake: &[u8]) -> Option<String> {
         return None;
     }
 
+    let record_len = u16::from_be_bytes([handshake[3], handshake[4]]) as usize;
+    if handshake.len() < 5 + record_len {
+        return None;
+    }
+
     let mut pos = 5; // after record header
     if handshake.get(pos).copied()? != 0x01 {
         return None; // not ClientHello
@@ -528,7 +610,9 @@ pub fn extract_sni_from_client_hello(handshake: &[u8]) -> Option<String> {
                 if name_type == 0 && name_len > 0
                     && let Ok(host) = std::str::from_utf8(&handshake[sn_pos..sn_pos + name_len])
                 {
-                    return Some(host.to_string());
+                    if is_valid_sni_hostname(host) {
+                        return Some(host.to_string());
+                    }
                 }
                 sn_pos += name_len;
             }
@@ -539,8 +623,46 @@ pub fn extract_sni_from_client_hello(handshake: &[u8]) -> Option<String> {
     None
 }
 
+fn is_valid_sni_hostname(host: &str) -> bool {
+    if host.is_empty() || host.len() > 253 {
+        return false;
+    }
+    if host.starts_with('.') || host.ends_with('.') {
+        return false;
+    }
+    if host.parse::<std::net::IpAddr>().is_ok() {
+        return false;
+    }
+
+    for label in host.split('.') {
+        if label.is_empty() || label.len() > 63 {
+            return false;
+        }
+        if label.starts_with('-') || label.ends_with('-') {
+            return false;
+        }
+        if !label
+            .bytes()
+            .all(|b| b.is_ascii_alphanumeric() || b == b'-')
+        {
+            return false;
+        }
+    }
+
+    true
+}
+
 /// Extract ALPN protocol list from ClientHello, return in offered order.
 pub fn extract_alpn_from_client_hello(handshake: &[u8]) -> Vec<Vec<u8>> {
+    if handshake.len() < 5 || handshake[0] != TLS_RECORD_HANDSHAKE {
+        return Vec::new();
+    }
+
+    let record_len = u16::from_be_bytes([handshake[3], handshake[4]]) as usize;
+    if handshake.len() < 5 + record_len {
+        return Vec::new();
+    }
+
     let mut pos = 5; // after record header
     if handshake.get(pos) != Some(&0x01) {
         return Vec::new();
@@ -592,13 +714,14 @@ pub fn is_tls_handshake(first_bytes: &[u8]) -> bool {
         return false;
     }
     
-    // TLS record header: 0x16 (handshake) 0x03 0x01 (TLS 1.0)
+    // TLS ClientHello commonly uses legacy record versions 0x0301 or 0x0303.
     first_bytes[0] == TLS_RECORD_HANDSHAKE 
         && first_bytes[1] == 0x03 
-        && first_bytes[2] == 0x01
+        && (first_bytes[2] == 0x01 || first_bytes[2] == 0x03)
 }
 
 /// Parse TLS record header, returns (record_type, length)
+
 pub fn parse_tls_record_header(header: &[u8; 5]) -> Option<(u8, u16)> {
     let record_type = header[0];
     let version = [header[1], header[2]];
@@ -667,291 +790,37 @@ fn validate_server_hello_structure(data: &[u8]) -> Result<(), ProxyError> {
     Ok(())
 }
 
-#[cfg(test)]
-mod tests {
-    use super::*;
-    
-    #[test]
-    fn test_is_tls_handshake() {
-        assert!(is_tls_handshake(&[0x16, 0x03, 0x01]));
-        assert!(is_tls_handshake(&[0x16, 0x03, 0x01, 0x02, 0x00]));
-        assert!(!is_tls_handshake(&[0x17, 0x03, 0x01])); // Application data
-        assert!(!is_tls_handshake(&[0x16, 0x03, 0x02])); // Wrong version
-        assert!(!is_tls_handshake(&[0x16, 0x03])); // Too short
-    }
-    
-    #[test]
-    fn test_parse_tls_record_header() {
-        let header = [0x16, 0x03, 0x01, 0x02, 0x00];
-        let result = parse_tls_record_header(&header).unwrap();
-        assert_eq!(result.0, TLS_RECORD_HANDSHAKE);
-        assert_eq!(result.1, 512);
-        
-        let header = [0x17, 0x03, 0x03, 0x40, 0x00];
-        let result = parse_tls_record_header(&header).unwrap();
-        assert_eq!(result.0, TLS_RECORD_APPLICATION);
-        assert_eq!(result.1, 16384);
-    }
-    
-    #[test]
-    fn test_gen_fake_x25519_key() {
-        let rng = SecureRandom::new();
-        let key1 = gen_fake_x25519_key(&rng);
-        let key2 = gen_fake_x25519_key(&rng);
-        
-        assert_eq!(key1.len(), 32);
-        assert_eq!(key2.len(), 32);
-        assert_ne!(key1, key2); // Should be random
-    }
+// ============= Compile-time Security Invariants =============
 
-    #[test]
-    fn test_fake_x25519_key_is_quadratic_residue() {
-        let rng = SecureRandom::new();
-        let key = gen_fake_x25519_key(&rng);
-        let p = curve25519_prime();
-        let k_num = BigUint::from_bytes_le(&key);
-        let exponent = (&p - BigUint::one()) >> 1;
-        let legendre = k_num.modpow(&exponent, &p);
-        assert_eq!(legendre, BigUint::one());
-    }
-    
-    #[test]
-    fn test_tls_extension_builder() {
-        let key = [0x42u8; 32];
-        
-        let mut builder = TlsExtensionBuilder::new();
-        builder.add_key_share(&key);
-        builder.add_supported_versions(0x0304);
-        
-        let result = builder.build();
-        
-        // Check length prefix
-        let len = u16::from_be_bytes([result[0], result[1]]) as usize;
-        assert_eq!(len, result.len() - 2);
-        
-        // Check key_share extension is present
-        assert!(result.len() > 40); // At least key share
-    }
-    
-    #[test]
-    fn test_server_hello_builder() {
-        let session_id = vec![0x01, 0x02, 0x03, 0x04];
-        let key = [0x55u8; 32];
-        
-        let builder = ServerHelloBuilder::new(session_id.clone())
-            .with_x25519_key(&key)
-            .with_tls13_version();
-        
-        let record = builder.build_record();
-        
-        // Validate structure
-        validate_server_hello_structure(&record).expect("Invalid ServerHello structure");
-        
-        // Check record type
-        assert_eq!(record[0], TLS_RECORD_HANDSHAKE);
-        
-        // Check version
-        assert_eq!(&record[1..3], &TLS_VERSION);
-        
-        // Check message type (ServerHello = 0x02)
-        assert_eq!(record[5], 0x02);
-    }
-    
-    #[test]
-    fn test_build_server_hello_structure() {
-        let secret = b"test secret";
-        let client_digest = [0x42u8; 32];
-        let session_id = vec![0xAA; 32];
-        
-        let rng = SecureRandom::new();
-        let response = build_server_hello(secret, &client_digest, &session_id, 2048, &rng, None, 0);
-        
-        // Should have at least 3 records
-        assert!(response.len() > 100);
-        
-        // First record should be ServerHello
-        assert_eq!(response[0], TLS_RECORD_HANDSHAKE);
-        
-        // Validate ServerHello structure
-        validate_server_hello_structure(&response).expect("Invalid ServerHello");
-        
-        // Find Change Cipher Spec
-        let server_hello_len = 5 + u16::from_be_bytes([response[3], response[4]]) as usize;
-        let ccs_start = server_hello_len;
-        
-        assert!(response.len() > ccs_start + 6);
-        assert_eq!(response[ccs_start], TLS_RECORD_CHANGE_CIPHER);
-        
-        // Find Application Data
-        let ccs_len = 5 + u16::from_be_bytes([response[ccs_start + 3], response[ccs_start + 4]]) as usize;
-        let app_start = ccs_start + ccs_len;
-        
-        assert!(response.len() > app_start + 5);
-        assert_eq!(response[app_start], TLS_RECORD_APPLICATION);
-    }
-    
-    #[test]
-    fn test_build_server_hello_digest() {
-        let secret = b"test secret key here";
-        let client_digest = [0x42u8; 32];
-        let session_id = vec![0xAA; 32];
-        
-        let rng = SecureRandom::new();
-        let response1 = build_server_hello(secret, &client_digest, &session_id, 1024, &rng, None, 0);
-        let response2 = build_server_hello(secret, &client_digest, &session_id, 1024, &rng, None, 0);
-        
-        // Digest position should have non-zero data
-        let digest1 = &response1[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN];
-        assert!(!digest1.iter().all(|&b| b == 0));
-        
-        // Different calls should have different digests (due to random cert)
-        let digest2 = &response2[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN];
-        assert_ne!(digest1, digest2);
-    }
-    
-    #[test]
-    fn test_server_hello_extensions_length() {
-        let session_id = vec![0x01; 32];
-        let key = [0x55u8; 32];
-        
-        let builder = ServerHelloBuilder::new(session_id)
-            .with_x25519_key(&key)
-            .with_tls13_version();
-        
-        let record = builder.build_record();
-        
-        // Parse to find extensions
-        let msg_start = 5; // After record header
-        let msg_len = u32::from_be_bytes([0, record[6], record[7], record[8]]) as usize;
-        
-        // Skip to session ID
-        let session_id_pos = msg_start + 4 + 2 + 32; // header(4) + version(2) + random(32)
-        let session_id_len = record[session_id_pos] as usize;
-        
-        // Skip to extensions
-        let ext_len_pos = session_id_pos + 1 + session_id_len + 2 + 1; // session_id + cipher(2) + compression(1)
-        let ext_len = u16::from_be_bytes([record[ext_len_pos], record[ext_len_pos + 1]]) as usize;
-        
-        // Verify extensions length matches actual data
-        let extensions_data = &record[ext_len_pos + 2..msg_start + 4 + msg_len];
-        assert_eq!(ext_len, extensions_data.len(), 
-            "Extension length mismatch: declared {}, actual {}", ext_len, extensions_data.len());
-    }
-    
-    #[test]
-    fn test_validate_tls_handshake_format() {
-        // Build a minimal ClientHello-like structure
-        let mut handshake = vec![0u8; 100];
-        
-        // Put a valid-looking digest at position 11
-        handshake[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN]
-            .copy_from_slice(&[0x42; 32]);
-        
-        // Session ID length
-        handshake[TLS_DIGEST_POS + TLS_DIGEST_LEN] = 32;
-        
-        // This won't validate (wrong HMAC) but shouldn't panic
-        let secrets = vec![("test".to_string(), b"secret".to_vec())];
-        let result = validate_tls_handshake(&handshake, &secrets, true);
-        
-        // Should return None (no match) but not panic
-        assert!(result.is_none());
-    }
+/// Compile-time checks that enforce invariants the rest of the code relies on.
+/// Using `static_assertions` ensures these can never silently break across
+/// refactors without a compile error.
+mod compile_time_security_checks {
+    use super::{TLS_DIGEST_LEN, TLS_DIGEST_HALF_LEN};
+    use static_assertions::const_assert;
 
-    fn build_client_hello_with_exts(exts: Vec<(u16, Vec<u8>)>, host: &str) -> Vec<u8> {
-        let mut body = Vec::new();
-        body.extend_from_slice(&TLS_VERSION); // legacy version
-        body.extend_from_slice(&[0u8; 32]); // random
-        body.push(0); // session id len
-        body.extend_from_slice(&2u16.to_be_bytes()); // cipher suites len
-        body.extend_from_slice(&[0x13, 0x01]); // TLS_AES_128_GCM_SHA256
-        body.push(1); // compression len
-        body.push(0); // null compression
+    // The digest must be exactly one SHA-256 output.
+    const_assert!(TLS_DIGEST_LEN == 32);
 
-        // Build SNI extension
-        let host_bytes = host.as_bytes();
-        let mut sni_ext = Vec::new();
-        sni_ext.extend_from_slice(&(host_bytes.len() as u16 + 3).to_be_bytes());
-        sni_ext.push(0);
-        sni_ext.extend_from_slice(&(host_bytes.len() as u16).to_be_bytes());
-        sni_ext.extend_from_slice(host_bytes);
+    // Replay-dedup stores the first half; verify it is literally half.
+    const_assert!(TLS_DIGEST_HALF_LEN * 2 == TLS_DIGEST_LEN);
 
-        let mut ext_blob = Vec::new();
-        for (typ, data) in exts {
-            ext_blob.extend_from_slice(&typ.to_be_bytes());
-            ext_blob.extend_from_slice(&(data.len() as u16).to_be_bytes());
-            ext_blob.extend_from_slice(&data);
-        }
-        // SNI last
-        ext_blob.extend_from_slice(&0x0000u16.to_be_bytes());
-        ext_blob.extend_from_slice(&(sni_ext.len() as u16).to_be_bytes());
-        ext_blob.extend_from_slice(&sni_ext);
-
-        body.extend_from_slice(&(ext_blob.len() as u16).to_be_bytes());
-        body.extend_from_slice(&ext_blob);
-
-        let mut handshake = Vec::new();
-        handshake.push(0x01); // ClientHello
-        let len_bytes = (body.len() as u32).to_be_bytes();
-        handshake.extend_from_slice(&len_bytes[1..4]);
-        handshake.extend_from_slice(&body);
-
-        let mut record = Vec::new();
-        record.push(TLS_RECORD_HANDSHAKE);
-        record.extend_from_slice(&[0x03, 0x01]);
-        record.extend_from_slice(&(handshake.len() as u16).to_be_bytes());
-        record.extend_from_slice(&handshake);
-        record
-    }
-
-    #[test]
-    fn test_extract_sni_with_grease_extension() {
-        // GREASE type 0x0a0a with zero length before SNI
-        let ch = build_client_hello_with_exts(vec![(0x0a0a, Vec::new())], "example.com");
-        let sni = extract_sni_from_client_hello(&ch);
-        assert_eq!(sni.as_deref(), Some("example.com"));
-    }
-
-    #[test]
-    fn test_extract_sni_tolerates_empty_unknown_extension() {
-        let ch = build_client_hello_with_exts(vec![(0x1234, Vec::new())], "test.local");
-        let sni = extract_sni_from_client_hello(&ch);
-        assert_eq!(sni.as_deref(), Some("test.local"));
-    }
-
-    #[test]
-    fn test_extract_alpn_single() {
-        let mut alpn_data = Vec::new();
-        // list length = 3 (1 length byte + "h2")
-        alpn_data.extend_from_slice(&3u16.to_be_bytes());
-        alpn_data.push(2);
-        alpn_data.extend_from_slice(b"h2");
-        let ch = build_client_hello_with_exts(vec![(0x0010, alpn_data)], "alpn.test");
-        let alpn = extract_alpn_from_client_hello(&ch);
-        let alpn_str: Vec<String> = alpn
-            .iter()
-            .map(|p| std::str::from_utf8(p).unwrap().to_string())
-            .collect();
-        assert_eq!(alpn_str, vec!["h2"]);
-    }
-
-    #[test]
-    fn test_extract_alpn_multiple() {
-        let mut alpn_data = Vec::new();
-        // list length = 11 (sum of per-proto lengths including length bytes)
-        alpn_data.extend_from_slice(&11u16.to_be_bytes());
-        alpn_data.push(2);
-        alpn_data.extend_from_slice(b"h2");
-        alpn_data.push(4);
-        alpn_data.extend_from_slice(b"spdy");
-        alpn_data.push(2);
-        alpn_data.extend_from_slice(b"h3");
-        let ch = build_client_hello_with_exts(vec![(0x0010, alpn_data)], "alpn.test");
-        let alpn = extract_alpn_from_client_hello(&ch);
-        let alpn_str: Vec<String> = alpn
-            .iter()
-            .map(|p| std::str::from_utf8(p).unwrap().to_string())
-            .collect();
-        assert_eq!(alpn_str, vec!["h2", "spdy", "h3"]);
-    }
+    // The HMAC check window (28 bytes) plus the embedded timestamp (4 bytes)
+    // must exactly fill the digest.  If TLS_DIGEST_LEN ever changes, these
+    // assertions will catch the mismatch before any timing-oracle fix is broke.
+    const_assert!(28 + 4 == TLS_DIGEST_LEN);
 }
+
+// ============= Security-focused regression tests =============
+
+#[cfg(test)]
+#[path = "tls_security_tests.rs"]
+mod security_tests;
+
+#[cfg(test)]
+#[path = "tls_adversarial_tests.rs"]
+mod adversarial_tests;
+
+#[cfg(test)]
+#[path = "tls_fuzz_security_tests.rs"]
+mod fuzz_security_tests;

src/protocol/tls_adversarial_tests.rs

@@ -0,0 +1,352 @@
+use super::*;
+use std::time::Instant;
+use crate::crypto::sha256_hmac;
+
+/// Helper to create a byte vector of specific length.
+fn make_garbage(len: usize) -> Vec<u8> {
+    vec![0x42u8; len]
+}
+
+/// Helper to create a valid-looking HMAC digest for test.
+fn make_digest(secret: &[u8], msg: &[u8], ts: u32) -> [u8; 32] {
+    let mut hmac = sha256_hmac(secret, msg);
+    let ts_bytes = ts.to_le_bytes();
+    for i in 0..4 {
+        hmac[28 + i] ^= ts_bytes[i];
+    }
+    hmac
+}
+
+fn make_valid_tls_handshake_with_session_id(
+    secret: &[u8],
+    timestamp: u32,
+    session_id: &[u8],
+) -> Vec<u8> {
+    let session_id_len = session_id.len();
+    let len = TLS_DIGEST_POS + TLS_DIGEST_LEN + 1 + session_id_len;
+    let mut handshake = vec![0x42u8; len];
+
+    handshake[TLS_DIGEST_POS + TLS_DIGEST_LEN] = session_id_len as u8;
+    let sid_start = TLS_DIGEST_POS + TLS_DIGEST_LEN + 1;
+    handshake[sid_start..sid_start + session_id_len].copy_from_slice(session_id);
+    handshake[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN].fill(0);
+
+    let digest = make_digest(secret, &handshake, timestamp);
+
+    handshake[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN]
+        .copy_from_slice(&digest);
+    handshake
+}
+
+fn make_valid_tls_handshake(secret: &[u8], timestamp: u32) -> Vec<u8> {
+    make_valid_tls_handshake_with_session_id(secret, timestamp, &[0x42; 32])
+}
+
+// ------------------------------------------------------------------
+// Truncated Packet Tests (OWASP ASVS 5.1.4, 5.1.5)
+// ------------------------------------------------------------------
+
+#[test]
+fn validate_tls_handshake_truncated_10_bytes_rejected() {
+    let secrets = vec![("user".to_string(), b"secret".to_vec())];
+    let truncated = make_garbage(10);
+    assert!(validate_tls_handshake(&truncated, &secrets, true).is_none());
+}
+
+#[test]
+fn validate_tls_handshake_truncated_at_digest_start_rejected() {
+    let secrets = vec![("user".to_string(), b"secret".to_vec())];
+    // TLS_DIGEST_POS = 11. 11 bytes should be rejected.
+    let truncated = make_garbage(TLS_DIGEST_POS);
+    assert!(validate_tls_handshake(&truncated, &secrets, true).is_none());
+}
+
+#[test]
+fn validate_tls_handshake_truncated_inside_digest_rejected() {
+    let secrets = vec![("user".to_string(), b"secret".to_vec())];
+    // TLS_DIGEST_POS + 16 (half digest)
+    let truncated = make_garbage(TLS_DIGEST_POS + 16);
+    assert!(validate_tls_handshake(&truncated, &secrets, true).is_none());
+}
+
+#[test]
+fn extract_sni_truncated_at_record_header_rejected() {
+    let truncated = make_garbage(3);
+    assert!(extract_sni_from_client_hello(&truncated).is_none());
+}
+
+#[test]
+fn extract_sni_truncated_at_handshake_header_rejected() {
+    let mut truncated = vec![TLS_RECORD_HANDSHAKE, 0x03, 0x03, 0x00, 0x05];
+    truncated.extend_from_slice(&[0x01, 0x00]); // ClientHello type but truncated length
+    assert!(extract_sni_from_client_hello(&truncated).is_none());
+}
+
+// ------------------------------------------------------------------
+// Malformed Extension Parsing Tests
+// ------------------------------------------------------------------
+
+#[test]
+fn extract_sni_with_overlapping_extension_lengths_rejected() {
+    let mut h = vec![0x16, 0x03, 0x03, 0x00, 0x60]; // Record header
+    h.push(0x01); // Handshake type: ClientHello
+    h.extend_from_slice(&[0x00, 0x00, 0x5C]); // Length: 92
+    h.extend_from_slice(&[0x03, 0x03]); // Version
+    h.extend_from_slice(&[0u8; 32]); // Random
+    h.push(0); // Session ID length: 0
+    h.extend_from_slice(&[0x00, 0x02, 0x13, 0x01]); // Cipher suites
+    h.extend_from_slice(&[0x01, 0x00]); // Compression
+    
+    // Extensions start
+    h.extend_from_slice(&[0x00, 0x20]); // Total Extensions length: 32
+    
+    // Extension 1: SNI (type 0)
+    h.extend_from_slice(&[0x00, 0x00]); 
+    h.extend_from_slice(&[0x00, 0x40]); // Claimed len: 64 (OVERFLOWS total extensions len 32)
+    h.extend_from_slice(&[0u8; 64]);
+    
+    assert!(extract_sni_from_client_hello(&h).is_none());
+}
+
+#[test]
+fn extract_sni_with_infinite_loop_potential_extension_rejected() {
+    let mut h = vec![0x16, 0x03, 0x03, 0x00, 0x60]; // Record header
+    h.push(0x01); // Handshake type: ClientHello
+    h.extend_from_slice(&[0x00, 0x00, 0x5C]); // Length: 92
+    h.extend_from_slice(&[0x03, 0x03]); // Version
+    h.extend_from_slice(&[0u8; 32]); // Random
+    h.push(0); // Session ID length: 0
+    h.extend_from_slice(&[0x00, 0x02, 0x13, 0x01]); // Cipher suites
+    h.extend_from_slice(&[0x01, 0x00]); // Compression
+    
+    // Extensions start
+    h.extend_from_slice(&[0x00, 0x10]); // Total Extensions length: 16
+    
+    // Extension: zero length but claims more? 
+    // If our parser didn't advance, it might loop.
+    // Telemt uses `pos += 4 + elen;` so it always advances.
+    h.extend_from_slice(&[0x12, 0x34]); // Unknown type
+    h.extend_from_slice(&[0x00, 0x00]); // Length 0
+    
+    // Fill the rest with garbage
+    h.extend_from_slice(&[0x42; 12]);
+    
+    // We expect it to finish without SNI found
+    assert!(extract_sni_from_client_hello(&h).is_none());
+}
+
+#[test]
+fn extract_sni_with_invalid_hostname_rejected() {
+    let host = b"invalid_host!%^";
+    let mut sni = Vec::new();
+    sni.extend_from_slice(&((host.len() + 3) as u16).to_be_bytes());
+    sni.push(0);
+    sni.extend_from_slice(&(host.len() as u16).to_be_bytes());
+    sni.extend_from_slice(host);
+    
+    let mut h = vec![0x16, 0x03, 0x03, 0x00, 0x60]; // Record header
+    h.push(0x01); // ClientHello
+    h.extend_from_slice(&[0x00, 0x00, 0x5C]);
+    h.extend_from_slice(&[0x03, 0x03]);
+    h.extend_from_slice(&[0u8; 32]);
+    h.push(0);
+    h.extend_from_slice(&[0x00, 0x02, 0x13, 0x01]);
+    h.extend_from_slice(&[0x01, 0x00]);
+    
+    let mut ext = Vec::new();
+    ext.extend_from_slice(&0x0000u16.to_be_bytes());
+    ext.extend_from_slice(&(sni.len() as u16).to_be_bytes());
+    ext.extend_from_slice(&sni);
+    
+    h.extend_from_slice(&(ext.len() as u16).to_be_bytes());
+    h.extend_from_slice(&ext);
+    
+    assert!(extract_sni_from_client_hello(&h).is_none(), "Invalid SNI hostname must be rejected");
+}
+
+// ------------------------------------------------------------------
+// Timing Neutrality Tests (OWASP ASVS 5.1.7)
+// ------------------------------------------------------------------
+
+#[test]
+fn validate_tls_handshake_timing_neutrality() {
+    let secret = b"timing_test_secret_32_bytes_long_";
+    let secrets = vec![("u".to_string(), secret.to_vec())];
+
+    let mut base = vec![0x42u8; 100];
+    base[TLS_DIGEST_POS + TLS_DIGEST_LEN] = 32;
+
+    const ITER: usize = 600;
+    const ROUNDS: usize = 7;
+
+    let mut per_round_avg_diff_ns = Vec::with_capacity(ROUNDS);
+
+    for round in 0..ROUNDS {
+        let mut success_h = base.clone();
+        let mut fail_h = base.clone();
+
+        let start_success = Instant::now();
+        for _ in 0..ITER {
+            let digest = make_digest(secret, &success_h, 0);
+            success_h[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN].copy_from_slice(&digest);
+            let _ = validate_tls_handshake_at_time(&success_h, &secrets, true, 0);
+        }
+        let success_elapsed = start_success.elapsed();
+
+        let start_fail = Instant::now();
+        for i in 0..ITER {
+            let mut digest = make_digest(secret, &fail_h, 0);
+            let flip_idx = (i + round) % (TLS_DIGEST_LEN - 4);
+            digest[flip_idx] ^= 0xFF;
+            fail_h[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN].copy_from_slice(&digest);
+            let _ = validate_tls_handshake_at_time(&fail_h, &secrets, true, 0);
+        }
+        let fail_elapsed = start_fail.elapsed();
+
+        let diff = if success_elapsed > fail_elapsed {
+            success_elapsed - fail_elapsed
+        } else {
+            fail_elapsed - success_elapsed
+        };
+        per_round_avg_diff_ns.push(diff.as_nanos() as f64 / ITER as f64);
+    }
+
+    per_round_avg_diff_ns.sort_by(|a, b| a.partial_cmp(b).unwrap());
+    let median_avg_diff_ns = per_round_avg_diff_ns[ROUNDS / 2];
+
+    // Keep this as a coarse side-channel guard only; noisy shared CI hosts can
+    // introduce microsecond-level jitter that should not fail deterministic suites.
+    assert!(
+        median_avg_diff_ns < 50_000.0,
+        "Median timing delta too large: {} ns/iter",
+        median_avg_diff_ns
+    );
+}
+
+// ------------------------------------------------------------------
+// Adversarial Fingerprinting / Active Probing Tests
+// ------------------------------------------------------------------
+
+#[test]
+fn is_tls_handshake_robustness_against_probing() {
+    // Valid TLS 1.0 ClientHello
+    assert!(is_tls_handshake(&[0x16, 0x03, 0x01]));
+    // Valid TLS 1.2/1.3 ClientHello (Legacy Record Layer)
+    assert!(is_tls_handshake(&[0x16, 0x03, 0x03]));
+    
+    // Invalid record type but matching version
+    assert!(!is_tls_handshake(&[0x17, 0x03, 0x03]));
+    // Plaintext HTTP request
+    assert!(!is_tls_handshake(b"GET / HTTP/1.1"));
+    // Short garbage
+    assert!(!is_tls_handshake(&[0x16, 0x03]));
+}
+
+#[test]
+fn validate_tls_handshake_at_time_strict_boundary() {
+    let secret = b"strict_boundary_secret_32_bytes_";
+    let secrets = vec![("u".to_string(), secret.to_vec())];
+    let now: i64 = 1_000_000_000;
+    
+    // Boundary: exactly TIME_SKEW_MAX (120s past)
+    let ts_past = (now - TIME_SKEW_MAX) as u32;
+    let h = make_valid_tls_handshake_with_session_id(secret, ts_past, &[0x42; 32]);
+    assert!(validate_tls_handshake_at_time(&h, &secrets, false, now).is_some());
+    
+    // Boundary + 1s: should be rejected
+    let ts_too_past = (now - TIME_SKEW_MAX - 1) as u32;
+    let h2 = make_valid_tls_handshake_with_session_id(secret, ts_too_past, &[0x42; 32]);
+    assert!(validate_tls_handshake_at_time(&h2, &secrets, false, now).is_none());
+}
+
+#[test]
+fn extract_sni_with_duplicate_extensions_rejected() {
+    // Construct a ClientHello with TWO SNI extensions
+    let host1 = b"first.com";
+    let mut sni1 = Vec::new();
+    sni1.extend_from_slice(&((host1.len() + 3) as u16).to_be_bytes());
+    sni1.push(0);
+    sni1.extend_from_slice(&(host1.len() as u16).to_be_bytes());
+    sni1.extend_from_slice(host1);
+    
+    let host2 = b"second.com";
+    let mut sni2 = Vec::new();
+    sni2.extend_from_slice(&((host2.len() + 3) as u16).to_be_bytes());
+    sni2.push(0);
+    sni2.extend_from_slice(&(host2.len() as u16).to_be_bytes());
+    sni2.extend_from_slice(host2);
+    
+    let mut ext = Vec::new();
+    // Ext 1: SNI
+    ext.extend_from_slice(&0x0000u16.to_be_bytes());
+    ext.extend_from_slice(&(sni1.len() as u16).to_be_bytes());
+    ext.extend_from_slice(&sni1);
+    // Ext 2: SNI again
+    ext.extend_from_slice(&0x0000u16.to_be_bytes());
+    ext.extend_from_slice(&(sni2.len() as u16).to_be_bytes());
+    ext.extend_from_slice(&sni2);
+    
+    let mut body = Vec::new();
+    body.extend_from_slice(&[0x03, 0x03]);
+    body.extend_from_slice(&[0u8; 32]);
+    body.push(0);
+    body.extend_from_slice(&[0x00, 0x02, 0x13, 0x01]);
+    body.extend_from_slice(&[0x01, 0x00]);
+    body.extend_from_slice(&(ext.len() as u16).to_be_bytes());
+    body.extend_from_slice(&ext);
+
+    let mut handshake = Vec::new();
+    handshake.push(0x01);
+    let body_len = (body.len() as u32).to_be_bytes();
+    handshake.extend_from_slice(&body_len[1..4]);
+    handshake.extend_from_slice(&body);
+
+    let mut h = Vec::new();
+    h.push(0x16);
+    h.extend_from_slice(&[0x03, 0x03]);
+    h.extend_from_slice(&(handshake.len() as u16).to_be_bytes());
+    h.extend_from_slice(&handshake);
+    
+    // Parser might return first, see second, or fail. OWASP ASVS prefers rejection of unexpected dups.
+    // Telemt's `extract_sni` returns the first one found.
+    assert!(extract_sni_from_client_hello(&h).is_some()); 
+}
+
+#[test]
+fn extract_alpn_with_malformed_list_rejected() {
+    let mut alpn_payload = Vec::new();
+    alpn_payload.extend_from_slice(&0x0005u16.to_be_bytes()); // Total len 5
+    alpn_payload.push(10); // Labeled len 10 (OVERFLOWS total 5)
+    alpn_payload.extend_from_slice(b"h2");
+    
+    let mut ext = Vec::new();
+    ext.extend_from_slice(&0x0010u16.to_be_bytes()); // Type: ALPN (16)
+    ext.extend_from_slice(&(alpn_payload.len() as u16).to_be_bytes());
+    ext.extend_from_slice(&alpn_payload);
+    
+    let mut h = vec![0x16, 0x03, 0x03, 0x00, 0x40, 0x01, 0x00, 0x00, 0x3C, 0x03, 0x03];
+    h.extend_from_slice(&[0u8; 32]);
+    h.push(0);
+    h.extend_from_slice(&[0x00, 0x02, 0x13, 0x01, 0x01, 0x00]);
+    h.extend_from_slice(&(ext.len() as u16).to_be_bytes());
+    h.extend_from_slice(&ext);
+    
+    let res = extract_alpn_from_client_hello(&h);
+    assert!(res.is_empty(), "Malformed ALPN list must return empty or fail");
+}
+
+#[test]
+fn extract_sni_with_huge_extension_header_rejected() {
+    let mut h = vec![0x16, 0x03, 0x03, 0x00, 0x00]; // Record header
+    h.push(0x01); // ClientHello
+    h.extend_from_slice(&[0x00, 0xFF, 0xFF]); // Huge length (65535) - overflows record
+    h.extend_from_slice(&[0x03, 0x03]);
+    h.extend_from_slice(&[0u8; 32]);
+    h.push(0);
+    h.extend_from_slice(&[0x00, 0x02, 0x13, 0x01, 0x01, 0x00]);
+    
+    // Extensions start
+    h.extend_from_slice(&[0xFF, 0xFF]); // Total extensions: 65535 (OVERFLOWS everything)
+    
+    assert!(extract_sni_from_client_hello(&h).is_none());
+}

src/protocol/tls_fuzz_security_tests.rs

@@ -0,0 +1,195 @@
+use super::*;
+use crate::crypto::sha256_hmac;
+use std::panic::catch_unwind;
+
+fn make_valid_tls_handshake_with_session_id(
+    secret: &[u8],
+    timestamp: u32,
+    session_id: &[u8],
+) -> Vec<u8> {
+    let session_id_len = session_id.len();
+    assert!(session_id_len <= u8::MAX as usize);
+
+    let len = TLS_DIGEST_POS + TLS_DIGEST_LEN + 1 + session_id_len;
+    let mut handshake = vec![0x42u8; len];
+    handshake[TLS_DIGEST_POS + TLS_DIGEST_LEN] = session_id_len as u8;
+    let sid_start = TLS_DIGEST_POS + TLS_DIGEST_LEN + 1;
+    handshake[sid_start..sid_start + session_id_len].copy_from_slice(session_id);
+    handshake[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN].fill(0);
+
+    let mut digest = sha256_hmac(secret, &handshake);
+    let ts = timestamp.to_le_bytes();
+    for idx in 0..4 {
+        digest[28 + idx] ^= ts[idx];
+    }
+
+    handshake[TLS_DIGEST_POS..TLS_DIGEST_POS + TLS_DIGEST_LEN].copy_from_slice(&digest);
+    handshake
+}
+
+fn make_valid_client_hello_record(host: &str, alpn_protocols: &[&[u8]]) -> Vec<u8> {
+    let mut body = Vec::new();
+    body.extend_from_slice(&TLS_VERSION);
+    body.extend_from_slice(&[0u8; 32]);
+    body.push(0);
+    body.extend_from_slice(&2u16.to_be_bytes());
+    body.extend_from_slice(&[0x13, 0x01]);
+    body.push(1);
+    body.push(0);
+
+    let mut ext_blob = Vec::new();
+
+    let host_bytes = host.as_bytes();
+    let mut sni_payload = Vec::new();
+    sni_payload.extend_from_slice(&((host_bytes.len() + 3) as u16).to_be_bytes());
+    sni_payload.push(0);
+    sni_payload.extend_from_slice(&(host_bytes.len() as u16).to_be_bytes());
+    sni_payload.extend_from_slice(host_bytes);
+    ext_blob.extend_from_slice(&0x0000u16.to_be_bytes());
+    ext_blob.extend_from_slice(&(sni_payload.len() as u16).to_be_bytes());
+    ext_blob.extend_from_slice(&sni_payload);
+
+    if !alpn_protocols.is_empty() {
+        let mut alpn_list = Vec::new();
+        for proto in alpn_protocols {
+            alpn_list.push(proto.len() as u8);
+            alpn_list.extend_from_slice(proto);
+        }
+        let mut alpn_data = Vec::new();
+        alpn_data.extend_from_slice(&(alpn_list.len() as u16).to_be_bytes());
+        alpn_data.extend_from_slice(&alpn_list);
+
+        ext_blob.extend_from_slice(&0x0010u16.to_be_bytes());
+        ext_blob.extend_from_slice(&(alpn_data.len() as u16).to_be_bytes());
+        ext_blob.extend_from_slice(&alpn_data);
+    }
+
+    body.extend_from_slice(&(ext_blob.len() as u16).to_be_bytes());
+    body.extend_from_slice(&ext_blob);
+
+    let mut handshake = Vec::new();
+    handshake.push(0x01);
+    let body_len = (body.len() as u32).to_be_bytes();
+    handshake.extend_from_slice(&body_len[1..4]);
+    handshake.extend_from_slice(&body);
+
+    let mut record = Vec::new();
+    record.push(TLS_RECORD_HANDSHAKE);
+    record.extend_from_slice(&[0x03, 0x01]);
+    record.extend_from_slice(&(handshake.len() as u16).to_be_bytes());
+    record.extend_from_slice(&handshake);
+    record
+}
+
+#[test]
+fn client_hello_fuzz_corpus_never_panics_or_accepts_corruption() {
+    let valid = make_valid_client_hello_record("example.com", &[b"h2", b"http/1.1"]);
+    assert_eq!(extract_sni_from_client_hello(&valid).as_deref(), Some("example.com"));
+    assert_eq!(
+        extract_alpn_from_client_hello(&valid),
+        vec![b"h2".to_vec(), b"http/1.1".to_vec()]
+    );
+    assert!(
+        extract_sni_from_client_hello(&make_valid_client_hello_record("127.0.0.1", &[])).is_none(),
+        "literal IP hostnames must be rejected"
+    );
+
+    let mut corpus = vec![
+        Vec::new(),
+        vec![0x16, 0x03, 0x03],
+        valid[..9].to_vec(),
+        valid[..valid.len() - 1].to_vec(),
+    ];
+
+    let mut wrong_type = valid.clone();
+    wrong_type[0] = 0x15;
+    corpus.push(wrong_type);
+
+    let mut wrong_handshake = valid.clone();
+    wrong_handshake[5] = 0x02;
+    corpus.push(wrong_handshake);
+
+    let mut wrong_length = valid.clone();
+    wrong_length[3] ^= 0x7f;
+    corpus.push(wrong_length);
+
+    for (idx, input) in corpus.iter().enumerate() {
+        assert!(catch_unwind(|| extract_sni_from_client_hello(input)).is_ok());
+        assert!(catch_unwind(|| extract_alpn_from_client_hello(input)).is_ok());
+
+        if idx == 0 {
+            continue;
+        }
+
+        assert!(extract_sni_from_client_hello(input).is_none(), "corpus item {idx} must fail closed for SNI");
+        assert!(extract_alpn_from_client_hello(input).is_empty(), "corpus item {idx} must fail closed for ALPN");
+    }
+}
+
+#[test]
+fn tls_handshake_fuzz_corpus_never_panics_and_rejects_digest_mutations() {
+    let secret = b"tls_fuzz_security_secret";
+    let now: i64 = 1_700_000_000;
+    let base = make_valid_tls_handshake_with_session_id(secret, now as u32, &[0x42; 32]);
+    let secrets = vec![("fuzz-user".to_string(), secret.to_vec())];
+
+    assert!(validate_tls_handshake_at_time(&base, &secrets, false, now).is_some());
+
+    let mut corpus = Vec::new();
+
+    let mut truncated = base.clone();
+    truncated.truncate(TLS_DIGEST_POS + 16);
+    corpus.push(truncated);
+
+    let mut digest_flip = base.clone();
+    digest_flip[TLS_DIGEST_POS + 7] ^= 0x80;
+    corpus.push(digest_flip);
+
+    let mut session_id_len_overflow = base.clone();
+    session_id_len_overflow[TLS_DIGEST_POS + TLS_DIGEST_LEN] = 33;
+    corpus.push(session_id_len_overflow);
+
+    let mut timestamp_far_past = base.clone();
+    timestamp_far_past[TLS_DIGEST_POS + 28..TLS_DIGEST_POS + 32]
+        .copy_from_slice(&((now - i64::from(TIME_SKEW_MAX) - 1) as u32).to_le_bytes());
+    corpus.push(timestamp_far_past);
+
+    let mut timestamp_far_future = base.clone();
+    timestamp_far_future[TLS_DIGEST_POS + 28..TLS_DIGEST_POS + 32]
+        .copy_from_slice(&((now - TIME_SKEW_MIN + 1) as u32).to_le_bytes());
+    corpus.push(timestamp_far_future);
+
+    let mut seed = 0xA5A5_5A5A_F00D_BAAD_u64;
+    for _ in 0..32 {
+        let mut mutated = base.clone();
+        for _ in 0..2 {
+            seed = seed.wrapping_mul(2862933555777941757).wrapping_add(3037000493);
+            let idx = TLS_DIGEST_POS + (seed as usize % TLS_DIGEST_LEN);
+            mutated[idx] ^= ((seed >> 17) as u8).wrapping_add(1);
+        }
+        corpus.push(mutated);
+    }
+
+    for (idx, handshake) in corpus.iter().enumerate() {
+        let result = catch_unwind(|| validate_tls_handshake_at_time(handshake, &secrets, false, now));
+        assert!(result.is_ok(), "corpus item {idx} must not panic");
+        assert!(result.unwrap().is_none(), "corpus item {idx} must fail closed");
+    }
+}
+
+#[test]
+fn tls_boot_time_acceptance_is_capped_by_replay_window() {
+    let secret = b"tls_boot_time_cap_secret";
+    let secrets = vec![("boot-user".to_string(), secret.to_vec())];
+    let boot_ts = 1u32;
+    let handshake = make_valid_tls_handshake_with_session_id(secret, boot_ts, &[0x42; 32]);
+
+    assert!(
+        validate_tls_handshake_with_replay_window(&handshake, &secrets, false, 300).is_some(),
+        "boot-time timestamp should be accepted while replay window permits it"
+    );
+    assert!(
+        validate_tls_handshake_with_replay_window(&handshake, &secrets, false, 0).is_none(),
+        "boot-time timestamp must be rejected when replay window disables the bypass"
+    );
+}

src/proxy/client.rs

@@ -4,7 +4,10 @@ use std::future::Future;
 use std::net::{IpAddr, SocketAddr};
 use std::pin::Pin;
 use std::sync::Arc;
+use std::sync::OnceLock;
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::time::Duration;
+use ipnetwork::IpNetwork;
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite};
 use tokio::net::TcpStream;
 use tokio::time::timeout;
@@ -21,9 +24,50 @@ enum HandshakeOutcome {
     Handled,
 }
 
+#[must_use = "UserConnectionReservation must be kept alive to retain user/IP reservation until release or drop"]
+struct UserConnectionReservation {
+    stats: Arc<Stats>,
+    ip_tracker: Arc<UserIpTracker>,
+    user: String,
+    ip: IpAddr,
+    active: bool,
+}
+
+impl UserConnectionReservation {
+    fn new(stats: Arc<Stats>, ip_tracker: Arc<UserIpTracker>, user: String, ip: IpAddr) -> Self {
+        Self {
+            stats,
+            ip_tracker,
+            user,
+            ip,
+            active: true,
+        }
+    }
+
+    async fn release(mut self) {
+        if !self.active {
+            return;
+        }
+        self.ip_tracker.remove_ip(&self.user, self.ip).await;
+        self.active = false;
+        self.stats.decrement_user_curr_connects(&self.user);
+    }
+}
+
+impl Drop for UserConnectionReservation {
+    fn drop(&mut self) {
+        if !self.active {
+            return;
+        }
+        self.active = false;
+        self.stats.decrement_user_curr_connects(&self.user);
+        self.ip_tracker.enqueue_cleanup(self.user.clone(), self.ip);
+    }
+}
+
 use crate::config::ProxyConfig;
 use crate::crypto::SecureRandom;
-use crate::error::{HandshakeResult, ProxyError, Result};
+use crate::error::{HandshakeResult, ProxyError, Result, StreamError};
 use crate::ip_tracker::UserIpTracker;
 use crate::protocol::constants::*;
 use crate::protocol::tls;
@@ -42,7 +86,19 @@ use crate::proxy::middle_relay::handle_via_middle_proxy;
 use crate::proxy::route_mode::{RelayRouteMode, RouteRuntimeController};
 
 fn beobachten_ttl(config: &ProxyConfig) -> Duration {
-    Duration::from_secs(config.general.beobachten_minutes.saturating_mul(60))
+    let minutes = config.general.beobachten_minutes;
+    if minutes == 0 {
+        static BEOBACHTEN_ZERO_MINUTES_WARNED: OnceLock<AtomicBool> = OnceLock::new();
+        let warned = BEOBACHTEN_ZERO_MINUTES_WARNED.get_or_init(|| AtomicBool::new(false));
+        if !warned.swap(true, Ordering::Relaxed) {
+            warn!(
+                "general.beobachten_minutes=0 is insecure because entries expire immediately; forcing minimum TTL to 1 minute"
+            );
+        }
+        return Duration::from_secs(60);
+    }
+
+    Duration::from_secs(minutes.saturating_mul(60))
 }
 
 fn record_beobachten_class(
@@ -63,14 +119,34 @@ fn record_handshake_failure_class(
     peer_ip: IpAddr,
     error: &ProxyError,
 ) {
-    let class = if error.to_string().contains("expected 64 bytes, got 0") {
-        "expected_64_got_0"
-    } else {
-        "other"
+    let class = match error {
+        ProxyError::Io(err) if err.kind() == std::io::ErrorKind::UnexpectedEof => {
+            "expected_64_got_0"
+        }
+        ProxyError::Stream(StreamError::UnexpectedEof) => "expected_64_got_0",
+        _ => "other",
     };
     record_beobachten_class(beobachten, config, peer_ip, class);
 }
 
+fn is_trusted_proxy_source(peer_ip: IpAddr, trusted: &[IpNetwork]) -> bool {
+    if trusted.is_empty() {
+        static EMPTY_PROXY_TRUST_WARNED: OnceLock<AtomicBool> = OnceLock::new();
+        let warned = EMPTY_PROXY_TRUST_WARNED.get_or_init(|| AtomicBool::new(false));
+        if !warned.swap(true, Ordering::Relaxed) {
+            warn!(
+                "PROXY protocol enabled but server.proxy_protocol_trusted_cidrs is empty; rejecting all PROXY headers by default"
+            );
+        }
+        return false;
+    }
+    trusted.iter().any(|cidr| cidr.contains(peer_ip))
+}
+
+fn synthetic_local_addr(port: u16) -> SocketAddr {
+    SocketAddr::from(([0, 0, 0, 0], port))
+}
+
 pub async fn handle_client_stream<S>(
     mut stream: S,
     peer: SocketAddr,
@@ -94,9 +170,7 @@ where
     let mut real_peer = normalize_ip(peer);
 
     // For non-TCP streams, use a synthetic local address; may be overridden by PROXY protocol dst
-    let mut local_addr: SocketAddr = format!("0.0.0.0:{}", config.server.port)
-        .parse()
-        .unwrap_or_else(|_| "0.0.0.0:443".parse().unwrap());
+    let mut local_addr = synthetic_local_addr(config.server.port);
 
     if proxy_protocol_enabled {
         let proxy_header_timeout = Duration::from_millis(
@@ -104,6 +178,17 @@ where
         );
         match timeout(proxy_header_timeout, parse_proxy_protocol(&mut stream, peer)).await {
             Ok(Ok(info)) => {
+                if !is_trusted_proxy_source(peer.ip(), &config.server.proxy_protocol_trusted_cidrs)
+                {
+                    stats.increment_connects_bad();
+                    warn!(
+                        peer = %peer,
+                        trusted = ?config.server.proxy_protocol_trusted_cidrs,
+                        "Rejecting PROXY protocol header from untrusted source"
+                    );
+                    record_beobachten_class(&beobachten, &config, peer.ip(), "other");
+                    return Err(ProxyError::InvalidProxyProtocol);
+                }
                 debug!(
                     peer = %peer,
                     client = %info.src_addr,
@@ -149,8 +234,13 @@ where
         if is_tls {
             let tls_len = u16::from_be_bytes([first_bytes[3], first_bytes[4]]) as usize;
 
-            if tls_len < 512 {
-                debug!(peer = %real_peer, tls_len = tls_len, "TLS handshake too short");
+// RFC 8446 §5.1 mandates that TLSPlaintext records must not exceed 2^14
+        // bytes (16_384). A client claiming a larger record is non-compliant and
+        // may be an active probe attempting to force large allocations.
+        //
+        // Also enforce a minimum record size to avoid trivial/garbage probes.
+        if !(512..=MAX_TLS_RECORD_SIZE).contains(&tls_len) {
+                debug!(peer = %real_peer, tls_len = tls_len, max_tls_len = MAX_TLS_RECORD_SIZE, "TLS handshake length out of bounds");
                 stats.increment_connects_bad();
                 let (reader, writer) = tokio::io::split(stream);
                 handle_bad_client(
@@ -204,9 +294,19 @@ where
                 &config, &replay_checker, true, Some(tls_user.as_str()),
             ).await {
                 HandshakeResult::Success(result) => result,
-                HandshakeResult::BadClient { reader: _, writer: _ } => {
+                HandshakeResult::BadClient { reader, writer } => {
                     stats.increment_connects_bad();
                     debug!(peer = %peer, "Valid TLS but invalid MTProto handshake");
+                    handle_bad_client(
+                        reader,
+                        writer,
+                        &handshake,
+                        real_peer,
+                        local_addr,
+                        &config,
+                        &beobachten,
+                    )
+                    .await;
                     return Ok(HandshakeOutcome::Handled);
                 }
                 HandshakeResult::Error(e) => return Err(e),
@@ -381,7 +481,6 @@ impl RunningClientHandler {
     pub async fn run(self) -> Result<()> {
         self.stats.increment_connects_all();
         let peer = self.peer;
-        let _ip_tracker = self.ip_tracker.clone();
         debug!(peer = %peer, "New connection");
 
         if let Err(e) = configure_client_socket(
@@ -445,6 +544,24 @@ impl RunningClientHandler {
             .await
             {
                 Ok(Ok(info)) => {
+                    if !is_trusted_proxy_source(
+                        self.peer.ip(),
+                        &self.config.server.proxy_protocol_trusted_cidrs,
+                    ) {
+                        self.stats.increment_connects_bad();
+                        warn!(
+                            peer = %self.peer,
+                            trusted = ?self.config.server.proxy_protocol_trusted_cidrs,
+                            "Rejecting PROXY protocol header from untrusted source"
+                        );
+                        record_beobachten_class(
+                            &self.beobachten,
+                            &self.config,
+                            self.peer.ip(),
+                            "other",
+                        );
+                        return Err(ProxyError::InvalidProxyProtocol);
+                    }
                     debug!(
                         peer = %self.peer,
                         client = %info.src_addr,
@@ -494,7 +611,6 @@ impl RunningClientHandler {
 
         let is_tls = tls::is_tls_handshake(&first_bytes[..3]);
         let peer = self.peer;
-        let _ip_tracker = self.ip_tracker.clone();
 
         debug!(peer = %peer, is_tls = is_tls, "Handshake type detected");
 
@@ -507,14 +623,15 @@ impl RunningClientHandler {
 
     async fn handle_tls_client(mut self, first_bytes: [u8; 5], local_addr: SocketAddr) -> Result<HandshakeOutcome> {
         let peer = self.peer;
-        let _ip_tracker = self.ip_tracker.clone();
 
         let tls_len = u16::from_be_bytes([first_bytes[3], first_bytes[4]]) as usize;
 
         debug!(peer = %peer, tls_len = tls_len, "Reading TLS handshake");
 
-        if tls_len < 512 {
-            debug!(peer = %peer, tls_len = tls_len, "TLS handshake too short");
+        // See RFC 8446 §5.1: TLSPlaintext records must not exceed 16_384 bytes.
+        // Treat too-small or too-large lengths as active probes and mask them.
+        if !(512..=MAX_TLS_RECORD_SIZE).contains(&tls_len) {
+            debug!(peer = %peer, tls_len = tls_len, max_tls_len = MAX_TLS_RECORD_SIZE, "TLS handshake length out of bounds");
             self.stats.increment_connects_bad();
             let (reader, writer) = self.stream.into_split();
             handle_bad_client(
@@ -590,12 +707,19 @@ impl RunningClientHandler {
         .await
         {
             HandshakeResult::Success(result) => result,
-            HandshakeResult::BadClient {
-                reader: _,
-                writer: _,
-            } => {
+            HandshakeResult::BadClient { reader, writer } => {
                 stats.increment_connects_bad();
                 debug!(peer = %peer, "Valid TLS but invalid MTProto handshake");
+                handle_bad_client(
+                    reader,
+                    writer,
+                    &handshake,
+                    peer,
+                    local_addr,
+                    &config,
+                    &self.beobachten,
+                )
+                .await;
                 return Ok(HandshakeOutcome::Handled);
             }
             HandshakeResult::Error(e) => return Err(e),
@@ -622,7 +746,6 @@ impl RunningClientHandler {
 
     async fn handle_direct_client(mut self, first_bytes: [u8; 5], local_addr: SocketAddr) -> Result<HandshakeOutcome> {
         let peer = self.peer;
-        let _ip_tracker = self.ip_tracker.clone();
 
         if !self.config.general.modes.classic && !self.config.general.modes.secure {
             debug!(peer = %peer, "Non-TLS modes disabled");
@@ -726,10 +849,22 @@ impl RunningClientHandler {
     {
         let user = success.user.clone();
 
-        if let Err(e) = Self::check_user_limits_static(&user, &config, &stats, peer_addr, &ip_tracker).await {
-            warn!(user = %user, error = %e, "User limit exceeded");
-            return Err(e);
-        }
+        let user_limit_reservation =
+            match Self::acquire_user_connection_reservation_static(
+                &user,
+                &config,
+                stats.clone(),
+                peer_addr,
+                ip_tracker,
+            )
+            .await
+            {
+                Ok(reservation) => reservation,
+                Err(e) => {
+                    warn!(user = %user, error = %e, "User admission check failed");
+                    return Err(e);
+                }
+            };
 
         let route_snapshot = route_runtime.snapshot();
         let session_id = rng.u64();
@@ -742,7 +877,7 @@ impl RunningClientHandler {
                     client_writer,
                     success,
                     pool.clone(),
-                    stats,
+                    stats.clone(),
                     config,
                     buffer_pool,
                     local_addr,
@@ -759,7 +894,7 @@ impl RunningClientHandler {
                     client_writer,
                     success,
                     upstream_manager,
-                    stats,
+                    stats.clone(),
                     config,
                     buffer_pool,
                     rng,
@@ -776,7 +911,7 @@ impl RunningClientHandler {
                 client_writer,
                 success,
                 upstream_manager,
-                stats,
+                stats.clone(),
                 config,
                 buffer_pool,
                 rng,
@@ -786,14 +921,68 @@ impl RunningClientHandler {
             )
             .await
         };
-
-        ip_tracker.remove_ip(&user, peer_addr.ip()).await;
+        user_limit_reservation.release().await;
         relay_result
     }
 
+    async fn acquire_user_connection_reservation_static(
+        user: &str,
+        config: &ProxyConfig,
+        stats: Arc<Stats>,
+        peer_addr: SocketAddr,
+        ip_tracker: Arc<UserIpTracker>,
+    ) -> Result<UserConnectionReservation> {
+        if let Some(expiration) = config.access.user_expirations.get(user)
+            && chrono::Utc::now() > *expiration
+        {
+            return Err(ProxyError::UserExpired {
+                user: user.to_string(),
+            });
+        }
+
+        if let Some(quota) = config.access.user_data_quota.get(user)
+            && stats.get_user_total_octets(user) >= *quota
+        {
+            return Err(ProxyError::DataQuotaExceeded {
+                user: user.to_string(),
+            });
+        }
+
+        let limit = config.access.user_max_tcp_conns.get(user).map(|v| *v as u64);
+        if !stats.try_acquire_user_curr_connects(user, limit) {
+            return Err(ProxyError::ConnectionLimitExceeded {
+                user: user.to_string(),
+            });
+        }
+
+        match ip_tracker.check_and_add(user, peer_addr.ip()).await {
+            Ok(()) => {}
+            Err(reason) => {
+                stats.decrement_user_curr_connects(user);
+                warn!(
+                    user = %user,
+                    ip = %peer_addr.ip(),
+                    reason = %reason,
+                    "IP limit exceeded"
+                );
+                return Err(ProxyError::ConnectionLimitExceeded {
+                    user: user.to_string(),
+                });
+            }
+        }
+
+        Ok(UserConnectionReservation::new(
+            stats,
+            ip_tracker,
+            user.to_string(),
+            peer_addr.ip(),
+        ))
+    }
+
+    #[cfg(test)]
     async fn check_user_limits_static(
-        user: &str, 
-        config: &ProxyConfig, 
+        user: &str,
+        config: &ProxyConfig,
         stats: &Stats,
         peer_addr: SocketAddr,
         ip_tracker: &UserIpTracker,
@@ -806,9 +995,32 @@ impl RunningClientHandler {
             });
         }
 
-        let ip_reserved = match ip_tracker.check_and_add(user, peer_addr.ip()).await {
-            Ok(()) => true,
+        if let Some(quota) = config.access.user_data_quota.get(user)
+            && stats.get_user_total_octets(user) >= *quota
+        {
+            return Err(ProxyError::DataQuotaExceeded {
+                user: user.to_string(),
+            });
+        }
+
+        let limit = config
+            .access
+            .user_max_tcp_conns
+            .get(user)
+            .map(|v| *v as u64);
+        if !stats.try_acquire_user_curr_connects(user, limit) {
+            return Err(ProxyError::ConnectionLimitExceeded {
+                user: user.to_string(),
+            });
+        }
+
+        match ip_tracker.check_and_add(user, peer_addr.ip()).await {
+            Ok(()) => {
+                ip_tracker.remove_ip(user, peer_addr.ip()).await;
+                stats.decrement_user_curr_connects(user);
+            }
             Err(reason) => {
+                stats.decrement_user_curr_connects(user);
                 warn!(
                     user = %user,
                     ip = %peer_addr.ip(),
@@ -819,33 +1031,15 @@ impl RunningClientHandler {
                     user: user.to_string(),
                 });
             }
-        };
-        // IP limit check
-
-        if let Some(limit) = config.access.user_max_tcp_conns.get(user)
-            && stats.get_user_curr_connects(user) >= *limit as u64
-        {
-            if ip_reserved {
-                ip_tracker.remove_ip(user, peer_addr.ip()).await;
-                stats.increment_ip_reservation_rollback_tcp_limit_total();
-            }
-            return Err(ProxyError::ConnectionLimitExceeded {
-                user: user.to_string(),
-            });
-        }
-
-        if let Some(quota) = config.access.user_data_quota.get(user)
-            && stats.get_user_total_octets(user) >= *quota
-        {
-            if ip_reserved {
-                ip_tracker.remove_ip(user, peer_addr.ip()).await;
-                stats.increment_ip_reservation_rollback_quota_limit_total();
-            }
-            return Err(ProxyError::DataQuotaExceeded {
-                user: user.to_string(),
-            });
         }
 
         Ok(())
     }
 }
+
+#[cfg(test)]
+#[path = "client_security_tests.rs"]
+mod security_tests;
+#[cfg(test)]
+#[path = "client_adversarial_tests.rs"]
+mod adversarial_tests;

src/proxy/client_adversarial_tests.rs

@@ -0,0 +1,109 @@
+use super::*;
+use crate::config::ProxyConfig;
+use crate::stats::Stats;
+use crate::ip_tracker::UserIpTracker;
+use crate::error::ProxyError;
+use std::sync::Arc;
+use std::net::{IpAddr, Ipv4Addr, SocketAddr};
+
+// ------------------------------------------------------------------
+// Priority 3: Massive Concurrency Stress (OWASP ASVS 5.1.6)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn client_stress_10k_connections_limit_strict() {
+    let user = "stress-user";
+    let limit = 512;
+    
+    let stats = Arc::new(Stats::new());
+    let ip_tracker = Arc::new(UserIpTracker::new());
+    
+    let mut config = ProxyConfig::default();
+    config.access.user_max_tcp_conns.insert(user.to_string(), limit);
+    
+    let iterations = 1000;
+    let mut tasks = Vec::new();
+
+    for i in 0..iterations {
+        let stats = Arc::clone(&stats);
+        let ip_tracker = Arc::clone(&ip_tracker);
+        let config = config.clone();
+        let user_str = user.to_string();
+        
+        tasks.push(tokio::spawn(async move {
+            let peer = SocketAddr::new(
+                IpAddr::V4(Ipv4Addr::new(127, 0, 0, (i % 254 + 1) as u8)),
+                10000 + (i % 1000) as u16,
+            );
+            
+            match RunningClientHandler::acquire_user_connection_reservation_static(
+                &user_str,
+                &config,
+                stats,
+                peer,
+                ip_tracker,
+            ).await {
+                Ok(res) => Ok(res),
+                Err(ProxyError::ConnectionLimitExceeded { .. }) => Err(()),
+                Err(e) => panic!("Unexpected error: {:?}", e),
+            }
+        }));
+    }
+
+    let results = futures::future::join_all(tasks).await;
+    let mut successes = 0;
+    let mut failures = 0;
+    let mut reservations = Vec::new();
+
+    for res in results {
+        match res.unwrap() {
+            Ok(r) => {
+                successes += 1;
+                reservations.push(r);
+            }
+            Err(_) => failures += 1,
+        }
+    }
+
+    assert_eq!(successes, limit, "Should allow exactly 'limit' connections");
+    assert_eq!(failures, iterations - limit, "Should fail the rest with LimitExceeded");
+    assert_eq!(stats.get_user_curr_connects(user), limit as u64);
+
+    drop(reservations);
+    
+    ip_tracker.drain_cleanup_queue().await;
+    
+    assert_eq!(stats.get_user_curr_connects(user), 0, "Stats must converge to 0 after all drops");
+    assert_eq!(ip_tracker.get_active_ip_count(user).await, 0, "IP tracker must converge to 0");
+}
+
+// ------------------------------------------------------------------
+// Priority 3: IP Tracker Race Stress
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn client_ip_tracker_race_condition_stress() {
+    let user = "race-user";
+    let ip_tracker = Arc::new(UserIpTracker::new());
+    ip_tracker.set_user_limit(user, 100).await;
+    
+    let iterations = 1000;
+    let mut tasks = Vec::new();
+
+    for i in 0..iterations {
+        let ip_tracker = Arc::clone(&ip_tracker);
+        let ip = IpAddr::V4(Ipv4Addr::new(10, 0, 0, (i % 254 + 1) as u8));
+        
+        tasks.push(tokio::spawn(async move {
+            for _ in 0..10 {
+                if let Ok(()) = ip_tracker.check_and_add("race-user", ip).await {
+                    ip_tracker.remove_ip("race-user", ip).await;
+                }
+            }
+        }));
+    }
+
+    futures::future::join_all(tasks).await;
+    
+    assert_eq!(ip_tracker.get_active_ip_count(user).await, 0, "IP count must be zero after balanced add/remove burst");
+}

src/proxy/direct_relay.rs

@@ -1,7 +1,11 @@
+use std::ffi::OsString;
 use std::fs::OpenOptions;
 use std::io::Write;
 use std::net::SocketAddr;
+use std::path::{Component, Path, PathBuf};
 use std::sync::Arc;
+use std::collections::HashSet;
+use std::sync::{Mutex, OnceLock};
 
 use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt};
 use tokio::net::TcpStream;
@@ -22,6 +26,151 @@ use crate::stats::Stats;
 use crate::stream::{BufferPool, CryptoReader, CryptoWriter};
 use crate::transport::UpstreamManager;
 
+#[cfg(unix)]
+use std::os::unix::fs::OpenOptionsExt;
+
+const UNKNOWN_DC_LOG_DISTINCT_LIMIT: usize = 1024;
+static LOGGED_UNKNOWN_DCS: OnceLock<Mutex<HashSet<i16>>> = OnceLock::new();
+const MAX_SCOPE_HINT_LEN: usize = 64;
+
+fn validated_scope_hint(user: &str) -> Option<&str> {
+    let scope = user.strip_prefix("scope_")?;
+    if scope.is_empty() || scope.len() > MAX_SCOPE_HINT_LEN {
+        return None;
+    }
+    if scope
+        .bytes()
+        .all(|b| b.is_ascii_alphanumeric() || b == b'-')
+    {
+        Some(scope)
+    } else {
+        None
+    }
+}
+
+#[derive(Clone)]
+struct SanitizedUnknownDcLogPath {
+    resolved_path: PathBuf,
+    allowed_parent: PathBuf,
+    file_name: OsString,
+}
+
+// In tests, this function shares global mutable state. Callers that also use
+// cache-reset helpers must hold `unknown_dc_test_lock()` to keep assertions
+// deterministic under parallel execution.
+fn should_log_unknown_dc(dc_idx: i16) -> bool {
+    let set = LOGGED_UNKNOWN_DCS.get_or_init(|| Mutex::new(HashSet::new()));
+    should_log_unknown_dc_with_set(set, dc_idx)
+}
+
+fn should_log_unknown_dc_with_set(set: &Mutex<HashSet<i16>>, dc_idx: i16) -> bool {
+    match set.lock() {
+        Ok(mut guard) => {
+            if guard.contains(&dc_idx) {
+                return false;
+            }
+            if guard.len() >= UNKNOWN_DC_LOG_DISTINCT_LIMIT {
+                return false;
+            }
+            guard.insert(dc_idx)
+        }
+        // Fail closed on poisoned state to avoid unbounded blocking log writes.
+        Err(_) => false,
+    }
+}
+
+fn sanitize_unknown_dc_log_path(path: &str) -> Option<SanitizedUnknownDcLogPath> {
+    let candidate = Path::new(path);
+    if candidate.as_os_str().is_empty() {
+        return None;
+    }
+    if candidate
+        .components()
+        .any(|component| matches!(component, Component::ParentDir))
+    {
+        return None;
+    }
+
+    let cwd = std::env::current_dir().ok()?;
+    let file_name = candidate.file_name()?;
+    let parent = candidate.parent().unwrap_or_else(|| Path::new("."));
+    let parent_path = if parent.is_absolute() {
+        parent.to_path_buf()
+    } else {
+        cwd.join(parent)
+    };
+    let canonical_parent = parent_path.canonicalize().ok()?;
+    if !canonical_parent.is_dir() {
+        return None;
+    }
+
+    Some(SanitizedUnknownDcLogPath {
+        resolved_path: canonical_parent.join(file_name),
+        allowed_parent: canonical_parent,
+        file_name: file_name.to_os_string(),
+    })
+}
+
+fn unknown_dc_log_path_is_still_safe(path: &SanitizedUnknownDcLogPath) -> bool {
+    let Some(parent) = path.resolved_path.parent() else {
+        return false;
+    };
+    let Ok(current_parent) = parent.canonicalize() else {
+        return false;
+    };
+    if current_parent != path.allowed_parent {
+        return false;
+    }
+
+    if let Ok(canonical_target) = path.resolved_path.canonicalize() {
+        let Some(target_parent) = canonical_target.parent() else {
+            return false;
+        };
+        let Some(target_name) = canonical_target.file_name() else {
+            return false;
+        };
+        if target_parent != path.allowed_parent || target_name != path.file_name {
+            return false;
+        }
+    }
+
+    true
+}
+
+fn open_unknown_dc_log_append(path: &Path) -> std::io::Result<std::fs::File> {
+    #[cfg(unix)]
+    {
+        OpenOptions::new()
+            .create(true)
+            .append(true)
+            .custom_flags(libc::O_NOFOLLOW)
+            .open(path)
+    }
+    #[cfg(not(unix))]
+    {
+        let _ = path;
+        Err(std::io::Error::new(
+            std::io::ErrorKind::PermissionDenied,
+            "unknown_dc_file_log_enabled requires unix O_NOFOLLOW support",
+        ))
+    }
+}
+
+#[cfg(test)]
+fn clear_unknown_dc_log_cache_for_testing() {
+    if let Some(set) = LOGGED_UNKNOWN_DCS.get()
+        && let Ok(mut guard) = set.lock()
+    {
+        guard.clear();
+    }
+}
+
+#[cfg(test)]
+fn unknown_dc_test_lock() -> &'static Mutex<()> {
+    static TEST_LOCK: OnceLock<Mutex<()>> = OnceLock::new();
+    TEST_LOCK.get_or_init(|| Mutex::new(()))
+}
+
 pub(crate) async fn handle_via_direct<R, W>(
     client_reader: CryptoReader<R>,
     client_writer: CryptoWriter<W>,
@@ -52,8 +201,15 @@ where
         "Connecting to Telegram DC"
     );
 
+    let scope_hint = validated_scope_hint(user);
+    if user.starts_with("scope_") && scope_hint.is_none() {
+        warn!(
+            user = %user,
+            "Ignoring invalid scope hint and falling back to default upstream selection"
+        );
+    }
     let tg_stream = upstream_manager
-        .connect(dc_addr, Some(success.dc_idx), user.strip_prefix("scope_").filter(|s| !s.is_empty()))
+        .connect(dc_addr, Some(success.dc_idx), scope_hint)
         .await?;
 
     debug!(peer = %success.peer, dc_addr = %dc_addr, "Connected, performing TG handshake");
@@ -64,8 +220,7 @@ where
     debug!(peer = %success.peer, "TG handshake complete, starting relay");
 
     stats.increment_user_connects(user);
-    stats.increment_user_curr_connects(user);
-    stats.increment_current_connections_direct();
+    let _direct_connection_lease = stats.acquire_direct_connection_lease();
 
     let relay_result = relay_bidirectional(
         client_reader,
@@ -76,6 +231,7 @@ where
         config.general.direct_relay_copy_buf_s2c_bytes,
         user,
         Arc::clone(&stats),
+        config.access.user_data_quota.get(user).copied(),
         buffer_pool,
     );
     tokio::pin!(relay_result);
@@ -108,9 +264,6 @@ where
         }
     };
 
-    stats.decrement_current_connections_direct();
-    stats.decrement_user_curr_connects(user);
-
     match &relay_result {
         Ok(()) => debug!(user = %user, "Direct relay completed"),
         Err(e) => debug!(user = %user, error = %e, "Direct relay ended with error"),
@@ -162,12 +315,19 @@ fn get_dc_addr_static(dc_idx: i16, config: &ProxyConfig) -> Result<SocketAddr> {
             && let Some(path) = &config.general.unknown_dc_log_path
             && let Ok(handle) = tokio::runtime::Handle::try_current()
         {
-            let path = path.clone();
-            handle.spawn_blocking(move || {
-                if let Ok(mut file) = OpenOptions::new().create(true).append(true).open(path) {
-                    let _ = writeln!(file, "dc_idx={dc_idx}");
+            if let Some(path) = sanitize_unknown_dc_log_path(path) {
+                if should_log_unknown_dc(dc_idx) {
+                    handle.spawn_blocking(move || {
+                        if unknown_dc_log_path_is_still_safe(&path)
+                            && let Ok(mut file) = open_unknown_dc_log_append(&path.resolved_path)
+                        {
+                            let _ = writeln!(file, "dc_idx={dc_idx}");
+                        }
+                    });
                 }
-            });
+            } else {
+                warn!(dc_idx = dc_idx, raw_path = %path, "Rejected unsafe unknown DC log path");
+            }
         }
     }
 
@@ -175,7 +335,7 @@ fn get_dc_addr_static(dc_idx: i16, config: &ProxyConfig) -> Result<SocketAddr> {
     let fallback_idx = if default_dc >= 1 && default_dc <= num_dcs {
         default_dc - 1
     } else {
-        1
+        0
     };
 
     info!(
@@ -203,8 +363,6 @@ async fn do_tg_handshake_static(
     let (nonce, _tg_enc_key, _tg_enc_iv, _tg_dec_key, _tg_dec_iv) = generate_tg_nonce(
         success.proto_tag,
         success.dc_idx,
-        &success.dec_key,
-        success.dec_iv,
         &success.enc_key,
         success.enc_iv,
         rng,
@@ -230,3 +388,7 @@ async fn do_tg_handshake_static(
         CryptoWriter::new(write_half, tg_encryptor, max_pending),
     ))
 }
+
+#[cfg(test)]
+#[path = "direct_relay_security_tests.rs"]
+mod security_tests;

src/proxy/handshake.rs

@@ -3,11 +3,18 @@
 #![allow(dead_code)]
 
 use std::net::SocketAddr;
+use std::collections::HashSet;
+use std::collections::hash_map::RandomState;
+use std::net::{IpAddr, Ipv6Addr};
 use std::sync::Arc;
-use std::time::Duration;
+use std::sync::{Mutex, OnceLock};
+use std::hash::{BuildHasher, Hash, Hasher};
+use std::time::{Duration, Instant};
+use dashmap::DashMap;
+use dashmap::mapref::entry::Entry;
 use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt};
 use tracing::{debug, warn, trace};
-use zeroize::Zeroize;
+use zeroize::{Zeroize, Zeroizing};
 
 use crate::crypto::{sha256, AesCtr, SecureRandom};
 use rand::Rng;
@@ -19,6 +26,463 @@ use crate::stats::ReplayChecker;
 use crate::config::ProxyConfig;
 use crate::tls_front::{TlsFrontCache, emulator};
 
+const ACCESS_SECRET_BYTES: usize = 16;
+static INVALID_SECRET_WARNED: OnceLock<Mutex<HashSet<(String, String)>>> = OnceLock::new();
+#[cfg(test)]
+const WARNED_SECRET_MAX_ENTRIES: usize = 64;
+#[cfg(not(test))]
+const WARNED_SECRET_MAX_ENTRIES: usize = 1_024;
+
+const AUTH_PROBE_TRACK_RETENTION_SECS: u64 = 10 * 60;
+#[cfg(test)]
+const AUTH_PROBE_TRACK_MAX_ENTRIES: usize = 256;
+#[cfg(not(test))]
+const AUTH_PROBE_TRACK_MAX_ENTRIES: usize = 65_536;
+const AUTH_PROBE_PRUNE_SCAN_LIMIT: usize = 1_024;
+const AUTH_PROBE_BACKOFF_START_FAILS: u32 = 4;
+const AUTH_PROBE_SATURATION_GRACE_FAILS: u32 = 2;
+
+#[cfg(test)]
+const AUTH_PROBE_BACKOFF_BASE_MS: u64 = 1;
+#[cfg(not(test))]
+const AUTH_PROBE_BACKOFF_BASE_MS: u64 = 25;
+
+#[cfg(test)]
+const AUTH_PROBE_BACKOFF_MAX_MS: u64 = 16;
+#[cfg(not(test))]
+const AUTH_PROBE_BACKOFF_MAX_MS: u64 = 1_000;
+
+#[derive(Clone, Copy)]
+struct AuthProbeState {
+    fail_streak: u32,
+    blocked_until: Instant,
+    last_seen: Instant,
+}
+
+#[derive(Clone, Copy)]
+struct AuthProbeSaturationState {
+    fail_streak: u32,
+    blocked_until: Instant,
+    last_seen: Instant,
+}
+
+static AUTH_PROBE_STATE: OnceLock<DashMap<IpAddr, AuthProbeState>> = OnceLock::new();
+static AUTH_PROBE_SATURATION_STATE: OnceLock<Mutex<Option<AuthProbeSaturationState>>> = OnceLock::new();
+static AUTH_PROBE_EVICTION_HASHER: OnceLock<RandomState> = OnceLock::new();
+
+fn auth_probe_state_map() -> &'static DashMap<IpAddr, AuthProbeState> {
+    AUTH_PROBE_STATE.get_or_init(DashMap::new)
+}
+
+fn auth_probe_saturation_state() -> &'static Mutex<Option<AuthProbeSaturationState>> {
+    AUTH_PROBE_SATURATION_STATE.get_or_init(|| Mutex::new(None))
+}
+
+fn normalize_auth_probe_ip(peer_ip: IpAddr) -> IpAddr {
+    match peer_ip {
+        IpAddr::V4(ip) => IpAddr::V4(ip),
+        IpAddr::V6(ip) => {
+            let [a, b, c, d, _, _, _, _] = ip.segments();
+            IpAddr::V6(Ipv6Addr::new(a, b, c, d, 0, 0, 0, 0))
+        }
+    }
+}
+
+fn auth_probe_backoff(fail_streak: u32) -> Duration {
+    if fail_streak < AUTH_PROBE_BACKOFF_START_FAILS {
+        return Duration::ZERO;
+    }
+    let shift = (fail_streak - AUTH_PROBE_BACKOFF_START_FAILS).min(10);
+    let multiplier = 1u64.checked_shl(shift).unwrap_or(u64::MAX);
+    let ms = AUTH_PROBE_BACKOFF_BASE_MS
+        .saturating_mul(multiplier)
+        .min(AUTH_PROBE_BACKOFF_MAX_MS);
+    Duration::from_millis(ms)
+}
+
+fn auth_probe_state_expired(state: &AuthProbeState, now: Instant) -> bool {
+    let retention = Duration::from_secs(AUTH_PROBE_TRACK_RETENTION_SECS);
+    now.duration_since(state.last_seen) > retention
+}
+
+fn auth_probe_eviction_offset(peer_ip: IpAddr, now: Instant) -> usize {
+    let hasher_state = AUTH_PROBE_EVICTION_HASHER.get_or_init(RandomState::new);
+    let mut hasher = hasher_state.build_hasher();
+    peer_ip.hash(&mut hasher);
+    now.hash(&mut hasher);
+    hasher.finish() as usize
+}
+
+fn auth_probe_is_throttled(peer_ip: IpAddr, now: Instant) -> bool {
+    let peer_ip = normalize_auth_probe_ip(peer_ip);
+    let state = auth_probe_state_map();
+    let Some(entry) = state.get(&peer_ip) else {
+        return false;
+    };
+    if auth_probe_state_expired(&entry, now) {
+        drop(entry);
+        state.remove(&peer_ip);
+        return false;
+    }
+    now < entry.blocked_until
+}
+
+fn auth_probe_saturation_grace_exhausted(peer_ip: IpAddr, now: Instant) -> bool {
+    let peer_ip = normalize_auth_probe_ip(peer_ip);
+    let state = auth_probe_state_map();
+    let Some(entry) = state.get(&peer_ip) else {
+        return false;
+    };
+    if auth_probe_state_expired(&entry, now) {
+        drop(entry);
+        state.remove(&peer_ip);
+        return false;
+    }
+
+    entry.fail_streak >= AUTH_PROBE_BACKOFF_START_FAILS + AUTH_PROBE_SATURATION_GRACE_FAILS
+}
+
+fn auth_probe_should_apply_preauth_throttle(peer_ip: IpAddr, now: Instant) -> bool {
+    if !auth_probe_is_throttled(peer_ip, now) {
+        return false;
+    }
+
+    if !auth_probe_saturation_is_throttled(now) {
+        return true;
+    }
+
+    auth_probe_saturation_grace_exhausted(peer_ip, now)
+}
+
+fn auth_probe_saturation_is_throttled(now: Instant) -> bool {
+    let saturation = auth_probe_saturation_state();
+    let mut guard = match saturation.lock() {
+        Ok(guard) => guard,
+        Err(_) => return false,
+    };
+
+    let Some(state) = guard.as_mut() else {
+        return false;
+    };
+
+    if now.duration_since(state.last_seen) > Duration::from_secs(AUTH_PROBE_TRACK_RETENTION_SECS) {
+        *guard = None;
+        return false;
+    }
+
+    if now < state.blocked_until {
+        return true;
+    }
+
+    false
+}
+
+fn auth_probe_note_saturation(now: Instant) {
+    let saturation = auth_probe_saturation_state();
+    let mut guard = match saturation.lock() {
+        Ok(guard) => guard,
+        Err(_) => return,
+    };
+
+    match guard.as_mut() {
+        Some(state)
+            if now.duration_since(state.last_seen)
+                <= Duration::from_secs(AUTH_PROBE_TRACK_RETENTION_SECS) =>
+        {
+            state.fail_streak = state.fail_streak.saturating_add(1);
+            state.last_seen = now;
+            state.blocked_until = now + auth_probe_backoff(state.fail_streak);
+        }
+        _ => {
+            let fail_streak = AUTH_PROBE_BACKOFF_START_FAILS;
+            *guard = Some(AuthProbeSaturationState {
+                fail_streak,
+                blocked_until: now + auth_probe_backoff(fail_streak),
+                last_seen: now,
+            });
+        }
+    }
+}
+
+fn auth_probe_record_failure(peer_ip: IpAddr, now: Instant) {
+    let peer_ip = normalize_auth_probe_ip(peer_ip);
+    let state = auth_probe_state_map();
+    auth_probe_record_failure_with_state(state, peer_ip, now);
+}
+
+fn auth_probe_record_failure_with_state(
+    state: &DashMap<IpAddr, AuthProbeState>,
+    peer_ip: IpAddr,
+    now: Instant,
+) {
+    let make_new_state = || AuthProbeState {
+        fail_streak: 1,
+        blocked_until: now + auth_probe_backoff(1),
+        last_seen: now,
+    };
+
+    let update_existing = |entry: &mut AuthProbeState| {
+        if auth_probe_state_expired(entry, now) {
+            *entry = make_new_state();
+        } else {
+            entry.fail_streak = entry.fail_streak.saturating_add(1);
+            entry.last_seen = now;
+            entry.blocked_until = now + auth_probe_backoff(entry.fail_streak);
+        }
+    };
+
+    match state.entry(peer_ip) {
+        Entry::Occupied(mut entry) => {
+            update_existing(entry.get_mut());
+            return;
+        }
+        Entry::Vacant(_) => {}
+    }
+
+    if state.len() >= AUTH_PROBE_TRACK_MAX_ENTRIES {
+        let mut rounds = 0usize;
+        while state.len() >= AUTH_PROBE_TRACK_MAX_ENTRIES {
+            rounds += 1;
+            if rounds > 8 {
+                auth_probe_note_saturation(now);
+                let mut eviction_candidate: Option<(IpAddr, u32, Instant)> = None;
+                for entry in state.iter().take(AUTH_PROBE_PRUNE_SCAN_LIMIT) {
+                    let key = *entry.key();
+                    let fail_streak = entry.value().fail_streak;
+                    let last_seen = entry.value().last_seen;
+                    match eviction_candidate {
+                        Some((_, current_fail, current_seen))
+                            if fail_streak > current_fail
+                                || (fail_streak == current_fail && last_seen >= current_seen) =>
+                        {
+                        }
+                        _ => eviction_candidate = Some((key, fail_streak, last_seen)),
+                    }
+                }
+
+                let Some((evict_key, _, _)) = eviction_candidate else {
+                    return;
+                };
+                state.remove(&evict_key);
+                break;
+            }
+
+            let mut stale_keys = Vec::new();
+            let mut eviction_candidate: Option<(IpAddr, u32, Instant)> = None;
+            let state_len = state.len();
+            let scan_limit = state_len.min(AUTH_PROBE_PRUNE_SCAN_LIMIT);
+            let start_offset = if state_len == 0 {
+                0
+            } else {
+                auth_probe_eviction_offset(peer_ip, now) % state_len
+            };
+
+            let mut scanned = 0usize;
+            for entry in state.iter().skip(start_offset) {
+                let key = *entry.key();
+                let fail_streak = entry.value().fail_streak;
+                let last_seen = entry.value().last_seen;
+                match eviction_candidate {
+                    Some((_, current_fail, current_seen))
+                        if fail_streak > current_fail
+                            || (fail_streak == current_fail && last_seen >= current_seen) =>
+                    {
+                    }
+                    _ => eviction_candidate = Some((key, fail_streak, last_seen)),
+                }
+                if auth_probe_state_expired(entry.value(), now) {
+                    stale_keys.push(key);
+                }
+                scanned += 1;
+                if scanned >= scan_limit {
+                    break;
+                }
+            }
+
+            if scanned < scan_limit {
+                for entry in state.iter().take(scan_limit - scanned) {
+                    let key = *entry.key();
+                    let fail_streak = entry.value().fail_streak;
+                    let last_seen = entry.value().last_seen;
+                    match eviction_candidate {
+                        Some((_, current_fail, current_seen))
+                            if fail_streak > current_fail
+                                || (fail_streak == current_fail && last_seen >= current_seen) =>
+                        {
+                        }
+                        _ => eviction_candidate = Some((key, fail_streak, last_seen)),
+                    }
+                    if auth_probe_state_expired(entry.value(), now) {
+                        stale_keys.push(key);
+                    }
+                }
+            }
+
+            for stale_key in stale_keys {
+                state.remove(&stale_key);
+            }
+
+            if state.len() < AUTH_PROBE_TRACK_MAX_ENTRIES {
+                break;
+            }
+
+            let Some((evict_key, _, _)) = eviction_candidate else {
+                auth_probe_note_saturation(now);
+                return;
+            };
+            state.remove(&evict_key);
+            auth_probe_note_saturation(now);
+        }
+    }
+
+    match state.entry(peer_ip) {
+        Entry::Occupied(mut entry) => {
+            update_existing(entry.get_mut());
+        }
+        Entry::Vacant(entry) => {
+            entry.insert(make_new_state());
+        }
+    }
+}
+
+fn auth_probe_record_success(peer_ip: IpAddr) {
+    let peer_ip = normalize_auth_probe_ip(peer_ip);
+    let state = auth_probe_state_map();
+    state.remove(&peer_ip);
+}
+
+#[cfg(test)]
+fn clear_auth_probe_state_for_testing() {
+    if let Some(state) = AUTH_PROBE_STATE.get() {
+        state.clear();
+    }
+    if let Some(saturation) = AUTH_PROBE_SATURATION_STATE.get()
+        && let Ok(mut guard) = saturation.lock()
+    {
+        *guard = None;
+    }
+}
+
+#[cfg(test)]
+fn auth_probe_fail_streak_for_testing(peer_ip: IpAddr) -> Option<u32> {
+    let peer_ip = normalize_auth_probe_ip(peer_ip);
+    let state = AUTH_PROBE_STATE.get()?;
+    state.get(&peer_ip).map(|entry| entry.fail_streak)
+}
+
+#[cfg(test)]
+fn auth_probe_is_throttled_for_testing(peer_ip: IpAddr) -> bool {
+    auth_probe_is_throttled(peer_ip, Instant::now())
+}
+
+#[cfg(test)]
+fn auth_probe_saturation_is_throttled_for_testing() -> bool {
+    auth_probe_saturation_is_throttled(Instant::now())
+}
+
+#[cfg(test)]
+fn auth_probe_saturation_is_throttled_at_for_testing(now: Instant) -> bool {
+    auth_probe_saturation_is_throttled(now)
+}
+
+#[cfg(test)]
+fn auth_probe_test_lock() -> &'static Mutex<()> {
+    static TEST_LOCK: OnceLock<Mutex<()>> = OnceLock::new();
+    TEST_LOCK.get_or_init(|| Mutex::new(()))
+}
+
+#[cfg(test)]
+fn clear_warned_secrets_for_testing() {
+    if let Some(warned) = INVALID_SECRET_WARNED.get()
+        && let Ok(mut guard) = warned.lock()
+    {
+        guard.clear();
+    }
+}
+
+#[cfg(test)]
+fn warned_secrets_test_lock() -> &'static Mutex<()> {
+    static TEST_LOCK: OnceLock<Mutex<()>> = OnceLock::new();
+    TEST_LOCK.get_or_init(|| Mutex::new(()))
+}
+
+fn warn_invalid_secret_once(name: &str, reason: &str, expected: usize, got: Option<usize>) {
+    let key = (name.to_string(), reason.to_string());
+    let warned = INVALID_SECRET_WARNED.get_or_init(|| Mutex::new(HashSet::new()));
+    let should_warn = match warned.lock() {
+        Ok(mut guard) => {
+            if !guard.contains(&key) && guard.len() >= WARNED_SECRET_MAX_ENTRIES {
+                false
+            } else {
+                guard.insert(key)
+            }
+        }
+        Err(_) => true,
+    };
+
+    if !should_warn {
+        return;
+    }
+
+    match got {
+        Some(actual) => {
+            warn!(
+                user = %name,
+                expected = expected,
+                got = actual,
+                "Skipping user: access secret has unexpected length"
+            );
+        }
+        None => {
+            warn!(
+                user = %name,
+                "Skipping user: access secret is not valid hex"
+            );
+        }
+    }
+}
+
+fn decode_user_secret(name: &str, secret_hex: &str) -> Option<Vec<u8>> {
+    match hex::decode(secret_hex) {
+        Ok(bytes) if bytes.len() == ACCESS_SECRET_BYTES => Some(bytes),
+        Ok(bytes) => {
+            warn_invalid_secret_once(
+                name,
+                "invalid_length",
+                ACCESS_SECRET_BYTES,
+                Some(bytes.len()),
+            );
+            None
+        }
+        Err(_) => {
+            warn_invalid_secret_once(name, "invalid_hex", ACCESS_SECRET_BYTES, None);
+            None
+        }
+    }
+}
+
+// Decide whether a client-supplied proto tag is allowed given the configured
+// proxy modes and the transport that carried the handshake.
+//
+// A common mistake is to treat `modes.tls` and `modes.secure` as interchangeable
+// even though they correspond to different transport profiles: `modes.tls` is
+// for the TLS-fronted (EE-TLS) path, while `modes.secure` is for direct MTProto
+// over TCP (DD). Enforcing this separation prevents an attacker from using a
+// TLS-capable client to bypass the operator intent for the direct MTProto mode,
+// and vice versa.
+fn mode_enabled_for_proto(config: &ProxyConfig, proto_tag: ProtoTag, is_tls: bool) -> bool {
+    match proto_tag {
+        ProtoTag::Secure => {
+            if is_tls {
+                config.general.modes.tls
+            } else {
+                config.general.modes.secure
+            }
+        }
+        ProtoTag::Intermediate | ProtoTag::Abridged => config.general.modes.classic,
+    }
+}
+
 fn decode_user_secrets(
     config: &ProxyConfig,
     preferred_user: Option<&str>,
@@ -27,7 +491,7 @@ fn decode_user_secrets(
 
     if let Some(preferred) = preferred_user
         && let Some(secret_hex) = config.access.users.get(preferred)
-        && let Ok(bytes) = hex::decode(secret_hex)
+        && let Some(bytes) = decode_user_secret(preferred, secret_hex)
     {
         secrets.push((preferred.to_string(), bytes));
     }
@@ -36,7 +500,7 @@ fn decode_user_secrets(
         if preferred_user.is_some_and(|preferred| preferred == name.as_str()) {
             continue;
         }
-        if let Ok(bytes) = hex::decode(secret_hex) {
+        if let Some(bytes) = decode_user_secret(name, secret_hex) {
             secrets.push((name.clone(), bytes));
         }
     }
@@ -44,11 +508,29 @@ fn decode_user_secrets(
     secrets
 }
 
+async fn maybe_apply_server_hello_delay(config: &ProxyConfig) {
+    if config.censorship.server_hello_delay_max_ms == 0 {
+        return;
+    }
+
+    let min = config.censorship.server_hello_delay_min_ms;
+    let max = config.censorship.server_hello_delay_max_ms.max(min);
+    let delay_ms = if max == min {
+        max
+    } else {
+        rand::rng().random_range(min..=max)
+    };
+
+    if delay_ms > 0 {
+        tokio::time::sleep(Duration::from_millis(delay_ms)).await;
+    }
+}
+
 /// Result of successful handshake
 ///
 /// Key material (`dec_key`, `dec_iv`, `enc_key`, `enc_iv`) is
 /// zeroized on drop.
-#[derive(Debug, Clone)]
+#[derive(Debug)]
 pub struct HandshakeSuccess {
     /// Authenticated user name
     pub user: String,
@@ -65,6 +547,7 @@ pub struct HandshakeSuccess {
     /// Client address
     pub peer: SocketAddr,
     /// Whether TLS was used
+    
     pub is_tls: bool,
 }
 
@@ -94,28 +577,33 @@ where
 {
     debug!(peer = %peer, handshake_len = handshake.len(), "Processing TLS handshake");
 
+    let throttle_now = Instant::now();
+    if auth_probe_should_apply_preauth_throttle(peer.ip(), throttle_now) {
+        maybe_apply_server_hello_delay(config).await;
+        debug!(peer = %peer, "TLS handshake rejected by pre-auth probe throttle");
+        return HandshakeResult::BadClient { reader, writer };
+    }
+
     if handshake.len() < tls::TLS_DIGEST_POS + tls::TLS_DIGEST_LEN + 1 {
+        auth_probe_record_failure(peer.ip(), Instant::now());
+        maybe_apply_server_hello_delay(config).await;
         debug!(peer = %peer, "TLS handshake too short");
         return HandshakeResult::BadClient { reader, writer };
     }
 
-    let digest = &handshake[tls::TLS_DIGEST_POS..tls::TLS_DIGEST_POS + tls::TLS_DIGEST_LEN];
-    let digest_half = &digest[..tls::TLS_DIGEST_HALF_LEN];
+    let client_sni = tls::extract_sni_from_client_hello(handshake);
+    let secrets = decode_user_secrets(config, client_sni.as_deref());
 
-    if replay_checker.check_and_add_tls_digest(digest_half) {
-        warn!(peer = %peer, "TLS replay attack detected (duplicate digest)");
-        return HandshakeResult::BadClient { reader, writer };
-    }
-
-    let secrets = decode_user_secrets(config, None);
-
-    let validation = match tls::validate_tls_handshake(
+    let validation = match tls::validate_tls_handshake_with_replay_window(
         handshake,
         &secrets,
         config.access.ignore_time_skew,
+        config.access.replay_window_secs,
     ) {
         Some(v) => v,
         None => {
+            auth_probe_record_failure(peer.ip(), Instant::now());
+            maybe_apply_server_hello_delay(config).await;
             debug!(
                 peer = %peer, 
                 ignore_time_skew = config.access.ignore_time_skew,
@@ -125,16 +613,29 @@ where
         }
     };
 
+    // Replay tracking is applied only after successful authentication to avoid
+    // letting unauthenticated probes evict valid entries from the replay cache.
+    let digest_half = &validation.digest[..tls::TLS_DIGEST_HALF_LEN];
+    if replay_checker.check_and_add_tls_digest(digest_half) {
+        auth_probe_record_failure(peer.ip(), Instant::now());
+        maybe_apply_server_hello_delay(config).await;
+        warn!(peer = %peer, "TLS replay attack detected (duplicate digest)");
+        return HandshakeResult::BadClient { reader, writer };
+    }
+
     let secret = match secrets.iter().find(|(name, _)| *name == validation.user) {
         Some((_, s)) => s,
-        None => return HandshakeResult::BadClient { reader, writer },
+        None => {
+            maybe_apply_server_hello_delay(config).await;
+            return HandshakeResult::BadClient { reader, writer };
+        }
     };
 
     let cached = if config.censorship.tls_emulation {
         if let Some(cache) = tls_cache.as_ref() {
-            let selected_domain = if let Some(sni) = tls::extract_sni_from_client_hello(handshake) {
+            let selected_domain = if let Some(sni) = client_sni.as_ref() {
                 if cache.contains_domain(&sni).await {
-                    sni
+                    sni.clone()
                 } else {
                     config.censorship.tls_domain.clone()
                 }
@@ -166,6 +667,10 @@ where
             Some(b"h2".to_vec())
         } else if alpn_list.iter().any(|p| p == b"http/1.1") {
             Some(b"http/1.1".to_vec())
+        } else if !alpn_list.is_empty() {
+            maybe_apply_server_hello_delay(config).await;
+            debug!(peer = %peer, "Client ALPN list has no supported protocol; using masking fallback");
+            return HandshakeResult::BadClient { reader, writer };
         } else {
             None
         }
@@ -196,19 +701,9 @@ where
         )
     };
 
-    // Optional anti-fingerprint delay before sending ServerHello.
-    if config.censorship.server_hello_delay_max_ms > 0 {
-        let min = config.censorship.server_hello_delay_min_ms;
-        let max = config.censorship.server_hello_delay_max_ms.max(min);
-        let delay_ms = if max == min {
-            max
-        } else {
-            rand::rng().random_range(min..=max)
-        };
-        if delay_ms > 0 {
-            tokio::time::sleep(std::time::Duration::from_millis(delay_ms)).await;
-        }
-    }
+    // Apply the same optional delay budget used by reject paths to reduce
+    // distinguishability between success and fail-closed handshakes.
+    maybe_apply_server_hello_delay(config).await;
 
     debug!(peer = %peer, response_len = response.len(), "Sending TLS ServerHello");
 
@@ -228,6 +723,8 @@ where
         "TLS handshake successful"
     );
 
+    auth_probe_record_success(peer.ip());
+
     HandshakeResult::Success((
         FakeTlsReader::new(reader),
         FakeTlsWriter::new(writer),
@@ -250,15 +747,25 @@ where
     R: AsyncRead + Unpin + Send,
     W: AsyncWrite + Unpin + Send,
 {
-    trace!(peer = %peer, handshake = ?hex::encode(handshake), "MTProto handshake bytes");
+    let handshake_fingerprint = {
+        let digest = sha256(&handshake[..8]);
+        hex::encode(&digest[..4])
+    };
+    trace!(
+        peer = %peer,
+        handshake_fingerprint = %handshake_fingerprint,
+        "MTProto handshake prefix"
+    );
 
-    let dec_prekey_iv = &handshake[SKIP_LEN..SKIP_LEN + PREKEY_LEN + IV_LEN];
-
-    if replay_checker.check_and_add_handshake(dec_prekey_iv) {
-        warn!(peer = %peer, "MTProto replay attack detected");
+    let throttle_now = Instant::now();
+    if auth_probe_should_apply_preauth_throttle(peer.ip(), throttle_now) {
+        maybe_apply_server_hello_delay(config).await;
+        debug!(peer = %peer, "MTProto handshake rejected by pre-auth probe throttle");
         return HandshakeResult::BadClient { reader, writer };
     }
 
+    let dec_prekey_iv = &handshake[SKIP_LEN..SKIP_LEN + PREKEY_LEN + IV_LEN];
+
     let enc_prekey_iv: Vec<u8> = dec_prekey_iv.iter().rev().copied().collect();
 
     let decoded_users = decode_user_secrets(config, preferred_user);
@@ -268,57 +775,66 @@ where
         let dec_prekey = &dec_prekey_iv[..PREKEY_LEN];
         let dec_iv_bytes = &dec_prekey_iv[PREKEY_LEN..];
 
-        let mut dec_key_input = Vec::with_capacity(PREKEY_LEN + secret.len());
+        let mut dec_key_input = Zeroizing::new(Vec::with_capacity(PREKEY_LEN + secret.len()));
         dec_key_input.extend_from_slice(dec_prekey);
         dec_key_input.extend_from_slice(&secret);
         let dec_key = sha256(&dec_key_input);
 
-        let dec_iv = u128::from_be_bytes(dec_iv_bytes.try_into().unwrap());
+        let mut dec_iv_arr = [0u8; IV_LEN];
+        dec_iv_arr.copy_from_slice(dec_iv_bytes);
+        let dec_iv = u128::from_be_bytes(dec_iv_arr);
 
         let mut decryptor = AesCtr::new(&dec_key, dec_iv);
         let decrypted = decryptor.decrypt(handshake);
 
-        let tag_bytes: [u8; 4] = decrypted[PROTO_TAG_POS..PROTO_TAG_POS + 4]
-            .try_into()
-            .unwrap();
+        let tag_bytes: [u8; 4] = [
+            decrypted[PROTO_TAG_POS],
+            decrypted[PROTO_TAG_POS + 1],
+            decrypted[PROTO_TAG_POS + 2],
+            decrypted[PROTO_TAG_POS + 3],
+        ];
 
         let proto_tag = match ProtoTag::from_bytes(tag_bytes) {
             Some(tag) => tag,
             None => continue,
         };
 
-        let mode_ok = match proto_tag {
-            ProtoTag::Secure => {
-                if is_tls {
-                    config.general.modes.tls || config.general.modes.secure
-                } else {
-                    config.general.modes.secure || config.general.modes.tls
-                }
-            }
-            ProtoTag::Intermediate | ProtoTag::Abridged => config.general.modes.classic,
-        };
+        let mode_ok = mode_enabled_for_proto(config, proto_tag, is_tls);
 
         if !mode_ok {
             debug!(peer = %peer, user = %user, proto = ?proto_tag, "Mode not enabled");
             continue;
         }
 
-        let dc_idx = i16::from_le_bytes(
-            decrypted[DC_IDX_POS..DC_IDX_POS + 2].try_into().unwrap()
-        );
+        let dc_idx = i16::from_le_bytes([decrypted[DC_IDX_POS], decrypted[DC_IDX_POS + 1]]);
 
         let enc_prekey = &enc_prekey_iv[..PREKEY_LEN];
         let enc_iv_bytes = &enc_prekey_iv[PREKEY_LEN..];
 
-        let mut enc_key_input = Vec::with_capacity(PREKEY_LEN + secret.len());
+        let mut enc_key_input = Zeroizing::new(Vec::with_capacity(PREKEY_LEN + secret.len()));
         enc_key_input.extend_from_slice(enc_prekey);
         enc_key_input.extend_from_slice(&secret);
         let enc_key = sha256(&enc_key_input);
 
-        let enc_iv = u128::from_be_bytes(enc_iv_bytes.try_into().unwrap());
+        let mut enc_iv_arr = [0u8; IV_LEN];
+        enc_iv_arr.copy_from_slice(enc_iv_bytes);
+        let enc_iv = u128::from_be_bytes(enc_iv_arr);
 
         let encryptor = AesCtr::new(&enc_key, enc_iv);
 
+// Apply replay tracking only after successful authentication.
+    //
+    // This ordering prevents an attacker from producing invalid handshakes that
+    // still collide with a valid handshake's replay slot and thus evict a valid
+    // entry from the cache. We accept the cost of performing the full
+    // authentication check first to avoid poisoning the replay cache.
+        if replay_checker.check_and_add_handshake(dec_prekey_iv) {
+            auth_probe_record_failure(peer.ip(), Instant::now());
+            maybe_apply_server_hello_delay(config).await;
+            warn!(peer = %peer, user = %user, "MTProto replay attack detected");
+            return HandshakeResult::BadClient { reader, writer };
+        }
+
         let success = HandshakeSuccess {
             user: user.clone(),
             dc_idx,
@@ -340,6 +856,8 @@ where
             "MTProto handshake successful"
         );
 
+        auth_probe_record_success(peer.ip());
+
         let max_pending = config.general.crypto_pending_buffer;
         return HandshakeResult::Success((
             CryptoReader::new(reader, decryptor),
@@ -348,6 +866,8 @@ where
         ));
     }
 
+    auth_probe_record_failure(peer.ip(), Instant::now());
+    maybe_apply_server_hello_delay(config).await;
     debug!(peer = %peer, "MTProto handshake: no matching user found");
     HandshakeResult::BadClient { reader, writer }
 }
@@ -356,8 +876,6 @@ where
 pub fn generate_tg_nonce(
     proto_tag: ProtoTag, 
     dc_idx: i16,
-    _client_dec_key: &[u8; 32],
-    _client_dec_iv: u128,
     client_enc_key: &[u8; 32],
     client_enc_iv: u128,
     rng: &SecureRandom,
@@ -365,14 +883,16 @@ pub fn generate_tg_nonce(
 ) -> ([u8; HANDSHAKE_LEN], [u8; 32], u128, [u8; 32], u128) {
     loop {
         let bytes = rng.bytes(HANDSHAKE_LEN);
-        let mut nonce: [u8; HANDSHAKE_LEN] = bytes.try_into().unwrap();
+        let Ok(mut nonce): Result<[u8; HANDSHAKE_LEN], _> = bytes.try_into() else {
+            continue;
+        };
 
         if RESERVED_NONCE_FIRST_BYTES.contains(&nonce[0]) { continue; }
 
-        let first_four: [u8; 4] = nonce[..4].try_into().unwrap();
+        let first_four: [u8; 4] = [nonce[0], nonce[1], nonce[2], nonce[3]];
         if RESERVED_NONCE_BEGINNINGS.contains(&first_four) { continue; }
 
-        let continue_four: [u8; 4] = nonce[4..8].try_into().unwrap();
+        let continue_four: [u8; 4] = [nonce[4], nonce[5], nonce[6], nonce[7]];
         if RESERVED_NONCE_CONTINUES.contains(&continue_four) { continue; }
 
         nonce[PROTO_TAG_POS..PROTO_TAG_POS + 4].copy_from_slice(&proto_tag.to_bytes());
@@ -380,7 +900,7 @@ pub fn generate_tg_nonce(
         nonce[DC_IDX_POS..DC_IDX_POS + 2].copy_from_slice(&dc_idx.to_le_bytes());
 
         if fast_mode {
-            let mut key_iv = Vec::with_capacity(KEY_LEN + IV_LEN);
+            let mut key_iv = Zeroizing::new(Vec::with_capacity(KEY_LEN + IV_LEN));
             key_iv.extend_from_slice(client_enc_key);
             key_iv.extend_from_slice(&client_enc_iv.to_be_bytes());
             key_iv.reverse(); // Python/C behavior: reversed enc_key+enc_iv in nonce
@@ -388,13 +908,19 @@ pub fn generate_tg_nonce(
         }
 
         let enc_key_iv = &nonce[SKIP_LEN..SKIP_LEN + KEY_LEN + IV_LEN];
-        let dec_key_iv: Vec<u8> = enc_key_iv.iter().rev().copied().collect();
+        let dec_key_iv = Zeroizing::new(enc_key_iv.iter().rev().copied().collect::<Vec<u8>>());
 
-        let tg_enc_key: [u8; 32] = enc_key_iv[..KEY_LEN].try_into().unwrap();
-        let tg_enc_iv = u128::from_be_bytes(enc_key_iv[KEY_LEN..].try_into().unwrap());
+        let mut tg_enc_key = [0u8; 32];
+        tg_enc_key.copy_from_slice(&enc_key_iv[..KEY_LEN]);
+        let mut tg_enc_iv_arr = [0u8; IV_LEN];
+        tg_enc_iv_arr.copy_from_slice(&enc_key_iv[KEY_LEN..]);
+        let tg_enc_iv = u128::from_be_bytes(tg_enc_iv_arr);
 
-        let tg_dec_key: [u8; 32] = dec_key_iv[..KEY_LEN].try_into().unwrap();
-        let tg_dec_iv = u128::from_be_bytes(dec_key_iv[KEY_LEN..].try_into().unwrap());
+        let mut tg_dec_key = [0u8; 32];
+        tg_dec_key.copy_from_slice(&dec_key_iv[..KEY_LEN]);
+        let mut tg_dec_iv_arr = [0u8; IV_LEN];
+        tg_dec_iv_arr.copy_from_slice(&dec_key_iv[KEY_LEN..]);
+        let tg_dec_iv = u128::from_be_bytes(tg_dec_iv_arr);
 
         return (nonce, tg_enc_key, tg_enc_iv, tg_dec_key, tg_dec_iv);
     }
@@ -403,13 +929,19 @@ pub fn generate_tg_nonce(
 /// Encrypt nonce for sending to Telegram and return cipher objects with correct counter state
 pub fn encrypt_tg_nonce_with_ciphers(nonce: &[u8; HANDSHAKE_LEN]) -> (Vec<u8>, AesCtr, AesCtr) {
     let enc_key_iv = &nonce[SKIP_LEN..SKIP_LEN + KEY_LEN + IV_LEN];
-    let dec_key_iv: Vec<u8> = enc_key_iv.iter().rev().copied().collect();
+    let dec_key_iv = Zeroizing::new(enc_key_iv.iter().rev().copied().collect::<Vec<u8>>());
 
-    let enc_key: [u8; 32] = enc_key_iv[..KEY_LEN].try_into().unwrap();
-    let enc_iv = u128::from_be_bytes(enc_key_iv[KEY_LEN..].try_into().unwrap());
+    let mut enc_key = [0u8; 32];
+    enc_key.copy_from_slice(&enc_key_iv[..KEY_LEN]);
+    let mut enc_iv_arr = [0u8; IV_LEN];
+    enc_iv_arr.copy_from_slice(&enc_key_iv[KEY_LEN..]);
+    let enc_iv = u128::from_be_bytes(enc_iv_arr);
 
-    let dec_key: [u8; 32] = dec_key_iv[..KEY_LEN].try_into().unwrap();
-    let dec_iv = u128::from_be_bytes(dec_key_iv[KEY_LEN..].try_into().unwrap());
+    let mut dec_key = [0u8; 32];
+    dec_key.copy_from_slice(&dec_key_iv[..KEY_LEN]);
+    let mut dec_iv_arr = [0u8; IV_LEN];
+    dec_iv_arr.copy_from_slice(&dec_key_iv[KEY_LEN..]);
+    let dec_iv = u128::from_be_bytes(dec_iv_arr);
 
     let mut encryptor = AesCtr::new(&enc_key, enc_iv);
     let encrypted_full = encryptor.encrypt(nonce);  // counter: 0 → 4
@@ -418,91 +950,37 @@ pub fn encrypt_tg_nonce_with_ciphers(nonce: &[u8; HANDSHAKE_LEN]) -> (Vec<u8>, A
     result.extend_from_slice(&encrypted_full[PROTO_TAG_POS..]);
 
     let decryptor = AesCtr::new(&dec_key, dec_iv);
+    enc_key.zeroize();
+    dec_key.zeroize();
 
     (result, encryptor, decryptor)
 }
 
 /// Encrypt nonce for sending to Telegram (legacy function for compatibility)
+
 pub fn encrypt_tg_nonce(nonce: &[u8; HANDSHAKE_LEN]) -> Vec<u8> {
     let (encrypted, _, _) = encrypt_tg_nonce_with_ciphers(nonce);
     encrypted
 }
 
 #[cfg(test)]
-mod tests {
-    use super::*;
+#[path = "handshake_security_tests.rs"]
+mod security_tests;
 
-    #[test]
-    fn test_generate_tg_nonce() {
-        let client_dec_key = [0x42u8; 32];
-        let client_dec_iv = 12345u128;
-        let client_enc_key = [0x24u8; 32];
-        let client_enc_iv = 54321u128;
+#[cfg(test)]
+#[path = "handshake_adversarial_tests.rs"]
+mod adversarial_tests;
 
-        let rng = SecureRandom::new();
-        let (nonce, _tg_enc_key, _tg_enc_iv, _tg_dec_key, _tg_dec_iv) = 
-            generate_tg_nonce(
-                ProtoTag::Secure,
-                2,
-                &client_dec_key,
-                client_dec_iv,
-                &client_enc_key,
-                client_enc_iv,
-                &rng,
-                false,
-            );
+#[cfg(test)]
+#[path = "handshake_fuzz_security_tests.rs"]
+mod fuzz_security_tests;
 
-        assert_eq!(nonce.len(), HANDSHAKE_LEN);
+/// Compile-time guard: HandshakeSuccess holds cryptographic key material and
+/// must never be Copy.  A Copy impl would allow silent key duplication,
+/// undermining the zeroize-on-drop guarantee.
+mod compile_time_security_checks {
+    use super::HandshakeSuccess;
+    use static_assertions::assert_not_impl_all;
 
-        let tag_bytes: [u8; 4] = nonce[PROTO_TAG_POS..PROTO_TAG_POS + 4].try_into().unwrap();
-        assert_eq!(ProtoTag::from_bytes(tag_bytes), Some(ProtoTag::Secure));
-    }
-
-    #[test]
-    fn test_encrypt_tg_nonce() {
-        let client_dec_key = [0x42u8; 32];
-        let client_dec_iv = 12345u128;
-        let client_enc_key = [0x24u8; 32];
-        let client_enc_iv = 54321u128;
-
-        let rng = SecureRandom::new();
-        let (nonce, _, _, _, _) = 
-            generate_tg_nonce(
-                ProtoTag::Secure,
-                2,
-                &client_dec_key,
-                client_dec_iv,
-                &client_enc_key,
-                client_enc_iv,
-                &rng,
-                false,
-            );
-
-        let encrypted = encrypt_tg_nonce(&nonce);
-
-        assert_eq!(encrypted.len(), HANDSHAKE_LEN);
-        assert_eq!(&encrypted[..PROTO_TAG_POS], &nonce[..PROTO_TAG_POS]);
-        assert_ne!(&encrypted[PROTO_TAG_POS..], &nonce[PROTO_TAG_POS..]);
-    }
-
-    #[test]
-    fn test_handshake_success_zeroize_on_drop() {
-        let success = HandshakeSuccess {
-            user: "test".to_string(),
-            dc_idx: 2,
-            proto_tag: ProtoTag::Secure,
-            dec_key: [0xAA; 32],
-            dec_iv: 0xBBBBBBBB,
-            enc_key: [0xCC; 32],
-            enc_iv: 0xDDDDDDDD,
-            peer: "127.0.0.1:1234".parse().unwrap(),
-            is_tls: true,
-        };
-
-        assert_eq!(success.dec_key, [0xAA; 32]);
-        assert_eq!(success.enc_key, [0xCC; 32]);
-
-        drop(success);
-        // Drop impl zeroizes key material without panic
-    }
+    assert_not_impl_all!(HandshakeSuccess: Copy, Clone);
 }

src/proxy/handshake_adversarial_tests.rs

@@ -0,0 +1,231 @@
+use super::*;
+use std::sync::Arc;
+use std::net::{IpAddr, Ipv4Addr};
+use std::time::{Duration, Instant};
+use crate::crypto::sha256;
+
+fn make_valid_mtproto_handshake(secret_hex: &str, proto_tag: ProtoTag, dc_idx: i16) -> [u8; HANDSHAKE_LEN] {
+    let secret = hex::decode(secret_hex).expect("secret hex must decode");
+    let mut handshake = [0x5Au8; HANDSHAKE_LEN];
+    for (idx, b) in handshake[SKIP_LEN..SKIP_LEN + PREKEY_LEN + IV_LEN]
+        .iter_mut()
+        .enumerate()
+    {
+        *b = (idx as u8).wrapping_add(1);
+    }
+
+    let dec_prekey = &handshake[SKIP_LEN..SKIP_LEN + PREKEY_LEN];
+    let dec_iv_bytes = &handshake[SKIP_LEN + PREKEY_LEN..SKIP_LEN + PREKEY_LEN + IV_LEN];
+
+    let mut dec_key_input = Vec::with_capacity(PREKEY_LEN + secret.len());
+    dec_key_input.extend_from_slice(dec_prekey);
+    dec_key_input.extend_from_slice(&secret);
+    let dec_key = sha256(&dec_key_input);
+
+    let mut dec_iv_arr = [0u8; IV_LEN];
+    dec_iv_arr.copy_from_slice(dec_iv_bytes);
+    let dec_iv = u128::from_be_bytes(dec_iv_arr);
+
+    let mut stream = AesCtr::new(&dec_key, dec_iv);
+    let keystream = stream.encrypt(&[0u8; HANDSHAKE_LEN]);
+
+    let mut target_plain = [0u8; HANDSHAKE_LEN];
+    target_plain[PROTO_TAG_POS..PROTO_TAG_POS + 4].copy_from_slice(&proto_tag.to_bytes());
+    target_plain[DC_IDX_POS..DC_IDX_POS + 2].copy_from_slice(&dc_idx.to_le_bytes());
+
+    for idx in PROTO_TAG_POS..HANDSHAKE_LEN {
+        handshake[idx] = target_plain[idx] ^ keystream[idx];
+    }
+
+    handshake
+}
+
+fn auth_probe_test_guard() -> std::sync::MutexGuard<'static, ()> {
+    auth_probe_test_lock()
+        .lock()
+        .unwrap_or_else(|poisoned| poisoned.into_inner())
+}
+
+fn test_config_with_secret_hex(secret_hex: &str) -> ProxyConfig {
+    let mut cfg = ProxyConfig::default();
+    cfg.access.users.clear();
+    cfg.access.users.insert("user".to_string(), secret_hex.to_string());
+    cfg.access.ignore_time_skew = true;
+    cfg.general.modes.secure = true;
+    cfg
+}
+
+// ------------------------------------------------------------------
+// Mutational Bit-Flipping Tests (OWASP ASVS 5.1.4)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn mtproto_handshake_bit_flip_anywhere_rejected() {
+    let _guard = auth_probe_test_guard();
+    clear_auth_probe_state_for_testing();
+
+    let secret_hex = "11223344556677889900aabbccddeeff";
+    let base = make_valid_mtproto_handshake(secret_hex, ProtoTag::Secure, 2);
+    let config = test_config_with_secret_hex(secret_hex);
+    let replay_checker = ReplayChecker::new(128, Duration::from_secs(60));
+    let peer: SocketAddr = "192.0.2.1:12345".parse().unwrap();
+
+    // Baseline check
+    let res = handle_mtproto_handshake(&base, tokio::io::empty(), tokio::io::sink(), peer, &config, &replay_checker, false, None).await;
+    match res {
+        HandshakeResult::Success(_) => {},
+        _ => panic!("Baseline failed: expected Success"),
+    }
+
+    // Flip bits in the encrypted part (beyond the key material)
+    for byte_pos in SKIP_LEN..HANDSHAKE_LEN {
+        let mut h = base;
+        h[byte_pos] ^= 0x01; // Flip 1 bit
+        let res = handle_mtproto_handshake(&h, tokio::io::empty(), tokio::io::sink(), peer, &config, &replay_checker, false, None).await;
+        assert!(matches!(res, HandshakeResult::BadClient { .. }), "Flip at byte {byte_pos} bit 0 must be rejected");
+    }
+}
+
+// ------------------------------------------------------------------
+// Adversarial Probing / Timing Neutrality (OWASP ASVS 5.1.7)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn mtproto_handshake_timing_neutrality_mocked() {
+    let secret_hex = "00112233445566778899aabbccddeeff";
+    let base = make_valid_mtproto_handshake(secret_hex, ProtoTag::Secure, 1);
+    let config = test_config_with_secret_hex(secret_hex);
+    let replay_checker = ReplayChecker::new(128, Duration::from_secs(60));
+    let peer: SocketAddr = "192.0.2.2:54321".parse().unwrap();
+
+    const ITER: usize = 50;
+    
+    let mut start = Instant::now();
+    for _ in 0..ITER {
+        let _ = handle_mtproto_handshake(&base, tokio::io::empty(), tokio::io::sink(), peer, &config, &replay_checker, false, None).await;
+    }
+    let duration_success = start.elapsed();
+
+    start = Instant::now();
+    for i in 0..ITER {
+        let mut h = base;
+        h[SKIP_LEN + (i % 48)] ^= 0xFF; 
+        let _ = handle_mtproto_handshake(&h, tokio::io::empty(), tokio::io::sink(), peer, &config, &replay_checker, false, None).await;
+    }
+    let duration_fail = start.elapsed();
+
+    let avg_diff_ms = (duration_success.as_millis() as f64 - duration_fail.as_millis() as f64).abs() / ITER as f64;
+    
+    // Threshold (loose for CI)
+    assert!(avg_diff_ms < 100.0, "Timing difference too large: {} ms/iter", avg_diff_ms);
+}
+
+// ------------------------------------------------------------------
+// Stress Tests (OWASP ASVS 5.1.6)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn auth_probe_throttle_saturation_stress() {
+    let _guard = auth_probe_test_guard();
+    clear_auth_probe_state_for_testing();
+
+    let now = Instant::now();
+    
+    // Record enough failures for one IP to trigger backoff
+    let target_ip = IpAddr::V4(Ipv4Addr::new(1, 1, 1, 1));
+    for _ in 0..AUTH_PROBE_BACKOFF_START_FAILS {
+        auth_probe_record_failure(target_ip, now);
+    }
+    
+    assert!(auth_probe_is_throttled(target_ip, now));
+
+    // Stress test with many unique IPs
+    for i in 0..500u32 {
+        let ip = IpAddr::V4(Ipv4Addr::new(203, 0, 113, (i % 256) as u8));
+        auth_probe_record_failure(ip, now);
+    }
+
+    let tracked = AUTH_PROBE_STATE
+        .get()
+        .map(|state| state.len())
+        .unwrap_or(0);
+    assert!(
+        tracked <= AUTH_PROBE_TRACK_MAX_ENTRIES,
+        "auth probe state grew past hard cap: {tracked} > {AUTH_PROBE_TRACK_MAX_ENTRIES}"
+    );
+}
+
+#[tokio::test]
+async fn mtproto_handshake_abridged_prefix_rejected() {
+    let _guard = auth_probe_test_guard();
+    clear_auth_probe_state_for_testing();
+
+    let mut handshake = [0x5Au8; HANDSHAKE_LEN];
+    handshake[0] = 0xef; // Abridged prefix
+    let config = ProxyConfig::default();
+    let replay_checker = ReplayChecker::new(128, Duration::from_secs(60));
+    let peer: SocketAddr = "192.0.2.3:12345".parse().unwrap();
+
+    let res = handle_mtproto_handshake(&handshake, tokio::io::empty(), tokio::io::sink(), peer, &config, &replay_checker, false, None).await;
+    // MTProxy stops immediately on 0xef
+    assert!(matches!(res, HandshakeResult::BadClient { .. }));
+}
+
+#[tokio::test]
+async fn mtproto_handshake_preferred_user_mismatch_continues() {
+    let _guard = auth_probe_test_guard();
+    clear_auth_probe_state_for_testing();
+
+    let secret1_hex = "11111111111111111111111111111111";
+    let secret2_hex = "22222222222222222222222222222222";
+    
+    let base = make_valid_mtproto_handshake(secret2_hex, ProtoTag::Secure, 1);
+    let mut config = ProxyConfig::default();
+    config.access.users.insert("user1".to_string(), secret1_hex.to_string());
+    config.access.users.insert("user2".to_string(), secret2_hex.to_string());
+    config.access.ignore_time_skew = true;
+    config.general.modes.secure = true;
+
+    let replay_checker = ReplayChecker::new(128, Duration::from_secs(60));
+    let peer: SocketAddr = "192.0.2.4:12345".parse().unwrap();
+
+    // Even if we prefer user1, if user2 matches, it should succeed.
+    let res = handle_mtproto_handshake(&base, tokio::io::empty(), tokio::io::sink(), peer, &config, &replay_checker, false, Some("user1")).await;
+    if let HandshakeResult::Success((_, _, success)) = res {
+        assert_eq!(success.user, "user2");
+    } else {
+        panic!("Handshake failed even though user2 matched");
+    }
+}
+
+#[tokio::test]
+async fn mtproto_handshake_concurrent_flood_stability() {
+    let _guard = auth_probe_test_guard();
+    clear_auth_probe_state_for_testing();
+
+    let secret_hex = "00112233445566778899aabbccddeeff";
+    let base = make_valid_mtproto_handshake(secret_hex, ProtoTag::Secure, 1);
+    let mut config = test_config_with_secret_hex(secret_hex);
+    config.access.ignore_time_skew = true;
+    let replay_checker = Arc::new(ReplayChecker::new(1024, Duration::from_secs(60)));
+    let config = Arc::new(config);
+    
+    let mut tasks = Vec::new();
+    for i in 0..50 {
+        let base = base;
+        let config = Arc::clone(&config);
+        let replay_checker = Arc::clone(&replay_checker);
+        let peer: SocketAddr = format!("192.0.2.{}:12345", (i % 254) + 1).parse().unwrap();
+        
+        tasks.push(tokio::spawn(async move {
+            let res = handle_mtproto_handshake(&base, tokio::io::empty(), tokio::io::sink(), peer, &config, &replay_checker, false, None).await;
+            matches!(res, HandshakeResult::Success(_))
+        }));
+    }
+    
+    // We don't necessarily care if they all succeed (some might fail due to replay if they hit the same chunk),
+    // but the system must not panic or hang.
+    for task in tasks {
+        let _ = task.await.unwrap();
+    }
+}

src/proxy/handshake_fuzz_security_tests.rs

@@ -0,0 +1,270 @@
+use super::*;
+use crate::config::ProxyConfig;
+use crate::crypto::AesCtr;
+use crate::crypto::sha256;
+use crate::protocol::constants::ProtoTag;
+use crate::stats::ReplayChecker;
+use std::net::SocketAddr;
+use std::sync::MutexGuard;
+use tokio::time::{timeout, Duration as TokioDuration};
+
+fn make_mtproto_handshake_with_proto_bytes(
+    secret_hex: &str,
+    proto_bytes: [u8; 4],
+    dc_idx: i16,
+) -> [u8; HANDSHAKE_LEN] {
+    let secret = hex::decode(secret_hex).expect("secret hex must decode");
+    let mut handshake = [0x5Au8; HANDSHAKE_LEN];
+    for (idx, b) in handshake[SKIP_LEN..SKIP_LEN + PREKEY_LEN + IV_LEN]
+        .iter_mut()
+        .enumerate()
+    {
+        *b = (idx as u8).wrapping_add(1);
+    }
+
+    let dec_prekey = &handshake[SKIP_LEN..SKIP_LEN + PREKEY_LEN];
+    let dec_iv_bytes = &handshake[SKIP_LEN + PREKEY_LEN..SKIP_LEN + PREKEY_LEN + IV_LEN];
+
+    let mut dec_key_input = Vec::with_capacity(PREKEY_LEN + secret.len());
+    dec_key_input.extend_from_slice(dec_prekey);
+    dec_key_input.extend_from_slice(&secret);
+    let dec_key = sha256(&dec_key_input);
+
+    let mut dec_iv_arr = [0u8; IV_LEN];
+    dec_iv_arr.copy_from_slice(dec_iv_bytes);
+    let dec_iv = u128::from_be_bytes(dec_iv_arr);
+
+    let mut stream = AesCtr::new(&dec_key, dec_iv);
+    let keystream = stream.encrypt(&[0u8; HANDSHAKE_LEN]);
+
+    let mut target_plain = [0u8; HANDSHAKE_LEN];
+    target_plain[PROTO_TAG_POS..PROTO_TAG_POS + 4].copy_from_slice(&proto_bytes);
+    target_plain[DC_IDX_POS..DC_IDX_POS + 2].copy_from_slice(&dc_idx.to_le_bytes());
+
+    for idx in PROTO_TAG_POS..HANDSHAKE_LEN {
+        handshake[idx] = target_plain[idx] ^ keystream[idx];
+    }
+
+    handshake
+}
+
+fn make_valid_mtproto_handshake(secret_hex: &str, proto_tag: ProtoTag, dc_idx: i16) -> [u8; HANDSHAKE_LEN] {
+    make_mtproto_handshake_with_proto_bytes(secret_hex, proto_tag.to_bytes(), dc_idx)
+}
+
+fn test_config_with_secret_hex(secret_hex: &str) -> ProxyConfig {
+    let mut cfg = ProxyConfig::default();
+    cfg.access.users.clear();
+    cfg.access.users.insert("user".to_string(), secret_hex.to_string());
+    cfg.access.ignore_time_skew = true;
+    cfg.general.modes.secure = true;
+    cfg
+}
+
+fn auth_probe_test_guard() -> MutexGuard<'static, ()> {
+    auth_probe_test_lock()
+        .lock()
+        .unwrap_or_else(|poisoned| poisoned.into_inner())
+}
+
+#[tokio::test]
+async fn mtproto_handshake_duplicate_digest_is_replayed_on_second_attempt() {
+    let _guard = auth_probe_test_guard();
+    clear_auth_probe_state_for_testing();
+
+    let secret_hex = "11223344556677889900aabbccddeeff";
+    let base = make_valid_mtproto_handshake(secret_hex, ProtoTag::Secure, 2);
+    let config = test_config_with_secret_hex(secret_hex);
+    let replay_checker = ReplayChecker::new(128, TokioDuration::from_secs(60));
+    let peer: SocketAddr = "192.0.2.1:12345".parse().unwrap();
+
+    let first = handle_mtproto_handshake(
+        &base,
+        tokio::io::empty(),
+        tokio::io::sink(),
+        peer,
+        &config,
+        &replay_checker,
+        false,
+        None,
+    )
+    .await;
+    assert!(matches!(first, HandshakeResult::Success(_)));
+
+    let second = handle_mtproto_handshake(
+        &base,
+        tokio::io::empty(),
+        tokio::io::sink(),
+        peer,
+        &config,
+        &replay_checker,
+        false,
+        None,
+    )
+    .await;
+    assert!(matches!(second, HandshakeResult::BadClient { .. }));
+
+    clear_auth_probe_state_for_testing();
+}
+
+#[tokio::test]
+async fn mtproto_handshake_fuzz_corpus_never_panics_and_stays_fail_closed() {
+    let _guard = auth_probe_test_guard();
+    clear_auth_probe_state_for_testing();
+
+    let secret_hex = "00112233445566778899aabbccddeeff";
+    let base = make_valid_mtproto_handshake(secret_hex, ProtoTag::Secure, 1);
+    let config = test_config_with_secret_hex(secret_hex);
+    let replay_checker = ReplayChecker::new(128, TokioDuration::from_secs(60));
+    let peer: SocketAddr = "192.0.2.2:54321".parse().unwrap();
+
+    let mut corpus = Vec::<[u8; HANDSHAKE_LEN]>::new();
+
+    corpus.push(make_mtproto_handshake_with_proto_bytes(
+        secret_hex,
+        [0x00, 0x00, 0x00, 0x00],
+        1,
+    ));
+    corpus.push(make_mtproto_handshake_with_proto_bytes(
+        secret_hex,
+        [0xff, 0xff, 0xff, 0xff],
+        1,
+    ));
+    corpus.push(make_valid_mtproto_handshake(
+        "ffeeddccbbaa99887766554433221100",
+        ProtoTag::Secure,
+        1,
+    ));
+
+    let mut seed = 0xF0F0_F00D_BAAD_CAFEu64;
+    for _ in 0..32 {
+        let mut mutated = base;
+        for _ in 0..4 {
+            seed = seed.wrapping_mul(2862933555777941757).wrapping_add(3037000493);
+            let idx = SKIP_LEN + (seed as usize % (PREKEY_LEN + IV_LEN));
+            mutated[idx] ^= ((seed >> 19) as u8).wrapping_add(1);
+        }
+        corpus.push(mutated);
+    }
+
+    for (idx, input) in corpus.into_iter().enumerate() {
+        let result = timeout(
+            TokioDuration::from_secs(1),
+            handle_mtproto_handshake(
+                &input,
+                tokio::io::empty(),
+                tokio::io::sink(),
+                peer,
+                &config,
+                &replay_checker,
+                false,
+                None,
+            ),
+        )
+        .await
+        .expect("fuzzed handshake must complete in time");
+
+        assert!(
+            matches!(result, HandshakeResult::BadClient { .. }),
+            "corpus item {idx} must fail closed"
+        );
+    }
+
+    clear_auth_probe_state_for_testing();
+}
+
+#[tokio::test]
+async fn mtproto_handshake_mixed_corpus_never_panics_and_exact_duplicates_are_rejected() {
+    let _guard = auth_probe_test_guard();
+    clear_auth_probe_state_for_testing();
+
+    let secret_hex = "99887766554433221100ffeeddccbbaa";
+    let base = make_valid_mtproto_handshake(secret_hex, ProtoTag::Secure, 4);
+    let config = test_config_with_secret_hex(secret_hex);
+    let replay_checker = ReplayChecker::new(256, TokioDuration::from_secs(60));
+    let peer: SocketAddr = "192.0.2.44:45444".parse().unwrap();
+
+    let first = timeout(
+        TokioDuration::from_secs(1),
+        handle_mtproto_handshake(
+            &base,
+            tokio::io::empty(),
+            tokio::io::sink(),
+            peer,
+            &config,
+            &replay_checker,
+            false,
+            None,
+        ),
+    )
+    .await
+    .expect("base handshake must not hang");
+    assert!(matches!(first, HandshakeResult::Success(_)));
+
+    let replay = timeout(
+        TokioDuration::from_secs(1),
+        handle_mtproto_handshake(
+            &base,
+            tokio::io::empty(),
+            tokio::io::sink(),
+            peer,
+            &config,
+            &replay_checker,
+            false,
+            None,
+        ),
+    )
+    .await
+    .expect("duplicate handshake must not hang");
+    assert!(matches!(replay, HandshakeResult::BadClient { .. }));
+
+    let mut corpus = Vec::<[u8; HANDSHAKE_LEN]>::new();
+
+    let mut prekey_flip = base;
+    prekey_flip[SKIP_LEN] ^= 0x80;
+    corpus.push(prekey_flip);
+
+    let mut iv_flip = base;
+    iv_flip[SKIP_LEN + PREKEY_LEN] ^= 0x01;
+    corpus.push(iv_flip);
+
+    let mut tail_flip = base;
+    tail_flip[SKIP_LEN + PREKEY_LEN + IV_LEN - 1] ^= 0x40;
+    corpus.push(tail_flip);
+
+    let mut seed = 0xBADC_0FFE_EE11_4242u64;
+    for _ in 0..24 {
+        let mut mutated = base;
+        for _ in 0..3 {
+            seed = seed.wrapping_mul(6364136223846793005).wrapping_add(1);
+            let idx = SKIP_LEN + (seed as usize % (PREKEY_LEN + IV_LEN));
+            mutated[idx] ^= ((seed >> 16) as u8).wrapping_add(1);
+        }
+        corpus.push(mutated);
+    }
+
+    for (idx, input) in corpus.iter().enumerate() {
+        let result = timeout(
+            TokioDuration::from_secs(1),
+            handle_mtproto_handshake(
+                input,
+                tokio::io::empty(),
+                tokio::io::sink(),
+                peer,
+                &config,
+                &replay_checker,
+                false,
+                None,
+            ),
+        )
+        .await
+        .expect("fuzzed handshake must complete in time");
+
+        assert!(
+            matches!(result, HandshakeResult::BadClient { .. }),
+            "mixed corpus item {idx} must fail closed"
+        );
+    }
+
+    clear_auth_probe_state_for_testing();
+}

src/proxy/masking.rs

@@ -7,19 +7,90 @@ use tokio::net::TcpStream;
 #[cfg(unix)]
 use tokio::net::UnixStream;
 use tokio::io::{AsyncRead, AsyncWrite, AsyncReadExt, AsyncWriteExt};
-use tokio::time::timeout;
+use tokio::time::{Instant, timeout};
 use tracing::debug;
 use crate::config::ProxyConfig;
 use crate::network::dns_overrides::resolve_socket_addr;
 use crate::stats::beobachten::BeobachtenStore;
 use crate::transport::proxy_protocol::{ProxyProtocolV1Builder, ProxyProtocolV2Builder};
 
+#[cfg(not(test))]
 const MASK_TIMEOUT: Duration = Duration::from_secs(5);
+#[cfg(test)]
+const MASK_TIMEOUT: Duration = Duration::from_millis(50);
 /// Maximum duration for the entire masking relay.
 /// Limits resource consumption from slow-loris attacks and port scanners.
+#[cfg(not(test))]
 const MASK_RELAY_TIMEOUT: Duration = Duration::from_secs(60);
+#[cfg(test)]
+const MASK_RELAY_TIMEOUT: Duration = Duration::from_millis(200);
+#[cfg(not(test))]
+const MASK_RELAY_IDLE_TIMEOUT: Duration = Duration::from_secs(5);
+#[cfg(test)]
+const MASK_RELAY_IDLE_TIMEOUT: Duration = Duration::from_millis(100);
 const MASK_BUFFER_SIZE: usize = 8192;
 
+async fn copy_with_idle_timeout<R, W>(reader: &mut R, writer: &mut W)
+where
+    R: AsyncRead + Unpin,
+    W: AsyncWrite + Unpin,
+{
+    let mut buf = [0u8; MASK_BUFFER_SIZE];
+    loop {
+        let read_res = timeout(MASK_RELAY_IDLE_TIMEOUT, reader.read(&mut buf)).await;
+        let n = match read_res {
+            Ok(Ok(n)) => n,
+            Ok(Err(_)) | Err(_) => break,
+        };
+        if n == 0 {
+            break;
+        }
+
+        let write_res = timeout(MASK_RELAY_IDLE_TIMEOUT, writer.write_all(&buf[..n])).await;
+        match write_res {
+            Ok(Ok(())) => {}
+            Ok(Err(_)) | Err(_) => break,
+        }
+    }
+}
+
+async fn write_proxy_header_with_timeout<W>(mask_write: &mut W, header: &[u8]) -> bool
+where
+    W: AsyncWrite + Unpin,
+{
+    match timeout(MASK_TIMEOUT, mask_write.write_all(header)).await {
+        Ok(Ok(())) => true,
+        Ok(Err(_)) => false,
+        Err(_) => {
+            debug!("Timeout writing proxy protocol header to mask backend");
+            false
+        }
+    }
+}
+
+async fn consume_client_data_with_timeout<R>(reader: R)
+where
+    R: AsyncRead + Unpin,
+{
+    if timeout(MASK_RELAY_TIMEOUT, consume_client_data(reader)).await.is_err() {
+        debug!("Timed out while consuming client data on masking fallback path");
+    }
+}
+
+async fn wait_mask_connect_budget(started: Instant) {
+    let elapsed = started.elapsed();
+    if elapsed < MASK_TIMEOUT {
+        tokio::time::sleep(MASK_TIMEOUT - elapsed).await;
+    }
+}
+
+async fn wait_mask_outcome_budget(started: Instant) {
+    let elapsed = started.elapsed();
+    if elapsed < MASK_TIMEOUT {
+        tokio::time::sleep(MASK_TIMEOUT - elapsed).await;
+    }
+}
+
 /// Detect client type based on initial data
 fn detect_client_type(data: &[u8]) -> &'static str {
     // Check for HTTP request
@@ -71,13 +142,15 @@ where
 
     if !config.censorship.mask {
         // Masking disabled, just consume data
-        consume_client_data(reader).await;
+        consume_client_data_with_timeout(reader).await;
         return;
     }
 
     // Connect via Unix socket or TCP
     #[cfg(unix)]
     if let Some(ref sock_path) = config.censorship.mask_unix_sock {
+        let outcome_started = Instant::now();
+        let connect_started = Instant::now();
         debug!(
             client_type = client_type,
             sock = %sock_path,
@@ -107,21 +180,26 @@ where
                     }
                 };
                 if let Some(header) = proxy_header {
-                    if mask_write.write_all(&header).await.is_err() {
+                    if !write_proxy_header_with_timeout(&mut mask_write, &header).await {
+                        wait_mask_outcome_budget(outcome_started).await;
                         return;
                     }
                 }
                 if timeout(MASK_RELAY_TIMEOUT, relay_to_mask(reader, writer, mask_read, mask_write, initial_data)).await.is_err() {
                     debug!("Mask relay timed out (unix socket)");
                 }
+                wait_mask_outcome_budget(outcome_started).await;
             }
             Ok(Err(e)) => {
+                wait_mask_connect_budget(connect_started).await;
                 debug!(error = %e, "Failed to connect to mask unix socket");
-                consume_client_data(reader).await;
+                consume_client_data_with_timeout(reader).await;
+                wait_mask_outcome_budget(outcome_started).await;
             }
             Err(_) => {
                 debug!("Timeout connecting to mask unix socket");
-                consume_client_data(reader).await;
+                consume_client_data_with_timeout(reader).await;
+                wait_mask_outcome_budget(outcome_started).await;
             }
         }
         return;
@@ -143,6 +221,8 @@ where
     let mask_addr = resolve_socket_addr(mask_host, mask_port)
         .map(|addr| addr.to_string())
         .unwrap_or_else(|| format!("{}:{}", mask_host, mask_port));
+    let outcome_started = Instant::now();
+    let connect_started = Instant::now();
     let connect_result = timeout(MASK_TIMEOUT, TcpStream::connect(&mask_addr)).await;
     match connect_result {
         Ok(Ok(stream)) => {
@@ -166,21 +246,26 @@ where
 
             let (mask_read, mut mask_write) = stream.into_split();
             if let Some(header) = proxy_header {
-                if mask_write.write_all(&header).await.is_err() {
+                if !write_proxy_header_with_timeout(&mut mask_write, &header).await {
+                    wait_mask_outcome_budget(outcome_started).await;
                     return;
                 }
             }
             if timeout(MASK_RELAY_TIMEOUT, relay_to_mask(reader, writer, mask_read, mask_write, initial_data)).await.is_err() {
                 debug!("Mask relay timed out");
             }
+            wait_mask_outcome_budget(outcome_started).await;
         }
         Ok(Err(e)) => {
+            wait_mask_connect_budget(connect_started).await;
             debug!(error = %e, "Failed to connect to mask host");
-            consume_client_data(reader).await;
+            consume_client_data_with_timeout(reader).await;
+            wait_mask_outcome_budget(outcome_started).await;
         }
         Err(_) => {
             debug!("Timeout connecting to mask host");
-            consume_client_data(reader).await;
+            consume_client_data_with_timeout(reader).await;
+            wait_mask_outcome_budget(outcome_started).await;
         }
     }
 }
@@ -203,47 +288,20 @@ where
     if mask_write.write_all(initial_data).await.is_err() {
         return;
     }
-
-    // Relay traffic
-    let c2m = tokio::spawn(async move {
-        let mut buf = vec![0u8; MASK_BUFFER_SIZE];
-        loop {
-            match reader.read(&mut buf).await {
-                Ok(0) | Err(_) => {
-                    let _ = mask_write.shutdown().await;
-                    break;
-                }
-                Ok(n) => {
-                    if mask_write.write_all(&buf[..n]).await.is_err() {
-                        break;
-                    }
-                }
-            }
-        }
-    });
-
-    let m2c = tokio::spawn(async move {
-        let mut buf = vec![0u8; MASK_BUFFER_SIZE];
-        loop {
-            match mask_read.read(&mut buf).await {
-                Ok(0) | Err(_) => {
-                    let _ = writer.shutdown().await;
-                    break;
-                }
-                Ok(n) => {
-                    if writer.write_all(&buf[..n]).await.is_err() {
-                        break;
-                    }
-                }
-            }
-        }
-    });
-
-    // Wait for either to complete
-    tokio::select! {
-        _ = c2m => {}
-        _ = m2c => {}
+    if mask_write.flush().await.is_err() {
+        return;
     }
+
+    let _ = tokio::join!(
+        async {
+            copy_with_idle_timeout(&mut reader, &mut mask_write).await;
+            let _ = mask_write.shutdown().await;
+        },
+        async {
+            copy_with_idle_timeout(&mut mask_read, &mut writer).await;
+            let _ = writer.shutdown().await;
+        }
+    );
 }
 
 /// Just consume all data from client without responding
@@ -255,3 +313,11 @@ async fn consume_client_data<R: AsyncRead + Unpin>(mut reader: R) {
         }
     }
 }
+
+#[cfg(test)]
+#[path = "masking_security_tests.rs"]
+mod security_tests;
+
+#[cfg(test)]
+#[path = "masking_adversarial_tests.rs"]
+mod adversarial_tests;

src/proxy/masking_adversarial_tests.rs

@@ -0,0 +1,213 @@
+use super::*;
+use std::sync::Arc;
+use tokio::io::duplex;
+use tokio::net::TcpListener;
+use tokio::time::{Instant, Duration};
+use crate::config::ProxyConfig;
+use crate::stats::beobachten::BeobachtenStore;
+
+// ------------------------------------------------------------------
+// Probing Indistinguishability (OWASP ASVS 5.1.7)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn masking_probes_indistinguishable_timing() {
+    let mut config = ProxyConfig::default();
+    config.censorship.mask = true;
+    config.censorship.mask_host = Some("127.0.0.1".to_string());
+    config.censorship.mask_port = 80; // Should timeout/refuse
+    
+    let peer: SocketAddr = "192.0.2.10:443".parse().unwrap();
+    let local_addr: SocketAddr = "127.0.0.1:443".parse().unwrap();
+    let beobachten = BeobachtenStore::new();
+
+    // Test different probe types
+    let probes = vec![
+        (b"GET / HTTP/1.1\r\nHost: x\r\n\r\n".to_vec(), "HTTP"),
+        (b"SSH-2.0-probe".to_vec(), "SSH"),
+        (vec![0x16, 0x03, 0x03, 0x00, 0x05, 0x01, 0x00, 0x00, 0x01, 0x00], "TLS-scanner"),
+        (vec![0x42; 5], "port-scanner"),
+    ];
+
+    for (probe, type_name) in probes {
+        let (client_reader, _client_writer) = duplex(256);
+        let (_client_visible_reader, client_visible_writer) = duplex(256);
+        
+        let start = Instant::now();
+        handle_bad_client(
+            client_reader,
+            client_visible_writer,
+            &probe,
+            peer,
+            local_addr,
+            &config,
+            &beobachten,
+        ).await;
+        
+        let elapsed = start.elapsed();
+        
+        // We expect any outcome to take roughly MASK_TIMEOUT (50ms in tests)
+        // to mask whether the backend was reachable or refused.
+        assert!(elapsed >= Duration::from_millis(30), "Probe {type_name} finished too fast: {elapsed:?}");
+    }
+}
+
+// ------------------------------------------------------------------
+// Masking Budget Stress Tests (OWASP ASVS 5.1.6)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn masking_budget_stress_under_load() {
+    let mut config = ProxyConfig::default();
+    config.censorship.mask = true;
+    config.censorship.mask_host = Some("127.0.0.1".to_string());
+    config.censorship.mask_port = 1; // Unlikely port
+
+    let peer: SocketAddr = "192.0.2.20:443".parse().unwrap();
+    let local_addr: SocketAddr = "127.0.0.1:443".parse().unwrap();
+    let beobachten = Arc::new(BeobachtenStore::new());
+
+    let mut tasks = Vec::new();
+    for _ in 0..50 {
+        let (client_reader, _client_writer) = duplex(256);
+        let (_client_visible_reader, client_visible_writer) = duplex(256);
+        let config = config.clone();
+        let beobachten = Arc::clone(&beobachten);
+        
+        tasks.push(tokio::spawn(async move {
+            let start = Instant::now();
+            handle_bad_client(
+                client_reader,
+                client_visible_writer,
+                b"probe",
+                peer,
+                local_addr,
+                &config,
+                &beobachten,
+            ).await;
+            start.elapsed()
+        }));
+    }
+
+    for task in tasks {
+        let elapsed = task.await.unwrap();
+        assert!(elapsed >= Duration::from_millis(30), "Stress probe finished too fast: {elapsed:?}");
+    }
+}
+
+// ------------------------------------------------------------------
+// detect_client_type Fingerprint Check
+// ------------------------------------------------------------------
+
+#[test]
+fn test_detect_client_type_boundary_cases() {
+    // 9 bytes = port-scanner
+    assert_eq!(detect_client_type(&[0x42; 9]), "port-scanner");
+    // 10 bytes = unknown
+    assert_eq!(detect_client_type(&[0x42; 10]), "unknown");
+    
+    // HTTP verbs without trailing space
+    assert_eq!(detect_client_type(b"GET/"), "port-scanner"); // because len < 10
+    assert_eq!(detect_client_type(b"GET /path"), "HTTP"); 
+}
+
+// ------------------------------------------------------------------
+// Priority 2: Slowloris and Slow Read Attacks (OWASP ASVS 5.1.5)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn masking_slowloris_client_idle_timeout_rejected() {
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let backend_addr = listener.local_addr().unwrap();
+    let initial = b"GET / HTTP/1.1\r\nHost: front.example\r\n\r\n".to_vec();
+
+    let accept_task = tokio::spawn({
+        let initial = initial.clone();
+        async move {
+            let (mut stream, _) = listener.accept().await.unwrap();
+            let mut observed = vec![0u8; initial.len()];
+            stream.read_exact(&mut observed).await.unwrap();
+            assert_eq!(observed, initial);
+
+            let mut drip = [0u8; 1];
+            let drip_read = tokio::time::timeout(Duration::from_millis(220), stream.read_exact(&mut drip)).await;
+            assert!(
+                drip_read.is_err() || drip_read.unwrap().is_err(),
+                "backend must not receive post-timeout slowloris drip bytes"
+            );
+        }
+    });
+
+    let mut config = ProxyConfig::default();
+    config.censorship.mask = true;
+    config.censorship.mask_host = Some("127.0.0.1".to_string());
+    config.censorship.mask_port = backend_addr.port();
+
+    let beobachten = BeobachtenStore::new();
+    let peer: SocketAddr = "192.0.2.10:12345".parse().unwrap();
+    let local: SocketAddr = "192.0.2.1:443".parse().unwrap();
+
+    let (mut client_writer, client_reader) = duplex(1024);
+    let (_client_visible_reader, client_visible_writer) = duplex(1024);
+
+    let handle = tokio::spawn(async move {
+        handle_bad_client(
+            client_reader,
+            client_visible_writer,
+            &initial,
+            peer,
+            local,
+            &config,
+            &beobachten,
+        )
+        .await;
+    });
+
+    tokio::time::sleep(Duration::from_millis(160)).await;
+    let _ = client_writer.write_all(b"X").await;
+
+    handle.await.unwrap();
+    accept_task.await.unwrap();
+}
+
+// ------------------------------------------------------------------
+// Priority 2: Fallback Server Down / Fingerprinting (OWASP ASVS 5.1.7)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn masking_fallback_down_mimics_timeout() {
+    let mut config = ProxyConfig::default();
+    config.censorship.mask = true;
+    config.censorship.mask_host = Some("127.0.0.1".to_string());
+    config.censorship.mask_port = 1; // Unlikely port
+    
+    let (server_reader, server_writer) = duplex(1024);
+    let beobachten = BeobachtenStore::new();
+    let peer: SocketAddr = "192.0.2.12:12345".parse().unwrap();
+    let local: SocketAddr = "192.0.2.1:443".parse().unwrap();
+
+    let start = Instant::now();
+    handle_bad_client(server_reader, server_writer, b"GET / HTTP/1.1\r\n", peer, local, &config, &beobachten).await;
+    
+    let elapsed = start.elapsed();
+    // It should wait for MASK_TIMEOUT (50ms in tests) even if connection was refused immediately
+    assert!(elapsed >= Duration::from_millis(40), "Must respect connect budget even on failure: {:?}", elapsed);
+}
+
+// ------------------------------------------------------------------
+// Priority 2: SSRF Prevention (OWASP ASVS 5.1.2)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn masking_ssrf_resolve_internal_ranges_blocked() {
+    use crate::network::dns_overrides::resolve_socket_addr;
+
+    let blocked_ips = ["127.0.0.1", "169.254.169.254", "10.0.0.1", "192.168.1.1", "0.0.0.0"];
+
+    for ip in blocked_ips {
+        assert!(
+            resolve_socket_addr(ip, 80).is_none(),
+            "runtime DNS overrides must not resolve unconfigured literal host targets"
+        );
+    }
+}

src/proxy/middle_relay.rs

@@ -1,14 +1,15 @@
-use std::collections::HashMap;
-use std::collections::hash_map::DefaultHasher;
+use std::collections::hash_map::RandomState;
+use std::hash::BuildHasher;
 use std::hash::{Hash, Hasher};
 use std::net::{IpAddr, SocketAddr};
-use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
 use std::sync::{Arc, Mutex, OnceLock};
 use std::time::{Duration, Instant};
 
-use bytes::Bytes;
+use dashmap::DashMap;
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
-use tokio::sync::{mpsc, oneshot, watch};
+use tokio::sync::{mpsc, oneshot, watch, Mutex as AsyncMutex};
+use tokio::time::timeout;
 use tracing::{debug, trace, warn};
 
 use crate::config::ProxyConfig;
@@ -21,22 +22,37 @@ use crate::proxy::route_mode::{
     cutover_stagger_delay,
 };
 use crate::stats::Stats;
-use crate::stream::{BufferPool, CryptoReader, CryptoWriter};
+use crate::stream::{BufferPool, CryptoReader, CryptoWriter, PooledBuffer};
 use crate::transport::middle_proxy::{MePool, MeResponse, proto_flags_for_tag};
 
 enum C2MeCommand {
-    Data { payload: Bytes, flags: u32 },
+    Data { payload: PooledBuffer, flags: u32 },
     Close,
 }
 
 const DESYNC_DEDUP_WINDOW: Duration = Duration::from_secs(60);
+const DESYNC_DEDUP_MAX_ENTRIES: usize = 65_536;
+const DESYNC_DEDUP_PRUNE_SCAN_LIMIT: usize = 1024;
+const DESYNC_FULL_CACHE_EMIT_MIN_INTERVAL: Duration = Duration::from_millis(1000);
 const DESYNC_ERROR_CLASS: &str = "frame_too_large_crypto_desync";
 const C2ME_CHANNEL_CAPACITY_FALLBACK: usize = 128;
 const C2ME_SOFT_PRESSURE_MIN_FREE_SLOTS: usize = 64;
 const C2ME_SENDER_FAIRNESS_BUDGET: usize = 32;
+#[cfg(test)]
+const C2ME_SEND_TIMEOUT: Duration = Duration::from_millis(50);
+#[cfg(not(test))]
+const C2ME_SEND_TIMEOUT: Duration = Duration::from_secs(5);
 const ME_D2C_FLUSH_BATCH_MAX_FRAMES_MIN: usize = 1;
 const ME_D2C_FLUSH_BATCH_MAX_BYTES_MIN: usize = 4096;
-static DESYNC_DEDUP: OnceLock<Mutex<HashMap<u64, Instant>>> = OnceLock::new();
+#[cfg(test)]
+const QUOTA_USER_LOCKS_MAX: usize = 64;
+#[cfg(not(test))]
+const QUOTA_USER_LOCKS_MAX: usize = 4_096;
+static DESYNC_DEDUP: OnceLock<DashMap<u64, Instant>> = OnceLock::new();
+static DESYNC_HASHER: OnceLock<RandomState> = OnceLock::new();
+static DESYNC_FULL_CACHE_LAST_EMIT_AT: OnceLock<Mutex<Option<Instant>>> = OnceLock::new();
+static DESYNC_DEDUP_EVER_SATURATED: OnceLock<AtomicBool> = OnceLock::new();
+static QUOTA_USER_LOCKS: OnceLock<DashMap<String, Arc<AsyncMutex<()>>>> = OnceLock::new();
 
 struct RelayForensicsState {
     trace_id: u64,
@@ -76,7 +92,8 @@ impl MeD2cFlushPolicy {
 }
 
 fn hash_value<T: Hash>(value: &T) -> u64 {
-    let mut hasher = DefaultHasher::new();
+    let state = DESYNC_HASHER.get_or_init(RandomState::new);
+    let mut hasher = state.build_hasher();
     value.hash(&mut hasher);
     hasher.finish()
 }
@@ -90,26 +107,122 @@ fn should_emit_full_desync(key: u64, all_full: bool, now: Instant) -> bool {
         return true;
     }
 
-    let dedup = DESYNC_DEDUP.get_or_init(|| Mutex::new(HashMap::new()));
-    let mut guard = dedup.lock().expect("desync dedup mutex poisoned");
-    guard.retain(|_, seen_at| now.duration_since(*seen_at) < DESYNC_DEDUP_WINDOW);
+    let dedup = DESYNC_DEDUP.get_or_init(DashMap::new);
+    let saturated_before = dedup.len() >= DESYNC_DEDUP_MAX_ENTRIES;
+    let ever_saturated = DESYNC_DEDUP_EVER_SATURATED.get_or_init(|| AtomicBool::new(false));
+    if saturated_before {
+        ever_saturated.store(true, Ordering::Relaxed);
+    }
 
-    match guard.get_mut(&key) {
-        Some(seen_at) => {
-            if now.duration_since(*seen_at) >= DESYNC_DEDUP_WINDOW {
-                *seen_at = now;
+    if let Some(mut seen_at) = dedup.get_mut(&key) {
+        if now.duration_since(*seen_at) >= DESYNC_DEDUP_WINDOW {
+            *seen_at = now;
+            return true;
+        }
+        return false;
+    }
+
+    if dedup.len() >= DESYNC_DEDUP_MAX_ENTRIES {
+        let mut stale_keys = Vec::new();
+        let mut oldest_candidate: Option<(u64, Instant)> = None;
+        for entry in dedup.iter().take(DESYNC_DEDUP_PRUNE_SCAN_LIMIT) {
+            let key = *entry.key();
+            let seen_at = *entry.value();
+
+            match oldest_candidate {
+                Some((_, oldest_seen)) if seen_at >= oldest_seen => {}
+                _ => oldest_candidate = Some((key, seen_at)),
+            }
+
+            if now.duration_since(seen_at) >= DESYNC_DEDUP_WINDOW {
+                stale_keys.push(*entry.key());
+            }
+        }
+        for stale_key in stale_keys {
+            dedup.remove(&stale_key);
+        }
+        if dedup.len() >= DESYNC_DEDUP_MAX_ENTRIES {
+            let Some((evict_key, _)) = oldest_candidate else {
+                return false;
+            };
+            dedup.remove(&evict_key);
+            dedup.insert(key, now);
+            return should_emit_full_desync_full_cache(now);
+        }
+    }
+
+    dedup.insert(key, now);
+    let saturated_after = dedup.len() >= DESYNC_DEDUP_MAX_ENTRIES;
+    // Preserve the first sequential insert that reaches capacity as a normal
+    // emit, while still gating concurrent newcomer churn after the cache has
+    // ever been observed at saturation.
+    let was_ever_saturated = if saturated_after {
+        ever_saturated.swap(true, Ordering::Relaxed)
+    } else {
+        ever_saturated.load(Ordering::Relaxed)
+    };
+
+    if saturated_before || (saturated_after && was_ever_saturated) {
+        should_emit_full_desync_full_cache(now)
+    } else {
+        true
+    }
+}
+
+fn should_emit_full_desync_full_cache(now: Instant) -> bool {
+    let gate = DESYNC_FULL_CACHE_LAST_EMIT_AT.get_or_init(|| Mutex::new(None));
+    let Ok(mut last_emit_at) = gate.lock() else {
+        return false;
+    };
+
+    match *last_emit_at {
+        None => {
+            *last_emit_at = Some(now);
+            true
+        }
+        Some(last) => {
+            let Some(elapsed) = now.checked_duration_since(last) else {
+                *last_emit_at = Some(now);
+                return true;
+            };
+            if elapsed >= DESYNC_FULL_CACHE_EMIT_MIN_INTERVAL {
+                *last_emit_at = Some(now);
                 true
             } else {
                 false
             }
         }
-        None => {
-            guard.insert(key, now);
-            true
+    }
+}
+
+#[cfg(test)]
+fn clear_desync_dedup_for_testing() {
+    if let Some(dedup) = DESYNC_DEDUP.get() {
+        dedup.clear();
+    }
+    if let Some(ever_saturated) = DESYNC_DEDUP_EVER_SATURATED.get() {
+        ever_saturated.store(false, Ordering::Relaxed);
+    }
+    if let Some(last_emit_at) = DESYNC_FULL_CACHE_LAST_EMIT_AT.get() {
+        match last_emit_at.lock() {
+            Ok(mut guard) => {
+                *guard = None;
+            }
+            Err(poisoned) => {
+                let mut guard = poisoned.into_inner();
+                *guard = None;
+                last_emit_at.clear_poison();
+            }
         }
     }
 }
 
+#[cfg(test)]
+fn desync_dedup_test_lock() -> &'static Mutex<()> {
+    static TEST_LOCK: OnceLock<Mutex<()>> = OnceLock::new();
+    TEST_LOCK.get_or_init(|| Mutex::new(()))
+}
+
 fn report_desync_frame_too_large(
     state: &RelayForensicsState,
     proto_tag: ProtoTag,
@@ -205,6 +318,46 @@ fn should_yield_c2me_sender(sent_since_yield: usize, has_backlog: bool) -> bool
     has_backlog && sent_since_yield >= C2ME_SENDER_FAIRNESS_BUDGET
 }
 
+fn quota_exceeded_for_user(stats: &Stats, user: &str, quota_limit: Option<u64>) -> bool {
+    quota_limit.is_some_and(|quota| stats.get_user_total_octets(user) >= quota)
+}
+
+fn quota_would_be_exceeded_for_user(
+    stats: &Stats,
+    user: &str,
+    quota_limit: Option<u64>,
+    bytes: u64,
+) -> bool {
+    quota_limit.is_some_and(|quota| {
+        let used = stats.get_user_total_octets(user);
+        used >= quota || bytes > quota.saturating_sub(used)
+    })
+}
+
+fn quota_user_lock(user: &str) -> Arc<AsyncMutex<()>> {
+    let locks = QUOTA_USER_LOCKS.get_or_init(DashMap::new);
+    if let Some(existing) = locks.get(user) {
+        return Arc::clone(existing.value());
+    }
+
+    if locks.len() >= QUOTA_USER_LOCKS_MAX {
+        locks.retain(|_, value| Arc::strong_count(value) > 1);
+    }
+
+    if locks.len() >= QUOTA_USER_LOCKS_MAX {
+        return Arc::new(AsyncMutex::new(()));
+    }
+
+    let created = Arc::new(AsyncMutex::new(()));
+    match locks.entry(user.to_string()) {
+        dashmap::mapref::entry::Entry::Occupied(entry) => Arc::clone(entry.get()),
+        dashmap::mapref::entry::Entry::Vacant(entry) => {
+            entry.insert(Arc::clone(&created));
+            created
+        }
+    }
+}
+
 async fn enqueue_c2me_command(
     tx: &mpsc::Sender<C2MeCommand>,
     cmd: C2MeCommand,
@@ -217,7 +370,14 @@ async fn enqueue_c2me_command(
             if tx.capacity() <= C2ME_SOFT_PRESSURE_MIN_FREE_SLOTS {
                 tokio::task::yield_now().await;
             }
-            tx.send(cmd).await
+            match timeout(C2ME_SEND_TIMEOUT, tx.reserve()).await {
+                Ok(Ok(permit)) => {
+                    permit.send(cmd);
+                    Ok(())
+                }
+                Ok(Err(_)) => Err(mpsc::error::SendError(cmd)),
+                Err(_) => Err(mpsc::error::SendError(cmd)),
+            }
         }
     }
 }
@@ -229,7 +389,7 @@ pub(crate) async fn handle_via_middle_proxy<R, W>(
     me_pool: Arc<MePool>,
     stats: Arc<Stats>,
     config: Arc<ProxyConfig>,
-    _buffer_pool: Arc<BufferPool>,
+    buffer_pool: Arc<BufferPool>,
     local_addr: SocketAddr,
     rng: Arc<SecureRandom>,
     mut route_rx: watch::Receiver<RouteCutoverState>,
@@ -241,6 +401,7 @@ where
     W: AsyncWrite + Unpin + Send + 'static,
 {
     let user = success.user.clone();
+    let quota_limit = config.access.user_data_quota.get(&user).copied();
     let peer = success.peer;
     let proto_tag = success.proto_tag;
     let pool_generation = me_pool.current_generation();
@@ -256,7 +417,7 @@ where
     );
 
     let (conn_id, me_rx) = me_pool.registry().register().await;
-    let trace_id = conn_id;
+    let trace_id = session_id;
     let bytes_me2c = Arc::new(AtomicU64::new(0));
     let mut forensics = RelayForensicsState {
         trace_id,
@@ -271,8 +432,7 @@ where
     };
 
     stats.increment_user_connects(&user);
-    stats.increment_user_curr_connects(&user);
-    stats.increment_current_connections_me();
+    let _me_connection_lease = stats.acquire_me_connection_lease();
 
     if let Some(cutover) = affected_cutover_state(
         &route_rx,
@@ -290,8 +450,6 @@ where
         tokio::time::sleep(delay).await;
         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(ROUTE_SWITCH_ERROR_MSG.to_string()));
     }
 
@@ -392,6 +550,7 @@ where
                         &mut frame_buf,
                         stats_clone.as_ref(),
                         &user_clone,
+                        quota_limit,
                         bytes_me2c_clone.as_ref(),
                         conn_id,
                         d2c_flush_policy.ack_flush_immediate,
@@ -424,6 +583,7 @@ where
                             &mut frame_buf,
                             stats_clone.as_ref(),
                             &user_clone,
+                            quota_limit,
                             bytes_me2c_clone.as_ref(),
                             conn_id,
                             d2c_flush_policy.ack_flush_immediate,
@@ -456,6 +616,7 @@ where
                                     &mut frame_buf,
                                     stats_clone.as_ref(),
                                     &user_clone,
+                                    quota_limit,
                                     bytes_me2c_clone.as_ref(),
                                     conn_id,
                                     d2c_flush_policy.ack_flush_immediate,
@@ -488,6 +649,7 @@ where
                                         &mut frame_buf,
                                         stats_clone.as_ref(),
                                         &user_clone,
+                                        quota_limit,
                                         bytes_me2c_clone.as_ref(),
                                         conn_id,
                                         d2c_flush_policy.ack_flush_immediate,
@@ -557,6 +719,8 @@ where
                 &mut crypto_reader,
                 proto_tag,
                 frame_limit,
+                Duration::from_secs(config.timeouts.client_handshake.max(1)),
+                &buffer_pool,
                 &forensics,
                 &mut frame_counter,
                 &stats,
@@ -567,7 +731,19 @@ where
                         forensics.bytes_c2me = forensics
                             .bytes_c2me
                             .saturating_add(payload.len() as u64);
-                        stats.add_user_octets_from(&user, payload.len() as u64);
+                        if let Some(limit) = quota_limit {
+                            let quota_lock = quota_user_lock(&user);
+                            let _quota_guard = quota_lock.lock().await;
+                            stats.add_user_octets_from(&user, payload.len() as u64);
+                            if quota_exceeded_for_user(stats.as_ref(), &user, Some(limit)) {
+                                main_result = Err(ProxyError::DataQuotaExceeded {
+                                    user: user.clone(),
+                                });
+                                break;
+                            }
+                        } else {
+                            stats.add_user_octets_from(&user, payload.len() as u64);
+                        }
                         let mut flags = proto_flags;
                         if quickack {
                             flags |= RPC_FLAG_QUICKACK;
@@ -637,8 +813,6 @@ where
         "ME relay cleanup"
     );
     me_pool.registry().unregister(conn_id).await;
-    stats.decrement_current_connections_me();
-    stats.decrement_user_curr_connects(&user);
     result
 }
 
@@ -646,30 +820,49 @@ async fn read_client_payload<R>(
     client_reader: &mut CryptoReader<R>,
     proto_tag: ProtoTag,
     max_frame: usize,
+    frame_read_timeout: Duration,
+    buffer_pool: &Arc<BufferPool>,
     forensics: &RelayForensicsState,
     frame_counter: &mut u64,
     stats: &Stats,
-) -> Result<Option<(Bytes, bool)>>
+) -> Result<Option<(PooledBuffer, bool)>>
 where
     R: AsyncRead + Unpin + Send + 'static,
 {
+    async fn read_exact_with_timeout<R>(
+        client_reader: &mut CryptoReader<R>,
+        buf: &mut [u8],
+        frame_read_timeout: Duration,
+    ) -> Result<()>
+    where
+        R: AsyncRead + Unpin + Send + 'static,
+    {
+        match timeout(frame_read_timeout, client_reader.read_exact(buf)).await {
+            Ok(Ok(_)) => Ok(()),
+            Ok(Err(e)) => Err(ProxyError::Io(e)),
+            Err(_) => Err(ProxyError::Io(std::io::Error::new(
+                std::io::ErrorKind::TimedOut,
+                "middle-relay client frame read timeout",
+            ))),
+        }
+    }
+
     loop {
         let (len, quickack, raw_len_bytes) = match proto_tag {
             ProtoTag::Abridged => {
                 let mut first = [0u8; 1];
-                match client_reader.read_exact(&mut first).await {
-                    Ok(_) => {}
-                    Err(e) if e.kind() == std::io::ErrorKind::UnexpectedEof => return Ok(None),
-                    Err(e) => return Err(ProxyError::Io(e)),
+                match read_exact_with_timeout(client_reader, &mut first, frame_read_timeout).await {
+                    Ok(()) => {}
+                    Err(ProxyError::Io(e)) if e.kind() == std::io::ErrorKind::UnexpectedEof => {
+                        return Ok(None);
+                    }
+                    Err(e) => return Err(e),
                 }
 
                 let quickack = (first[0] & 0x80) != 0;
                 let len_words = if (first[0] & 0x7f) == 0x7f {
                     let mut ext = [0u8; 3];
-                    client_reader
-                        .read_exact(&mut ext)
-                        .await
-                        .map_err(ProxyError::Io)?;
+                    read_exact_with_timeout(client_reader, &mut ext, frame_read_timeout).await?;
                     u32::from_le_bytes([ext[0], ext[1], ext[2], 0]) as usize
                 } else {
                     (first[0] & 0x7f) as usize
@@ -682,10 +875,12 @@ where
             }
             ProtoTag::Intermediate | ProtoTag::Secure => {
                 let mut len_buf = [0u8; 4];
-                match client_reader.read_exact(&mut len_buf).await {
-                    Ok(_) => {}
-                    Err(e) if e.kind() == std::io::ErrorKind::UnexpectedEof => return Ok(None),
-                    Err(e) => return Err(ProxyError::Io(e)),
+                match read_exact_with_timeout(client_reader, &mut len_buf, frame_read_timeout).await {
+                    Ok(()) => {}
+                    Err(ProxyError::Io(e)) if e.kind() == std::io::ErrorKind::UnexpectedEof => {
+                        return Ok(None);
+                    }
+                    Err(e) => return Err(e),
                 }
                 let quickack = (len_buf[3] & 0x80) != 0;
                 (
@@ -737,18 +932,21 @@ where
             len
         };
 
-        let mut payload = vec![0u8; len];
-        client_reader
-            .read_exact(&mut payload)
-            .await
-            .map_err(ProxyError::Io)?;
+        let mut payload = buffer_pool.get();
+        payload.clear();
+        let current_cap = payload.capacity();
+        if current_cap < len {
+            payload.reserve(len - current_cap);
+        }
+        payload.resize(len, 0);
+        read_exact_with_timeout(client_reader, &mut payload[..len], frame_read_timeout).await?;
 
         // Secure Intermediate: strip validated trailing padding bytes.
         if proto_tag == ProtoTag::Secure {
             payload.truncate(secure_payload_len);
         }
         *frame_counter += 1;
-        return Ok(Some((Bytes::from(payload), quickack)));
+        return Ok(Some((payload, quickack)));
     }
 }
 
@@ -769,6 +967,7 @@ async fn process_me_writer_response<W>(
     frame_buf: &mut Vec<u8>,
     stats: &Stats,
     user: &str,
+    quota_limit: Option<u64>,
     bytes_me2c: &AtomicU64,
     conn_id: u64,
     ack_flush_immediate: bool,
@@ -784,17 +983,47 @@ where
             } else {
                 trace!(conn_id, bytes = data.len(), flags, "ME->C data");
             }
-            bytes_me2c.fetch_add(data.len() as u64, Ordering::Relaxed);
-            stats.add_user_octets_to(user, data.len() as u64);
-            write_client_payload(
-                client_writer,
-                proto_tag,
-                flags,
-                &data,
-                rng,
-                frame_buf,
-            )
-            .await?;
+            let data_len = data.len() as u64;
+            if let Some(limit) = quota_limit {
+                let quota_lock = quota_user_lock(user);
+                let _quota_guard = quota_lock.lock().await;
+                if quota_would_be_exceeded_for_user(stats, user, Some(limit), data_len) {
+                    return Err(ProxyError::DataQuotaExceeded {
+                        user: user.to_string(),
+                    });
+                }
+                write_client_payload(
+                    client_writer,
+                    proto_tag,
+                    flags,
+                    &data,
+                    rng,
+                    frame_buf,
+                )
+                .await?;
+
+                bytes_me2c.fetch_add(data.len() as u64, Ordering::Relaxed);
+                stats.add_user_octets_to(user, data.len() as u64);
+
+                if quota_exceeded_for_user(stats, user, Some(limit)) {
+                    return Err(ProxyError::DataQuotaExceeded {
+                        user: user.to_string(),
+                    });
+                }
+            } else {
+                write_client_payload(
+                    client_writer,
+                    proto_tag,
+                    flags,
+                    &data,
+                    rng,
+                    frame_buf,
+                )
+                .await?;
+
+                bytes_me2c.fetch_add(data.len() as u64, Ordering::Relaxed);
+                stats.add_user_octets_to(user, data.len() as u64);
+            }
 
             Ok(MeWriterResponseOutcome::Continue {
                 frames: 1,
@@ -940,82 +1169,5 @@ where
 }
 
 #[cfg(test)]
-mod tests {
-    use super::*;
-    use tokio::time::{Duration as TokioDuration, timeout};
-
-    #[test]
-    fn should_yield_sender_only_on_budget_with_backlog() {
-        assert!(!should_yield_c2me_sender(0, true));
-        assert!(!should_yield_c2me_sender(C2ME_SENDER_FAIRNESS_BUDGET - 1, true));
-        assert!(!should_yield_c2me_sender(C2ME_SENDER_FAIRNESS_BUDGET, false));
-        assert!(should_yield_c2me_sender(C2ME_SENDER_FAIRNESS_BUDGET, true));
-    }
-
-    #[tokio::test]
-    async fn enqueue_c2me_command_uses_try_send_fast_path() {
-        let (tx, mut rx) = mpsc::channel::<C2MeCommand>(2);
-        enqueue_c2me_command(
-            &tx,
-            C2MeCommand::Data {
-                payload: Bytes::from_static(&[1, 2, 3]),
-                flags: 0,
-            },
-        )
-        .await
-        .unwrap();
-
-        let recv = timeout(TokioDuration::from_millis(50), rx.recv())
-            .await
-            .unwrap()
-            .unwrap();
-        match recv {
-            C2MeCommand::Data { payload, flags } => {
-                assert_eq!(payload.as_ref(), &[1, 2, 3]);
-                assert_eq!(flags, 0);
-            }
-            C2MeCommand::Close => panic!("unexpected close command"),
-        }
-    }
-
-    #[tokio::test]
-    async fn enqueue_c2me_command_falls_back_to_send_when_queue_is_full() {
-        let (tx, mut rx) = mpsc::channel::<C2MeCommand>(1);
-        tx.send(C2MeCommand::Data {
-            payload: Bytes::from_static(&[9]),
-            flags: 9,
-        })
-        .await
-        .unwrap();
-
-        let tx2 = tx.clone();
-        let producer = tokio::spawn(async move {
-            enqueue_c2me_command(
-                &tx2,
-                C2MeCommand::Data {
-                    payload: Bytes::from_static(&[7, 7]),
-                    flags: 7,
-                },
-            )
-            .await
-            .unwrap();
-        });
-
-        let _ = timeout(TokioDuration::from_millis(100), rx.recv())
-            .await
-            .unwrap();
-        producer.await.unwrap();
-
-        let recv = timeout(TokioDuration::from_millis(100), rx.recv())
-            .await
-            .unwrap()
-            .unwrap();
-        match recv {
-            C2MeCommand::Data { payload, flags } => {
-                assert_eq!(payload.as_ref(), &[7, 7]);
-                assert_eq!(flags, 7);
-            }
-            C2MeCommand::Close => panic!("unexpected close command"),
-        }
-    }
-}
+#[path = "middle_relay_security_tests.rs"]
+mod security_tests;

src/proxy/relay.rs

@@ -53,16 +53,17 @@
 
 use std::io;
 use std::pin::Pin;
-use std::sync::Arc;
-use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex, OnceLock};
+use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
 use std::task::{Context, Poll};
 use std::time::Duration;
+use dashmap::DashMap;
 use tokio::io::{
     AsyncRead, AsyncWrite, AsyncWriteExt, ReadBuf, copy_bidirectional_with_sizes,
 };
 use tokio::time::Instant;
 use tracing::{debug, trace, warn};
-use crate::error::Result;
+use crate::error::{ProxyError, Result};
 use crate::stats::Stats;
 use crate::stream::BufferPool;
 
@@ -205,6 +206,10 @@ struct StatsIo<S> {
     counters: Arc<SharedCounters>,
     stats: Arc<Stats>,
     user: String,
+    quota_limit: Option<u64>,
+    quota_exceeded: Arc<AtomicBool>,
+    quota_read_wake_scheduled: bool,
+    quota_write_wake_scheduled: bool,
     epoch: Instant,
 }
 
@@ -214,11 +219,64 @@ impl<S> StatsIo<S> {
         counters: Arc<SharedCounters>,
         stats: Arc<Stats>,
         user: String,
+        quota_limit: Option<u64>,
+        quota_exceeded: Arc<AtomicBool>,
         epoch: Instant,
     ) -> Self {
         // Mark initial activity so the watchdog doesn't fire before data flows
         counters.touch(Instant::now(), epoch);
-        Self { inner, counters, stats, user, epoch }
+        Self {
+            inner,
+            counters,
+            stats,
+            user,
+            quota_limit,
+            quota_exceeded,
+            quota_read_wake_scheduled: false,
+            quota_write_wake_scheduled: false,
+            epoch,
+        }
+    }
+}
+
+#[derive(Debug)]
+struct QuotaIoSentinel;
+
+impl std::fmt::Display for QuotaIoSentinel {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str("user data quota exceeded")
+    }
+}
+
+impl std::error::Error for QuotaIoSentinel {}
+
+fn quota_io_error() -> io::Error {
+    io::Error::new(io::ErrorKind::PermissionDenied, QuotaIoSentinel)
+}
+
+fn is_quota_io_error(err: &io::Error) -> bool {
+    err.kind() == io::ErrorKind::PermissionDenied
+        && err
+            .get_ref()
+            .and_then(|source| source.downcast_ref::<QuotaIoSentinel>())
+            .is_some()
+}
+
+static QUOTA_USER_LOCKS: OnceLock<DashMap<String, Arc<Mutex<()>>>> = OnceLock::new();
+
+fn quota_user_lock(user: &str) -> Arc<Mutex<()>> {
+    let locks = QUOTA_USER_LOCKS.get_or_init(DashMap::new);
+    if let Some(existing) = locks.get(user) {
+        return Arc::clone(existing.value());
+    }
+
+    let created = Arc::new(Mutex::new(()));
+    match locks.entry(user.to_string()) {
+        dashmap::mapref::entry::Entry::Occupied(entry) => Arc::clone(entry.get()),
+        dashmap::mapref::entry::Entry::Vacant(entry) => {
+            entry.insert(Arc::clone(&created));
+            created
+        }
     }
 }
 
@@ -229,6 +287,42 @@ impl<S: AsyncRead + Unpin> AsyncRead for StatsIo<S> {
         buf: &mut ReadBuf<'_>,
     ) -> Poll<io::Result<()>> {
         let this = self.get_mut();
+        if this.quota_exceeded.load(Ordering::Relaxed) {
+            return Poll::Ready(Err(quota_io_error()));
+        }
+
+        let quota_lock = this
+            .quota_limit
+            .is_some()
+            .then(|| quota_user_lock(&this.user));
+        let _quota_guard = if let Some(lock) = quota_lock.as_ref() {
+            match lock.try_lock() {
+                Ok(guard) => {
+                    this.quota_read_wake_scheduled = false;
+                    Some(guard)
+                }
+                Err(_) => {
+                    if !this.quota_read_wake_scheduled {
+                        this.quota_read_wake_scheduled = true;
+                        let waker = cx.waker().clone();
+                        tokio::task::spawn(async move {
+                            tokio::task::yield_now().await;
+                            waker.wake();
+                        });
+                    }
+                    return Poll::Pending;
+                }
+            }
+        } else {
+            None
+        };
+
+        if let Some(limit) = this.quota_limit
+            && this.stats.get_user_total_octets(&this.user) >= limit
+        {
+            this.quota_exceeded.store(true, Ordering::Relaxed);
+            return Poll::Ready(Err(quota_io_error()));
+        }
         let before = buf.filled().len();
 
         match Pin::new(&mut this.inner).poll_read(cx, buf) {
@@ -243,6 +337,13 @@ impl<S: AsyncRead + Unpin> AsyncRead for StatsIo<S> {
                     this.stats.add_user_octets_from(&this.user, n as u64);
                     this.stats.increment_user_msgs_from(&this.user);
 
+                    if let Some(limit) = this.quota_limit
+                        && this.stats.get_user_total_octets(&this.user) >= limit
+                    {
+                        this.quota_exceeded.store(true, Ordering::Relaxed);
+                        return Poll::Ready(Err(quota_io_error()));
+                    }
+
                     trace!(user = %this.user, bytes = n, "C->S");
                 }
                 Poll::Ready(Ok(()))
@@ -259,8 +360,56 @@ impl<S: AsyncWrite + Unpin> AsyncWrite for StatsIo<S> {
         buf: &[u8],
     ) -> Poll<io::Result<usize>> {
         let this = self.get_mut();
+        if this.quota_exceeded.load(Ordering::Relaxed) {
+            return Poll::Ready(Err(quota_io_error()));
+        }
 
-        match Pin::new(&mut this.inner).poll_write(cx, buf) {
+        let quota_lock = this
+            .quota_limit
+            .is_some()
+            .then(|| quota_user_lock(&this.user));
+        let _quota_guard = if let Some(lock) = quota_lock.as_ref() {
+            match lock.try_lock() {
+                Ok(guard) => {
+                    this.quota_write_wake_scheduled = false;
+                    Some(guard)
+                }
+                Err(_) => {
+                    if !this.quota_write_wake_scheduled {
+                        this.quota_write_wake_scheduled = true;
+                        let waker = cx.waker().clone();
+                        tokio::task::spawn(async move {
+                            tokio::task::yield_now().await;
+                            waker.wake();
+                        });
+                    }
+                    return Poll::Pending;
+                }
+            }
+        } else {
+            None
+        };
+
+        let write_buf = if let Some(limit) = this.quota_limit {
+            let used = this.stats.get_user_total_octets(&this.user);
+            if used >= limit {
+                this.quota_exceeded.store(true, Ordering::Relaxed);
+                return Poll::Ready(Err(quota_io_error()));
+            }
+
+            let remaining = (limit - used) as usize;
+            if buf.len() > remaining {
+                // Fail closed: do not emit partial S->C payload when remaining
+                // quota cannot accommodate the pending write request.
+                this.quota_exceeded.store(true, Ordering::Relaxed);
+                return Poll::Ready(Err(quota_io_error()));
+            }
+            buf
+        } else {
+            buf
+        };
+
+        match Pin::new(&mut this.inner).poll_write(cx, write_buf) {
             Poll::Ready(Ok(n)) => {
                 if n > 0 {
                     // S→C: data written to client
@@ -271,6 +420,13 @@ impl<S: AsyncWrite + Unpin> AsyncWrite for StatsIo<S> {
                     this.stats.add_user_octets_to(&this.user, n as u64);
                     this.stats.increment_user_msgs_to(&this.user);
 
+                    if let Some(limit) = this.quota_limit
+                        && this.stats.get_user_total_octets(&this.user) >= limit
+                    {
+                        this.quota_exceeded.store(true, Ordering::Relaxed);
+                        return Poll::Ready(Err(quota_io_error()));
+                    }
+
                     trace!(user = %this.user, bytes = n, "S->C");
                 }
                 Poll::Ready(Ok(n))
@@ -307,7 +463,8 @@ impl<S: AsyncWrite + Unpin> AsyncWrite for StatsIo<S> {
 /// - Per-user stats: bytes and ops counted per direction
 /// - Periodic rate logging: every 10 seconds when active
 /// - Clean shutdown: both write sides are shut down on exit
-/// - Error propagation: I/O errors are returned as `ProxyError::Io`
+/// - Error propagation: quota exits return `ProxyError::DataQuotaExceeded`,
+///   other I/O failures are returned as `ProxyError::Io`
 pub async fn relay_bidirectional<CR, CW, SR, SW>(
     client_reader: CR,
     client_writer: CW,
@@ -317,6 +474,7 @@ pub async fn relay_bidirectional<CR, CW, SR, SW>(
     s2c_buf_size: usize,
     user: &str,
     stats: Arc<Stats>,
+    quota_limit: Option<u64>,
     _buffer_pool: Arc<BufferPool>,
 ) -> Result<()>
 where
@@ -327,6 +485,7 @@ where
 {
     let epoch = Instant::now();
     let counters = Arc::new(SharedCounters::new());
+    let quota_exceeded = Arc::new(AtomicBool::new(false));
     let user_owned = user.to_string();
 
     // ── Combine split halves into bidirectional streams ──────────────
@@ -339,12 +498,15 @@ where
         Arc::clone(&counters),
         Arc::clone(&stats),
         user_owned.clone(),
+        quota_limit,
+        Arc::clone(&quota_exceeded),
         epoch,
     );
 
     // ── Watchdog: activity timeout + periodic rate logging ──────────
     let wd_counters = Arc::clone(&counters);
     let wd_user = user_owned.clone();
+    let wd_quota_exceeded = Arc::clone(&quota_exceeded);
 
     let watchdog = async {
         let mut prev_c2s: u64 = 0;
@@ -356,6 +518,11 @@ where
             let now = Instant::now();
             let idle = wd_counters.idle_duration(now, epoch);
 
+            if wd_quota_exceeded.load(Ordering::Relaxed) {
+                warn!(user = %wd_user, "User data quota reached, closing relay");
+                return;
+            }
+
             // ── Activity timeout ────────────────────────────────────
             if idle >= ACTIVITY_TIMEOUT {
                 let c2s = wd_counters.c2s_bytes.load(Ordering::Relaxed);
@@ -439,6 +606,22 @@ where
             );
             Ok(())
         }
+        Some(Err(e)) if is_quota_io_error(&e) => {
+            let c2s = counters.c2s_bytes.load(Ordering::Relaxed);
+            let s2c = counters.s2c_bytes.load(Ordering::Relaxed);
+            warn!(
+                user = %user_owned,
+                c2s_bytes = c2s,
+                s2c_bytes = s2c,
+                c2s_msgs = c2s_ops,
+                s2c_msgs = s2c_ops,
+                duration_secs = duration.as_secs(),
+                "Data quota reached, closing relay"
+            );
+            Err(ProxyError::DataQuotaExceeded {
+                user: user_owned.clone(),
+            })
+        }
         Some(Err(e)) => {
             // I/O error in one of the directions
             let c2s = counters.c2s_bytes.load(Ordering::Relaxed);
@@ -472,3 +655,9 @@ where
         }
     }
 }
+
+#[cfg(test)]
+#[path = "relay_security_tests.rs"]
+mod security_tests;
+#[path = "relay_adversarial_tests.rs"]
+mod adversarial_tests;

src/proxy/relay_adversarial_tests.rs

@@ -0,0 +1,122 @@
+use super::*;
+use crate::error::ProxyError;
+use crate::stats::Stats;
+use crate::stream::BufferPool;
+use std::sync::Arc;
+use tokio::io::{duplex, AsyncReadExt, AsyncWriteExt};
+use tokio::time::{Duration, Instant, timeout};
+
+// ------------------------------------------------------------------
+// Priority 3: Async Relay HOL Blocking Prevention (OWASP ASVS 5.1.5)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn relay_hol_blocking_prevention_regression() {
+    let stats = Arc::new(Stats::new());
+    let user = "hol-user";
+    
+    let (client_peer, relay_client) = duplex(65536);
+    let (relay_server, server_peer) = duplex(65536);
+
+    let (client_reader, client_writer) = tokio::io::split(relay_client);
+    let (server_reader, server_writer) = tokio::io::split(relay_server);
+    let (mut cp_reader, mut cp_writer) = tokio::io::split(client_peer);
+    let (mut sp_reader, mut sp_writer) = tokio::io::split(server_peer);
+
+    let relay_task = tokio::spawn(relay_bidirectional(
+        client_reader,
+        client_writer,
+        server_reader,
+        server_writer,
+        8192,
+        8192,
+        user,
+        Arc::clone(&stats),
+        None,
+        Arc::new(BufferPool::new()),
+    ));
+
+    let payload_size = 1024 * 10;
+    let s2c_payload = vec![0x41; payload_size];
+    let c2s_payload = vec![0x42; payload_size];
+
+    let s2c_handle = tokio::spawn(async move {
+        sp_writer.write_all(&s2c_payload).await.unwrap();
+        
+        let mut total_read = 0;
+        let mut buf = [0u8; 10];
+        while total_read < payload_size {
+            let n = cp_reader.read(&mut buf).await.unwrap();
+            total_read += n;
+            tokio::time::sleep(Duration::from_millis(100)).await;
+        }
+    });
+
+    let start = Instant::now();
+    cp_writer.write_all(&c2s_payload).await.unwrap();
+    
+    let mut server_buf = vec![0u8; payload_size];
+    sp_reader.read_exact(&mut server_buf).await.unwrap();
+    let elapsed = start.elapsed();
+
+    assert!(elapsed < Duration::from_millis(1000), "C->S must not be blocked by slow S->C (HOL blocking): {:?}", elapsed);
+    assert_eq!(server_buf, c2s_payload);
+
+    s2c_handle.abort();
+    relay_task.abort();
+}
+
+// ------------------------------------------------------------------
+// Priority 3: Data Quota Mid-Session Cutoff (OWASP ASVS 5.1.6)
+// ------------------------------------------------------------------
+
+#[tokio::test]
+async fn relay_quota_mid_session_cutoff() {
+    let stats = Arc::new(Stats::new());
+    let user = "quota-mid-user";
+    let quota = 5000;
+    
+    let (client_peer, relay_client) = duplex(8192);
+    let (relay_server, server_peer) = duplex(8192);
+
+    let (client_reader, client_writer) = tokio::io::split(relay_client);
+    let (server_reader, server_writer) = tokio::io::split(relay_server);
+    let (mut _cp_reader, mut cp_writer) = tokio::io::split(client_peer);
+    let (mut sp_reader, _sp_writer) = tokio::io::split(server_peer);
+
+    let relay_task = tokio::spawn(relay_bidirectional(
+        client_reader,
+        client_writer,
+        server_reader,
+        server_writer,
+        1024,
+        1024,
+        user,
+        Arc::clone(&stats),
+        Some(quota),
+        Arc::new(BufferPool::new()),
+    ));
+
+    // Send 4000 bytes (Ok)
+    let buf1 = vec![0x42; 4000];
+    cp_writer.write_all(&buf1).await.unwrap();
+    let mut server_recv = vec![0u8; 4000];
+    sp_reader.read_exact(&mut server_recv).await.unwrap();
+
+    // Send another 2000 bytes (Total 6000 > 5000)
+    let buf2 = vec![0x42; 2000];
+    let _ = cp_writer.write_all(&buf2).await;
+    
+    let relay_res = timeout(Duration::from_secs(1), relay_task).await.unwrap();
+    
+    match relay_res {
+        Ok(Err(ProxyError::DataQuotaExceeded { .. })) => {
+            // Expected
+        }
+        other => panic!("Expected DataQuotaExceeded error, got: {:?}", other),
+    }
+
+    let mut small_buf = [0u8; 1];
+    let n = sp_reader.read(&mut small_buf).await.unwrap();
+    assert_eq!(n, 0, "Server must see EOF after quota reached");
+}

src/proxy/route_mode.rs

@@ -1,10 +1,10 @@
 use std::sync::Arc;
-use std::sync::atomic::{AtomicU8, AtomicU64, Ordering};
+use std::sync::atomic::{AtomicU64, Ordering};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
 
 use tokio::sync::watch;
 
-pub(crate) const ROUTE_SWITCH_ERROR_MSG: &str = "Route mode switched by cutover";
+pub(crate) const ROUTE_SWITCH_ERROR_MSG: &str = "Session terminated";
 
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 #[repr(u8)]
@@ -14,17 +14,6 @@ pub(crate) enum RelayRouteMode {
 }
 
 impl RelayRouteMode {
-    pub(crate) fn as_u8(self) -> u8 {
-        self as u8
-    }
-
-    pub(crate) fn from_u8(value: u8) -> Self {
-        match value {
-            1 => Self::Middle,
-            _ => Self::Direct,
-        }
-    }
-
     pub(crate) fn as_str(self) -> &'static str {
         match self {
             Self::Direct => "direct",
@@ -41,8 +30,6 @@ pub(crate) struct RouteCutoverState {
 
 #[derive(Clone)]
 pub(crate) struct RouteRuntimeController {
-    mode: Arc<AtomicU8>,
-    generation: Arc<AtomicU64>,
     direct_since_epoch_secs: Arc<AtomicU64>,
     tx: watch::Sender<RouteCutoverState>,
 }
@@ -60,18 +47,13 @@ impl RouteRuntimeController {
             0
         };
         Self {
-            mode: Arc::new(AtomicU8::new(initial_mode.as_u8())),
-            generation: Arc::new(AtomicU64::new(0)),
             direct_since_epoch_secs: Arc::new(AtomicU64::new(direct_since_epoch_secs)),
             tx,
         }
     }
 
     pub(crate) fn snapshot(&self) -> RouteCutoverState {
-        RouteCutoverState {
-            mode: RelayRouteMode::from_u8(self.mode.load(Ordering::Relaxed)),
-            generation: self.generation.load(Ordering::Relaxed),
-        }
+        *self.tx.borrow()
     }
 
     pub(crate) fn subscribe(&self) -> watch::Receiver<RouteCutoverState> {
@@ -84,20 +66,29 @@ impl RouteRuntimeController {
     }
 
     pub(crate) fn set_mode(&self, mode: RelayRouteMode) -> Option<RouteCutoverState> {
-        let previous = self.mode.swap(mode.as_u8(), Ordering::Relaxed);
-        if previous == mode.as_u8() {
+        let mut next = None;
+        let changed = self.tx.send_if_modified(|state| {
+            if state.mode == mode {
+                return false;
+            }
+            state.mode = mode;
+            state.generation = state.generation.saturating_add(1);
+            next = Some(*state);
+            true
+        });
+
+        if !changed {
             return None;
         }
+
         if matches!(mode, RelayRouteMode::Direct) {
             self.direct_since_epoch_secs
                 .store(now_epoch_secs(), Ordering::Relaxed);
         } else {
             self.direct_since_epoch_secs.store(0, Ordering::Relaxed);
         }
-        let generation = self.generation.fetch_add(1, Ordering::Relaxed) + 1;
-        let next = RouteCutoverState { mode, generation };
-        self.tx.send_replace(next);
-        Some(next)
+
+        next
     }
 }
 
@@ -110,10 +101,10 @@ fn now_epoch_secs() -> u64 {
 
 pub(crate) fn is_session_affected_by_cutover(
     current: RouteCutoverState,
-    _session_mode: RelayRouteMode,
+    session_mode: RelayRouteMode,
     session_generation: u64,
 ) -> bool {
-    current.generation > session_generation
+    current.generation > session_generation && current.mode != session_mode
 }
 
 pub(crate) fn affected_cutover_state(
@@ -140,3 +131,7 @@ pub(crate) fn cutover_stagger_delay(session_id: u64, generation: u64) -> Duratio
     let ms = 1000 + (value % 1000);
     Duration::from_millis(ms)
 }
+
+#[cfg(test)]
+#[path = "route_mode_security_tests.rs"]
+mod security_tests;

src/proxy/route_mode_security_tests.rs

@@ -0,0 +1,406 @@
+use super::*;
+use rand::{Rng, SeedableRng};
+use rand::rngs::StdRng;
+use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
+
+#[test]
+fn cutover_stagger_delay_is_deterministic_for_same_inputs() {
+    let d1 = cutover_stagger_delay(0x0123_4567_89ab_cdef, 42);
+    let d2 = cutover_stagger_delay(0x0123_4567_89ab_cdef, 42);
+    assert_eq!(
+        d1, d2,
+        "stagger delay must be deterministic for identical session/generation inputs"
+    );
+}
+
+#[test]
+fn cutover_stagger_delay_stays_within_budget_bounds() {
+    // Black-hat model: censors trigger many cutovers and correlate disconnect timing.
+    // Keep delay inside a narrow coarse window to avoid long-tail spikes.
+    for generation in [0u64, 1, 2, 3, 16, 128, u32::MAX as u64, u64::MAX] {
+        for session_id in [
+            0u64,
+            1,
+            2,
+            0xdead_beef,
+            0xfeed_face_cafe_babe,
+            u64::MAX,
+        ] {
+            let delay = cutover_stagger_delay(session_id, generation);
+            assert!(
+                (1000..=1999).contains(&delay.as_millis()),
+                "stagger delay must remain in fixed 1000..=1999ms budget"
+            );
+        }
+    }
+}
+
+#[test]
+fn cutover_stagger_delay_changes_with_generation_for_same_session() {
+    let session_id = 0x0123_4567_89ab_cdef;
+    let first = cutover_stagger_delay(session_id, 100);
+    let second = cutover_stagger_delay(session_id, 101);
+    assert_ne!(
+        first, second,
+        "adjacent cutover generations should decorrelate disconnect delays"
+    );
+}
+
+#[test]
+fn route_runtime_set_mode_is_idempotent_for_same_mode() {
+    let runtime = RouteRuntimeController::new(RelayRouteMode::Direct);
+    let first = runtime.snapshot();
+    let changed = runtime.set_mode(RelayRouteMode::Direct);
+    let second = runtime.snapshot();
+
+    assert!(
+        changed.is_none(),
+        "setting already-active mode must not produce a cutover event"
+    );
+    assert_eq!(
+        first.generation, second.generation,
+        "idempotent mode set must not bump generation"
+    );
+}
+
+#[test]
+fn affected_cutover_state_triggers_only_for_newer_generation() {
+    let runtime = RouteRuntimeController::new(RelayRouteMode::Direct);
+    let rx = runtime.subscribe();
+    let initial = runtime.snapshot();
+
+    assert!(
+        affected_cutover_state(&rx, RelayRouteMode::Direct, initial.generation).is_none(),
+        "current generation must not be considered a cutover for existing session"
+    );
+
+    let next = runtime
+        .set_mode(RelayRouteMode::Middle)
+        .expect("mode change must produce cutover state");
+    let seen = affected_cutover_state(&rx, RelayRouteMode::Direct, initial.generation)
+        .expect("newer generation must be observed as cutover");
+
+    assert_eq!(seen.generation, next.generation);
+    assert_eq!(seen.mode, RelayRouteMode::Middle);
+}
+
+#[test]
+fn integration_watch_and_snapshot_follow_same_transition_sequence() {
+    let runtime = RouteRuntimeController::new(RelayRouteMode::Direct);
+    let rx = runtime.subscribe();
+
+    let sequence = [
+        RelayRouteMode::Middle,
+        RelayRouteMode::Middle,
+        RelayRouteMode::Direct,
+        RelayRouteMode::Direct,
+        RelayRouteMode::Middle,
+    ];
+
+    let mut expected_generation = 0u64;
+    let mut expected_mode = RelayRouteMode::Direct;
+
+    for target in sequence {
+        let changed = runtime.set_mode(target);
+        if target == expected_mode {
+            assert!(changed.is_none(), "idempotent transition must return none");
+        } else {
+            expected_mode = target;
+            expected_generation = expected_generation.saturating_add(1);
+            let emitted = changed.expect("real transition must emit cutover state");
+            assert_eq!(emitted.mode, expected_mode);
+            assert_eq!(emitted.generation, expected_generation);
+        }
+
+        let snap = runtime.snapshot();
+        let watched = *rx.borrow();
+        assert_eq!(snap, watched, "snapshot and watch state must stay aligned");
+        assert_eq!(snap.mode, expected_mode);
+        assert_eq!(snap.generation, expected_generation);
+    }
+}
+
+#[test]
+fn session_is_not_affected_when_mode_matches_even_if_generation_advanced() {
+    let session_mode = RelayRouteMode::Direct;
+    let current = RouteCutoverState {
+        mode: RelayRouteMode::Direct,
+        generation: 2,
+    };
+    let session_generation = 0;
+
+    assert!(
+        !is_session_affected_by_cutover(current, session_mode, session_generation),
+        "session on matching final route mode should not be force-cut over on intermediate generation bumps"
+    );
+}
+
+#[test]
+fn cutover_predicate_rejects_equal_generation_even_if_mode_differs() {
+    let current = RouteCutoverState {
+        mode: RelayRouteMode::Middle,
+        generation: 77,
+    };
+    assert!(
+        !is_session_affected_by_cutover(current, RelayRouteMode::Direct, 77),
+        "equal generation must never trigger cutover regardless of mode mismatch"
+    );
+}
+
+#[test]
+fn adversarial_route_oscillation_only_cuts_over_sessions_with_different_final_mode() {
+    let runtime = RouteRuntimeController::new(RelayRouteMode::Direct);
+    let rx = runtime.subscribe();
+    let session_generation = runtime.snapshot().generation;
+
+    runtime
+        .set_mode(RelayRouteMode::Middle)
+        .expect("direct->middle must transition");
+    runtime
+        .set_mode(RelayRouteMode::Direct)
+        .expect("middle->direct must transition");
+
+    assert!(
+        affected_cutover_state(&rx, RelayRouteMode::Direct, session_generation).is_none(),
+        "direct session should survive when final mode returns to direct"
+    );
+    assert!(
+        affected_cutover_state(&rx, RelayRouteMode::Middle, session_generation).is_some(),
+        "middle session should be cut over when final mode is direct"
+    );
+}
+
+#[test]
+fn light_fuzz_cutover_predicate_matches_reference_oracle() {
+    let mut rng = StdRng::seed_from_u64(0xC0DEC0DE5EED);
+    for _ in 0..20_000 {
+        let current = RouteCutoverState {
+            mode: if rng.random::<bool>() {
+                RelayRouteMode::Direct
+            } else {
+                RelayRouteMode::Middle
+            },
+            generation: rng.random_range(0u64..1_000_000),
+        };
+        let session_mode = if rng.random::<bool>() {
+            RelayRouteMode::Direct
+        } else {
+            RelayRouteMode::Middle
+        };
+        let session_generation = rng.random_range(0u64..1_000_000);
+
+        let expected = current.generation > session_generation && current.mode != session_mode;
+        let actual = is_session_affected_by_cutover(current, session_mode, session_generation);
+        assert_eq!(
+            actual, expected,
+            "cutover predicate must match mode-aware generation oracle"
+        );
+    }
+}
+
+#[test]
+fn light_fuzz_set_mode_generation_tracks_only_real_transitions() {
+    let runtime = RouteRuntimeController::new(RelayRouteMode::Direct);
+    let mut rng = StdRng::seed_from_u64(0x0DDC0FFE);
+
+    let mut expected_mode = RelayRouteMode::Direct;
+    let mut expected_generation = 0u64;
+
+    for _ in 0..10_000 {
+        let candidate = if rng.random::<bool>() {
+            RelayRouteMode::Direct
+        } else {
+            RelayRouteMode::Middle
+        };
+        let changed = runtime.set_mode(candidate);
+
+        if candidate == expected_mode {
+            assert!(changed.is_none(), "idempotent set_mode must not emit cutover state");
+        } else {
+            expected_mode = candidate;
+            expected_generation = expected_generation.saturating_add(1);
+            let next = changed.expect("mode transition must emit cutover state");
+            assert_eq!(next.mode, expected_mode);
+            assert_eq!(next.generation, expected_generation);
+        }
+    }
+
+    let final_state = runtime.snapshot();
+    assert_eq!(final_state.mode, expected_mode);
+    assert_eq!(final_state.generation, expected_generation);
+}
+
+#[test]
+fn stress_snapshot_and_watch_state_remain_consistent_under_concurrent_switch_storm() {
+    let runtime = Arc::new(RouteRuntimeController::new(RelayRouteMode::Direct));
+
+    std::thread::scope(|scope| {
+        let mut writers = Vec::new();
+        for worker in 0..4usize {
+            let runtime = Arc::clone(&runtime);
+            writers.push(scope.spawn(move || {
+                for step in 0..20_000usize {
+                    let mode = if (worker + step) % 2 == 0 {
+                        RelayRouteMode::Direct
+                    } else {
+                        RelayRouteMode::Middle
+                    };
+                    let _ = runtime.set_mode(mode);
+                }
+            }));
+        }
+
+        for writer in writers {
+            writer
+                .join()
+                .expect("route mode writer thread must not panic");
+        }
+
+        let rx = runtime.subscribe();
+        for _ in 0..128 {
+            assert_eq!(
+                runtime.snapshot(),
+                *rx.borrow(),
+                "snapshot and watch state must converge after concurrent set_mode churn"
+            );
+            std::thread::yield_now();
+        }
+    });
+}
+
+#[test]
+fn stress_concurrent_transition_count_matches_final_generation() {
+    let runtime = Arc::new(RouteRuntimeController::new(RelayRouteMode::Direct));
+    let successful_transitions = Arc::new(AtomicU64::new(0));
+
+    std::thread::scope(|scope| {
+        let mut workers = Vec::new();
+        for worker in 0..6usize {
+            let runtime = Arc::clone(&runtime);
+            let successful_transitions = Arc::clone(&successful_transitions);
+            workers.push(scope.spawn(move || {
+                let mut state = (worker as u64 + 1).wrapping_mul(0x9E37_79B9_7F4A_7C15);
+                for _ in 0..25_000usize {
+                    state ^= state << 7;
+                    state ^= state >> 9;
+                    state ^= state << 8;
+                    let mode = if (state & 1) == 0 {
+                        RelayRouteMode::Direct
+                    } else {
+                        RelayRouteMode::Middle
+                    };
+                    if runtime.set_mode(mode).is_some() {
+                        successful_transitions.fetch_add(1, Ordering::Relaxed);
+                    }
+                }
+            }));
+        }
+
+        for worker in workers {
+            worker.join().expect("route mode transition worker must not panic");
+        }
+    });
+
+    let final_state = runtime.snapshot();
+    assert_eq!(
+        final_state.generation,
+        successful_transitions.load(Ordering::Relaxed),
+        "final generation must equal number of accepted mode transitions"
+    );
+    assert_eq!(
+        final_state,
+        *runtime.subscribe().borrow(),
+        "watch and snapshot state must match after concurrent transition accounting"
+    );
+}
+
+#[test]
+fn light_fuzz_cutover_stagger_delay_distribution_stays_in_fixed_window() {
+    // Deterministic xorshift fuzzing keeps this test stable across runs.
+    let mut s: u64 = 0x9E37_79B9_7F4A_7C15;
+
+    for _ in 0..20_000 {
+        s ^= s << 7;
+        s ^= s >> 9;
+        s ^= s << 8;
+        let session_id = s;
+
+        s ^= s << 7;
+        s ^= s >> 9;
+        s ^= s << 8;
+        let generation = s;
+
+        let delay = cutover_stagger_delay(session_id, generation);
+        assert!(
+            (1000..=1999).contains(&delay.as_millis()),
+            "fuzzed inputs must always map into fixed stagger window"
+        );
+    }
+}
+
+#[test]
+fn cutover_stagger_delay_distribution_has_no_empty_buckets_under_sequential_sessions() {
+    let mut buckets = [0usize; 1000];
+    let generation = 4242u64;
+
+    for session_id in 0..250_000u64 {
+        let delay_ms = cutover_stagger_delay(session_id, generation).as_millis() as usize;
+        let idx = delay_ms - 1000;
+        buckets[idx] += 1;
+    }
+
+    let empty = buckets.iter().filter(|&&count| count == 0).count();
+    assert_eq!(
+        empty, 0,
+        "all 1000 delay buckets must be exercised to avoid cutover herd clustering"
+    );
+}
+
+#[test]
+fn light_fuzz_cutover_stagger_delay_distribution_stays_reasonably_uniform() {
+    let mut buckets = [0usize; 1000];
+    let mut s: u64 = 0x1BAD_B002_CAFE_F00D;
+
+    for _ in 0..300_000usize {
+        s ^= s << 7;
+        s ^= s >> 9;
+        s ^= s << 8;
+        let session_id = s;
+
+        s ^= s << 7;
+        s ^= s >> 9;
+        s ^= s << 8;
+        let generation = s;
+
+        let delay_ms = cutover_stagger_delay(session_id, generation).as_millis() as usize;
+        buckets[delay_ms - 1000] += 1;
+    }
+
+    let min = *buckets.iter().min().unwrap_or(&0);
+    let max = *buckets.iter().max().unwrap_or(&0);
+    assert!(min > 0, "fuzzed distribution must not leave empty buckets");
+    assert!(
+        max <= min.saturating_mul(3),
+        "bucket skew is too high for anti-herd staggering (max={max}, min={min})"
+    );
+}
+
+#[test]
+fn stress_cutover_stagger_delay_distribution_remains_stable_across_generations() {
+    for generation in [0u64, 1, 7, 31, 255, 1024, u32::MAX as u64, u64::MAX - 1] {
+        let mut buckets = [0usize; 1000];
+        for session_id in 0..100_000u64 {
+            let delay_ms = cutover_stagger_delay(session_id ^ 0x9E37_79B9, generation)
+                .as_millis() as usize;
+            buckets[delay_ms - 1000] += 1;
+        }
+
+        let min = *buckets.iter().min().unwrap_or(&0);
+        let max = *buckets.iter().max().unwrap_or(&0);
+        assert!(
+            max <= min.saturating_mul(4).max(1),
+            "generation={generation}: distribution collapsed (max={max}, min={min})"
+        );
+    }
+}

src/stats/connection_lease_security_tests.rs

@@ -0,0 +1,265 @@
+use super::*;
+use std::panic::{self, AssertUnwindSafe};
+use std::sync::Arc;
+use std::time::Duration;
+use tokio::sync::Barrier;
+
+#[test]
+fn direct_connection_lease_balances_on_drop() {
+    let stats = Arc::new(Stats::new());
+    assert_eq!(stats.get_current_connections_direct(), 0);
+
+    {
+        let _lease = stats.acquire_direct_connection_lease();
+        assert_eq!(stats.get_current_connections_direct(), 1);
+    }
+
+    assert_eq!(stats.get_current_connections_direct(), 0);
+}
+
+#[test]
+fn middle_connection_lease_balances_on_drop() {
+    let stats = Arc::new(Stats::new());
+    assert_eq!(stats.get_current_connections_me(), 0);
+
+    {
+        let _lease = stats.acquire_me_connection_lease();
+        assert_eq!(stats.get_current_connections_me(), 1);
+    }
+
+    assert_eq!(stats.get_current_connections_me(), 0);
+}
+
+#[test]
+fn connection_lease_disarm_prevents_double_release() {
+    let stats = Arc::new(Stats::new());
+
+    let mut lease = stats.acquire_direct_connection_lease();
+    assert_eq!(stats.get_current_connections_direct(), 1);
+
+    stats.decrement_current_connections_direct();
+    assert_eq!(stats.get_current_connections_direct(), 0);
+
+    lease.disarm();
+    drop(lease);
+
+    assert_eq!(stats.get_current_connections_direct(), 0);
+}
+
+#[test]
+fn direct_connection_lease_balances_on_panic_unwind() {
+    let stats = Arc::new(Stats::new());
+    let stats_for_panic = stats.clone();
+
+    let panic_result = panic::catch_unwind(AssertUnwindSafe(move || {
+        let _lease = stats_for_panic.acquire_direct_connection_lease();
+        panic!("intentional panic to verify lease drop path");
+    }));
+
+    assert!(panic_result.is_err(), "panic must propagate from test closure");
+    assert_eq!(
+        stats.get_current_connections_direct(),
+        0,
+        "panic unwind must release direct route gauge"
+    );
+}
+
+#[test]
+fn middle_connection_lease_balances_on_panic_unwind() {
+    let stats = Arc::new(Stats::new());
+    let stats_for_panic = stats.clone();
+
+    let panic_result = panic::catch_unwind(AssertUnwindSafe(move || {
+        let _lease = stats_for_panic.acquire_me_connection_lease();
+        panic!("intentional panic to verify middle lease drop path");
+    }));
+
+    assert!(panic_result.is_err(), "panic must propagate from test closure");
+    assert_eq!(
+        stats.get_current_connections_me(),
+        0,
+        "panic unwind must release middle route gauge"
+    );
+}
+
+#[tokio::test]
+async fn concurrent_mixed_route_lease_churn_balances_to_zero() {
+    const TASKS: usize = 48;
+    const ITERATIONS_PER_TASK: usize = 256;
+
+    let stats = Arc::new(Stats::new());
+    let barrier = Arc::new(Barrier::new(TASKS));
+    let mut workers = Vec::with_capacity(TASKS);
+
+    for task_idx in 0..TASKS {
+        let stats_for_task = stats.clone();
+        let barrier_for_task = barrier.clone();
+        workers.push(tokio::spawn(async move {
+            barrier_for_task.wait().await;
+            for iter in 0..ITERATIONS_PER_TASK {
+                if (task_idx + iter) % 2 == 0 {
+                    let _lease = stats_for_task.acquire_direct_connection_lease();
+                    tokio::task::yield_now().await;
+                } else {
+                    let _lease = stats_for_task.acquire_me_connection_lease();
+                    tokio::task::yield_now().await;
+                }
+            }
+        }));
+    }
+
+    for worker in workers {
+        worker
+            .await
+            .expect("lease churn worker must not panic");
+    }
+
+    assert_eq!(
+        stats.get_current_connections_direct(),
+        0,
+        "direct route gauge must return to zero after concurrent lease churn"
+    );
+    assert_eq!(
+        stats.get_current_connections_me(),
+        0,
+        "middle route gauge must return to zero after concurrent lease churn"
+    );
+}
+
+#[tokio::test]
+async fn abort_storm_mixed_route_leases_returns_all_gauges_to_zero() {
+    const TASKS: usize = 64;
+
+    let stats = Arc::new(Stats::new());
+    let mut workers = Vec::with_capacity(TASKS);
+
+    for task_idx in 0..TASKS {
+        let stats_for_task = stats.clone();
+        workers.push(tokio::spawn(async move {
+            if task_idx % 2 == 0 {
+                let _lease = stats_for_task.acquire_direct_connection_lease();
+                tokio::time::sleep(Duration::from_secs(60)).await;
+            } else {
+                let _lease = stats_for_task.acquire_me_connection_lease();
+                tokio::time::sleep(Duration::from_secs(60)).await;
+            }
+        }));
+    }
+
+    tokio::time::timeout(Duration::from_secs(2), async {
+        loop {
+            let total = stats.get_current_connections_direct() + stats.get_current_connections_me();
+            if total == TASKS as u64 {
+                break;
+            }
+            tokio::time::sleep(Duration::from_millis(10)).await;
+        }
+    })
+    .await
+    .expect("all storm tasks must acquire route leases before abort");
+
+    for worker in &workers {
+        worker.abort();
+    }
+    for worker in workers {
+        let joined = worker.await;
+        assert!(joined.is_err(), "aborted worker must return join error");
+    }
+
+    tokio::time::timeout(Duration::from_secs(2), async {
+        loop {
+            if stats.get_current_connections_direct() == 0 && stats.get_current_connections_me() == 0 {
+                break;
+            }
+            tokio::time::sleep(Duration::from_millis(10)).await;
+        }
+    })
+    .await
+    .expect("all route gauges must drain to zero after abort storm");
+}
+
+#[test]
+fn saturating_route_decrements_do_not_underflow_under_race() {
+    const THREADS: usize = 16;
+    const DECREMENTS_PER_THREAD: usize = 4096;
+
+    let stats = Arc::new(Stats::new());
+    let mut workers = Vec::with_capacity(THREADS);
+
+    for _ in 0..THREADS {
+        let stats_for_thread = stats.clone();
+        workers.push(std::thread::spawn(move || {
+            for _ in 0..DECREMENTS_PER_THREAD {
+                stats_for_thread.decrement_current_connections_direct();
+                stats_for_thread.decrement_current_connections_me();
+            }
+        }));
+    }
+
+    for worker in workers {
+        worker
+            .join()
+            .expect("decrement race worker must not panic");
+    }
+
+    assert_eq!(
+        stats.get_current_connections_direct(),
+        0,
+        "direct route decrement races must never underflow"
+    );
+    assert_eq!(
+        stats.get_current_connections_me(),
+        0,
+        "middle route decrement races must never underflow"
+    );
+}
+
+#[tokio::test]
+async fn direct_connection_lease_balances_on_task_abort() {
+    let stats = Arc::new(Stats::new());
+    let stats_for_task = stats.clone();
+
+    let task = tokio::spawn(async move {
+        let _lease = stats_for_task.acquire_direct_connection_lease();
+        tokio::time::sleep(Duration::from_secs(60)).await;
+    });
+
+    tokio::time::sleep(Duration::from_millis(20)).await;
+    assert_eq!(stats.get_current_connections_direct(), 1);
+
+    task.abort();
+    let joined = task.await;
+    assert!(joined.is_err(), "aborted task must return a join error");
+
+    tokio::time::sleep(Duration::from_millis(20)).await;
+    assert_eq!(
+        stats.get_current_connections_direct(),
+        0,
+        "aborted task must release direct route gauge"
+    );
+}
+
+#[tokio::test]
+async fn middle_connection_lease_balances_on_task_abort() {
+    let stats = Arc::new(Stats::new());
+    let stats_for_task = stats.clone();
+
+    let task = tokio::spawn(async move {
+        let _lease = stats_for_task.acquire_me_connection_lease();
+        tokio::time::sleep(Duration::from_secs(60)).await;
+    });
+
+    tokio::time::sleep(Duration::from_millis(20)).await;
+    assert_eq!(stats.get_current_connections_me(), 1);
+
+    task.abort();
+    let joined = task.await;
+    assert!(joined.is_err(), "aborted task must return a join error");
+
+    tokio::time::sleep(Duration::from_millis(20)).await;
+    assert_eq!(
+        stats.get_current_connections_me(),
+        0,
+        "aborted task must release middle route gauge"
+    );
+}

src/stats/mod.rs

@@ -6,6 +6,7 @@ pub mod beobachten;
 pub mod telemetry;
 
 use std::sync::atomic::{AtomicBool, AtomicU8, AtomicU64, Ordering};
+use std::sync::Arc;
 use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
 use dashmap::DashMap;
 use parking_lot::Mutex;
@@ -19,6 +20,46 @@ use tracing::debug;
 use crate::config::{MeTelemetryLevel, MeWriterPickMode};
 use self::telemetry::TelemetryPolicy;
 
+#[derive(Clone, Copy)]
+enum RouteConnectionGauge {
+    Direct,
+    Middle,
+}
+
+#[must_use = "RouteConnectionLease must be kept alive to hold the connection gauge increment"]
+pub struct RouteConnectionLease {
+    stats: Arc<Stats>,
+    gauge: RouteConnectionGauge,
+    active: bool,
+}
+
+impl RouteConnectionLease {
+    fn new(stats: Arc<Stats>, gauge: RouteConnectionGauge) -> Self {
+        Self {
+            stats,
+            gauge,
+            active: true,
+        }
+    }
+
+    #[cfg(test)]
+    fn disarm(&mut self) {
+        self.active = false;
+    }
+}
+
+impl Drop for RouteConnectionLease {
+    fn drop(&mut self) {
+        if !self.active {
+            return;
+        }
+        match self.gauge {
+            RouteConnectionGauge::Direct => self.stats.decrement_current_connections_direct(),
+            RouteConnectionGauge::Middle => self.stats.decrement_current_connections_me(),
+        }
+    }
+}
+
 // ============= Stats =============
 
 #[derive(Default)]
@@ -285,6 +326,16 @@ impl Stats {
     pub fn decrement_current_connections_me(&self) {
         Self::decrement_atomic_saturating(&self.current_connections_me);
     }
+
+    pub fn acquire_direct_connection_lease(self: &Arc<Self>) -> RouteConnectionLease {
+        self.increment_current_connections_direct();
+        RouteConnectionLease::new(self.clone(), RouteConnectionGauge::Direct)
+    }
+
+    pub fn acquire_me_connection_lease(self: &Arc<Self>) -> RouteConnectionLease {
+        self.increment_current_connections_me();
+        RouteConnectionLease::new(self.clone(), RouteConnectionGauge::Middle)
+    }
     pub fn increment_handshake_timeouts(&self) {
         if self.telemetry_core_enabled() {
             self.handshake_timeouts.fetch_add(1, Ordering::Relaxed);
@@ -1256,6 +1307,33 @@ impl Stats {
         Self::touch_user_stats(stats.value());
         stats.curr_connects.fetch_add(1, Ordering::Relaxed);
     }
+
+    pub fn try_acquire_user_curr_connects(&self, user: &str, limit: Option<u64>) -> bool {
+        if !self.telemetry_user_enabled() {
+            return true;
+        }
+
+        self.maybe_cleanup_user_stats();
+        let stats = self.user_stats.entry(user.to_string()).or_default();
+        Self::touch_user_stats(stats.value());
+
+        let counter = &stats.curr_connects;
+        let mut current = counter.load(Ordering::Relaxed);
+        loop {
+            if let Some(max) = limit && current >= max {
+                return false;
+            }
+            match counter.compare_exchange_weak(
+                current,
+                current.saturating_add(1),
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => return true,
+                Err(actual) => current = actual,
+            }
+        }
+    }
     
     pub fn decrement_user_curr_connects(&self, user: &str) {
         self.maybe_cleanup_user_stats();
@@ -1430,9 +1508,11 @@ impl Stats {
 // ============= Replay Checker =============
 
 pub struct ReplayChecker {
-    shards: Vec<Mutex<ReplayShard>>,
+    handshake_shards: Vec<Mutex<ReplayShard>>,
+    tls_shards: Vec<Mutex<ReplayShard>>,
     shard_mask: usize,
     window: Duration,
+    tls_window: Duration,
     checks: AtomicU64,
     hits: AtomicU64,
     additions: AtomicU64,
@@ -1509,19 +1589,24 @@ impl ReplayShard {
 
 impl ReplayChecker {
     pub fn new(total_capacity: usize, window: Duration) -> Self {
+        const MIN_TLS_REPLAY_WINDOW: Duration = Duration::from_secs(120);
         let num_shards = 64;
         let shard_capacity = (total_capacity / num_shards).max(1);
         let cap = NonZeroUsize::new(shard_capacity).unwrap();
 
-        let mut shards = Vec::with_capacity(num_shards);
+        let mut handshake_shards = Vec::with_capacity(num_shards);
+        let mut tls_shards = Vec::with_capacity(num_shards);
         for _ in 0..num_shards {
-            shards.push(Mutex::new(ReplayShard::new(cap)));
+            handshake_shards.push(Mutex::new(ReplayShard::new(cap)));
+            tls_shards.push(Mutex::new(ReplayShard::new(cap)));
         }
 
         Self {
-            shards,
+            handshake_shards,
+            tls_shards,
             shard_mask: num_shards - 1,
             window,
+            tls_window: window.max(MIN_TLS_REPLAY_WINDOW),
             checks: AtomicU64::new(0),
             hits: AtomicU64::new(0),
             additions: AtomicU64::new(0),
@@ -1535,46 +1620,60 @@ impl ReplayChecker {
         (hasher.finish() as usize) & self.shard_mask
     }
 
-    fn check_and_add_internal(&self, data: &[u8]) -> bool {
+    fn check_and_add_internal(
+        &self,
+        data: &[u8],
+        shards: &[Mutex<ReplayShard>],
+        window: Duration,
+    ) -> bool {
         self.checks.fetch_add(1, Ordering::Relaxed);
         let idx = self.get_shard_idx(data);
-        let mut shard = self.shards[idx].lock();
+        let mut shard = shards[idx].lock();
         let now = Instant::now();
-        let found = shard.check(data, now, self.window);
+        let found = shard.check(data, now, window);
         if found {
             self.hits.fetch_add(1, Ordering::Relaxed);
         } else {
-            shard.add(data, now, self.window);
+            shard.add(data, now, window);
             self.additions.fetch_add(1, Ordering::Relaxed);
         }
         found
     }
 
-    fn add_only(&self, data: &[u8]) {
+    fn add_only(&self, data: &[u8], shards: &[Mutex<ReplayShard>], window: Duration) {
         self.additions.fetch_add(1, Ordering::Relaxed);
         let idx = self.get_shard_idx(data);
-        let mut shard = self.shards[idx].lock();
-        shard.add(data, Instant::now(), self.window);
+        let mut shard = shards[idx].lock();
+        shard.add(data, Instant::now(), window);
     }
 
     pub fn check_and_add_handshake(&self, data: &[u8]) -> bool {
-        self.check_and_add_internal(data)
+        self.check_and_add_internal(data, &self.handshake_shards, self.window)
     }
 
     pub fn check_and_add_tls_digest(&self, data: &[u8]) -> bool {
-        self.check_and_add_internal(data)
+        self.check_and_add_internal(data, &self.tls_shards, self.tls_window)
     }
 
     // Compatibility helpers (non-atomic split operations) — prefer check_and_add_*.
     pub fn check_handshake(&self, data: &[u8]) -> bool { self.check_and_add_handshake(data) }
-    pub fn add_handshake(&self, data: &[u8]) { self.add_only(data) }
+    pub fn add_handshake(&self, data: &[u8]) {
+        self.add_only(data, &self.handshake_shards, self.window)
+    }
     pub fn check_tls_digest(&self, data: &[u8]) -> bool { self.check_and_add_tls_digest(data) }
-    pub fn add_tls_digest(&self, data: &[u8]) { self.add_only(data) }
+    pub fn add_tls_digest(&self, data: &[u8]) {
+        self.add_only(data, &self.tls_shards, self.tls_window)
+    }
     
     pub fn stats(&self) -> ReplayStats {
         let mut total_entries = 0;
         let mut total_queue_len = 0;
-        for shard in &self.shards {
+        for shard in &self.handshake_shards {
+            let s = shard.lock();
+            total_entries += s.cache.len();
+            total_queue_len += s.queue.len();
+        }
+        for shard in &self.tls_shards {
             let s = shard.lock();
             total_entries += s.cache.len();
             total_queue_len += s.queue.len();
@@ -1587,7 +1686,7 @@ impl ReplayChecker {
             total_hits: self.hits.load(Ordering::Relaxed),
             total_additions: self.additions.load(Ordering::Relaxed),
             total_cleanups: self.cleanups.load(Ordering::Relaxed),
-            num_shards: self.shards.len(),
+            num_shards: self.handshake_shards.len() + self.tls_shards.len(),
             window_secs: self.window.as_secs(),
         }
     }
@@ -1605,13 +1704,20 @@ impl ReplayChecker {
             let now = Instant::now();
             let mut cleaned = 0usize;
             
-            for shard_mutex in &self.shards {
+            for shard_mutex in &self.handshake_shards {
                 let mut shard = shard_mutex.lock();
                 let before = shard.len();
                 shard.cleanup(now, self.window);
                 let after = shard.len();
                 cleaned += before.saturating_sub(after);
             }
+            for shard_mutex in &self.tls_shards {
+                let mut shard = shard_mutex.lock();
+                let before = shard.len();
+                shard.cleanup(now, self.tls_window);
+                let after = shard.len();
+                cleaned += before.saturating_sub(after);
+            }
             
             self.cleanups.fetch_add(1, Ordering::Relaxed);
             
@@ -1737,7 +1843,7 @@ mod tests {
     fn test_replay_checker_many_keys() {
         let checker = ReplayChecker::new(10_000, Duration::from_secs(60));
         for i in 0..500u32 {
-            checker.add_only(&i.to_le_bytes());
+            checker.add_handshake(&i.to_le_bytes());
         }
         for i in 0..500u32 {
             assert!(checker.check_handshake(&i.to_le_bytes()));
@@ -1745,3 +1851,11 @@ mod tests {
         assert_eq!(checker.stats().total_entries, 500);
     }
 }
+
+#[cfg(test)]
+#[path = "connection_lease_security_tests.rs"]
+mod connection_lease_security_tests;
+
+#[cfg(test)]
+#[path = "replay_checker_security_tests.rs"]
+mod replay_checker_security_tests;

src/stats/replay_checker_security_tests.rs

@@ -0,0 +1,80 @@
+use super::*;
+use std::time::Duration;
+
+#[test]
+fn replay_checker_keeps_tls_and_handshake_domains_isolated_for_same_key() {
+    let checker = ReplayChecker::new(128, Duration::from_millis(20));
+    let key = b"same-key-domain-separation";
+
+    assert!(
+        !checker.check_and_add_handshake(key),
+        "first handshake use should be fresh"
+    );
+    assert!(
+        !checker.check_and_add_tls_digest(key),
+        "same bytes in TLS domain should still be fresh"
+    );
+
+    assert!(
+        checker.check_and_add_handshake(key),
+        "second handshake use should be replay-hit"
+    );
+    assert!(
+        checker.check_and_add_tls_digest(key),
+        "second TLS use should be replay-hit independently"
+    );
+}
+
+#[test]
+fn replay_checker_tls_window_is_clamped_beyond_small_handshake_window() {
+    let checker = ReplayChecker::new(128, Duration::from_millis(20));
+    let handshake_key = b"short-window-handshake";
+    let tls_key = b"short-window-tls";
+
+    assert!(!checker.check_and_add_handshake(handshake_key));
+    assert!(!checker.check_and_add_tls_digest(tls_key));
+
+    std::thread::sleep(Duration::from_millis(80));
+
+    assert!(
+        !checker.check_and_add_handshake(handshake_key),
+        "handshake key should expire under short configured window"
+    );
+    assert!(
+        checker.check_and_add_tls_digest(tls_key),
+        "TLS key should still be replay-hit because TLS window is clamped to a secure minimum"
+    );
+}
+
+#[test]
+fn replay_checker_compat_add_paths_do_not_cross_pollute_domains() {
+    let checker = ReplayChecker::new(128, Duration::from_secs(1));
+    let key = b"compat-domain-separation";
+
+    checker.add_handshake(key);
+    assert!(
+        checker.check_and_add_handshake(key),
+        "handshake add helper must populate handshake domain"
+    );
+    assert!(
+        !checker.check_and_add_tls_digest(key),
+        "handshake add helper must not pollute TLS domain"
+    );
+
+    checker.add_tls_digest(key);
+    assert!(
+        checker.check_and_add_tls_digest(key),
+        "TLS add helper must populate TLS domain"
+    );
+}
+
+#[test]
+fn replay_checker_stats_reflect_dual_shard_domains() {
+    let checker = ReplayChecker::new(128, Duration::from_secs(1));
+    let stats = checker.stats();
+
+    assert_eq!(
+        stats.num_shards, 128,
+        "stats should expose both shard domains (handshake + TLS)"
+    );
+}

src/stream/frame_codec.rs

@@ -513,6 +513,7 @@ impl FrameCodecTrait for SecureCodec {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use std::collections::HashSet;
     use tokio_util::codec::{FramedRead, FramedWrite};
     use tokio::io::duplex;
     use futures::{SinkExt, StreamExt};
@@ -630,4 +631,31 @@ mod tests {
         let result = codec.decode(&mut buf);
         assert!(result.is_err());
     }
+
+    #[test]
+    fn secure_codec_always_adds_padding_and_jitters_wire_length() {
+        let codec = SecureCodec::new(Arc::new(SecureRandom::new()));
+        let payload = Bytes::from_static(&[1, 2, 3, 4, 5, 6, 7, 8]);
+        let mut wire_lens = HashSet::new();
+
+        for _ in 0..64 {
+            let frame = Frame::new(payload.clone());
+            let mut out = BytesMut::new();
+            codec.encode(&frame, &mut out).unwrap();
+
+            assert!(out.len() >= 4 + payload.len() + 1);
+            let wire_len = u32::from_le_bytes([out[0], out[1], out[2], out[3]]) as usize;
+            assert!(
+                (payload.len() + 1..=payload.len() + 3).contains(&wire_len),
+                "Secure wire length must be payload+1..3, got {wire_len}"
+            );
+            assert_ne!(wire_len % 4, 0, "Secure wire length must be non-4-aligned");
+            wire_lens.insert(wire_len);
+        }
+
+        assert!(
+            wire_lens.len() >= 2,
+            "Secure padding should create observable wire-length jitter"
+        );
+    }
 }

src/tls_front/emulator.rs

@@ -117,15 +117,6 @@ pub fn build_emulated_server_hello(
     extensions.extend_from_slice(&0x002bu16.to_be_bytes());
     extensions.extend_from_slice(&(2u16).to_be_bytes());
     extensions.extend_from_slice(&0x0304u16.to_be_bytes());
-    if let Some(alpn_proto) = &alpn {
-        extensions.extend_from_slice(&0x0010u16.to_be_bytes());
-        let list_len: u16 = 1 + alpn_proto.len() as u16;
-        let ext_len: u16 = 2 + list_len;
-        extensions.extend_from_slice(&ext_len.to_be_bytes());
-        extensions.extend_from_slice(&list_len.to_be_bytes());
-        extensions.push(alpn_proto.len() as u8);
-        extensions.extend_from_slice(alpn_proto);
-    }
     let extensions_len = extensions.len() as u16;
 
     let body_len = 2 + // version
@@ -207,8 +198,22 @@ pub fn build_emulated_server_hello(
     }
 
     let mut app_data = Vec::new();
+    let alpn_marker = alpn
+        .as_ref()
+        .filter(|p| !p.is_empty() && p.len() <= u8::MAX as usize)
+        .map(|proto| {
+            let proto_list_len = 1usize + proto.len();
+            let ext_data_len = 2usize + proto_list_len;
+            let mut marker = Vec::with_capacity(4 + ext_data_len);
+            marker.extend_from_slice(&0x0010u16.to_be_bytes());
+            marker.extend_from_slice(&(ext_data_len as u16).to_be_bytes());
+            marker.extend_from_slice(&(proto_list_len as u16).to_be_bytes());
+            marker.push(proto.len() as u8);
+            marker.extend_from_slice(proto);
+            marker
+        });
     let mut payload_offset = 0usize;
-    for size in sizes {
+    for (idx, size) in sizes.into_iter().enumerate() {
         let mut rec = Vec::with_capacity(5 + size);
         rec.push(TLS_RECORD_APPLICATION);
         rec.extend_from_slice(&TLS_VERSION);
@@ -233,7 +238,20 @@ pub fn build_emulated_server_hello(
             }
         } else if size > 17 {
             let body_len = size - 17;
-            rec.extend_from_slice(&rng.bytes(body_len));
+            let mut body = Vec::with_capacity(body_len);
+            if idx == 0 && let Some(marker) = &alpn_marker {
+                if marker.len() <= body_len {
+                    body.extend_from_slice(marker);
+                    if body_len > marker.len() {
+                        body.extend_from_slice(&rng.bytes(body_len - marker.len()));
+                    }
+                } else {
+                    body.extend_from_slice(&rng.bytes(body_len));
+                }
+            } else {
+                body.extend_from_slice(&rng.bytes(body_len));
+            }
+            rec.extend_from_slice(&body);
             rec.push(0x16); // inner content type marker (handshake)
             rec.extend_from_slice(&rng.bytes(16)); // AEAD-like tag
         } else {
@@ -245,8 +263,9 @@ pub fn build_emulated_server_hello(
     // --- Combine ---
     // Optional NewSessionTicket mimic records (opaque ApplicationData for fingerprint).
     let mut tickets = Vec::new();
-    if new_session_tickets > 0 {
-        for _ in 0..new_session_tickets {
+    let ticket_count = new_session_tickets.min(4);
+    if ticket_count > 0 {
+        for _ in 0..ticket_count {
             let ticket_len: usize = rng.range(48) + 48;
             let mut rec = Vec::with_capacity(5 + ticket_len);
             rec.push(TLS_RECORD_APPLICATION);
@@ -273,6 +292,10 @@ pub fn build_emulated_server_hello(
     response
 }
 
+#[cfg(test)]
+#[path = "emulator_security_tests.rs"]
+mod security_tests;
+
 #[cfg(test)]
 mod tests {
     use std::time::SystemTime;

src/tls_front/emulator_security_tests.rs

@@ -0,0 +1,136 @@
+use std::time::SystemTime;
+
+use crate::crypto::SecureRandom;
+use crate::protocol::constants::{TLS_RECORD_APPLICATION, TLS_RECORD_CHANGE_CIPHER, TLS_RECORD_HANDSHAKE};
+use crate::tls_front::emulator::build_emulated_server_hello;
+use crate::tls_front::types::{
+    CachedTlsData, ParsedServerHello, TlsBehaviorProfile, TlsCertPayload, TlsProfileSource,
+};
+
+fn make_cached(cert_payload: Option<crate::tls_front::types::TlsCertPayload>) -> CachedTlsData {
+    CachedTlsData {
+        server_hello_template: ParsedServerHello {
+            version: [0x03, 0x03],
+            random: [0u8; 32],
+            session_id: Vec::new(),
+            cipher_suite: [0x13, 0x01],
+            compression: 0,
+            extensions: Vec::new(),
+        },
+        cert_info: None,
+        cert_payload,
+        app_data_records_sizes: vec![64],
+        total_app_data_len: 64,
+        behavior_profile: TlsBehaviorProfile {
+            change_cipher_spec_count: 1,
+            app_data_record_sizes: vec![64],
+            ticket_record_sizes: Vec::new(),
+            source: TlsProfileSource::Default,
+        },
+        fetched_at: SystemTime::now(),
+        domain: "example.com".to_string(),
+    }
+}
+
+fn first_app_data_payload(response: &[u8]) -> &[u8] {
+    let hello_len = u16::from_be_bytes([response[3], response[4]]) as usize;
+    let ccs_start = 5 + hello_len;
+    let ccs_len = u16::from_be_bytes([response[ccs_start + 3], response[ccs_start + 4]]) as usize;
+    let app_start = ccs_start + 5 + ccs_len;
+    let app_len = u16::from_be_bytes([response[app_start + 3], response[app_start + 4]]) as usize;
+    &response[app_start + 5..app_start + 5 + app_len]
+}
+
+#[test]
+fn emulated_server_hello_ignores_oversized_alpn_when_marker_would_not_fit() {
+    let cached = make_cached(None);
+    let rng = SecureRandom::new();
+    let oversized_alpn = vec![0xAB; u8::MAX as usize + 1];
+
+    let response = build_emulated_server_hello(
+        b"secret",
+        &[0x11; 32],
+        &[0x22; 16],
+        &cached,
+        true,
+        &rng,
+        Some(oversized_alpn),
+        0,
+    );
+
+    assert_eq!(response[0], TLS_RECORD_HANDSHAKE);
+    let hello_len = u16::from_be_bytes([response[3], response[4]]) as usize;
+    let ccs_start = 5 + hello_len;
+    assert_eq!(response[ccs_start], TLS_RECORD_CHANGE_CIPHER);
+    let app_start = ccs_start + 6;
+    assert_eq!(response[app_start], TLS_RECORD_APPLICATION);
+
+    let payload = first_app_data_payload(&response);
+    let mut marker_prefix = Vec::new();
+    marker_prefix.extend_from_slice(&0x0010u16.to_be_bytes());
+    marker_prefix.extend_from_slice(&0x0102u16.to_be_bytes());
+    marker_prefix.extend_from_slice(&0x0100u16.to_be_bytes());
+    marker_prefix.push(0xff);
+    marker_prefix.extend_from_slice(&[0xab; 8]);
+    assert!(
+        !payload.starts_with(&marker_prefix),
+        "oversized ALPN must not be partially embedded into the emulated first application record"
+    );
+}
+
+#[test]
+fn emulated_server_hello_embeds_full_alpn_marker_when_body_can_fit() {
+    let cached = make_cached(None);
+    let rng = SecureRandom::new();
+
+    let response = build_emulated_server_hello(
+        b"secret",
+        &[0x31; 32],
+        &[0x41; 16],
+        &cached,
+        true,
+        &rng,
+        Some(b"h2".to_vec()),
+        0,
+    );
+
+    let payload = first_app_data_payload(&response);
+    let expected = [0x00u8, 0x10, 0x00, 0x05, 0x00, 0x03, 0x02, b'h', b'2'];
+    assert!(
+        payload.starts_with(&expected),
+        "when body has enough capacity, emulated first application record must include full ALPN marker"
+    );
+}
+
+#[test]
+fn emulated_server_hello_prefers_cert_payload_over_alpn_marker() {
+    let cert_msg = vec![0x0b, 0x00, 0x00, 0x05, 0x00, 0xaa, 0xbb, 0xcc, 0xdd];
+    let cached = make_cached(Some(TlsCertPayload {
+        cert_chain_der: vec![vec![0x30, 0x01, 0x00]],
+        certificate_message: cert_msg.clone(),
+    }));
+    let rng = SecureRandom::new();
+
+    let response = build_emulated_server_hello(
+        b"secret",
+        &[0x32; 32],
+        &[0x42; 16],
+        &cached,
+        true,
+        &rng,
+        Some(b"h2".to_vec()),
+        0,
+    );
+
+    let payload = first_app_data_payload(&response);
+    let alpn_marker = [0x00u8, 0x10, 0x00, 0x05, 0x00, 0x03, 0x02, b'h', b'2'];
+
+    assert!(
+        payload.starts_with(&cert_msg),
+        "when certificate payload is available, first record must start with cert payload bytes"
+    );
+    assert!(
+        !payload.starts_with(&alpn_marker),
+        "ALPN marker must not displace selected certificate payload"
+    );
+}

src/transport/middle_proxy/health.rs

@@ -124,12 +124,12 @@ pub(super) async fn reap_draining_writers(
     let drain_threshold = pool
         .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 draining_writers = Vec::<DrainingWriterSnapshot>::new();
     let mut empty_writer_ids = Vec::<u64>::new();
     let mut force_close_writer_ids = Vec::<u64>::new();
-    for writer in writers {
+    let writers = pool.writers.read().await;
+    for writer in writers.iter() {
         if !writer.draining.load(std::sync::atomic::Ordering::Relaxed) {
             continue;
         }
@@ -143,23 +143,38 @@ pub(super) async fn reap_draining_writers(
             empty_writer_ids.push(writer.id);
             continue;
         }
-        draining_writers.push(writer);
+        draining_writers.push(DrainingWriterSnapshot {
+            id: writer.id,
+            writer_dc: writer.writer_dc,
+            addr: writer.addr,
+            generation: writer.generation,
+            created_at: writer.created_at,
+            draining_started_at_epoch_secs: writer
+                .draining_started_at_epoch_secs
+                .load(std::sync::atomic::Ordering::Relaxed),
+            drain_deadline_epoch_secs: writer
+                .drain_deadline_epoch_secs
+                .load(std::sync::atomic::Ordering::Relaxed),
+            allow_drain_fallback: writer
+                .allow_drain_fallback
+                .load(std::sync::atomic::Ordering::Relaxed),
+        });
     }
+    drop(writers);
 
-    if drain_threshold > 0 && draining_writers.len() > drain_threshold as usize {
+    let overflow = if drain_threshold > 0 && draining_writers.len() > drain_threshold as usize {
+        draining_writers.len().saturating_sub(drain_threshold as usize)
+    } else {
+        0
+    };
+
+    if overflow > 0 {
         draining_writers.sort_by(|left, right| {
-            let left_started = left
-                .draining_started_at_epoch_secs
-                .load(std::sync::atomic::Ordering::Relaxed);
-            let right_started = right
-                .draining_started_at_epoch_secs
-                .load(std::sync::atomic::Ordering::Relaxed);
-            left_started
-                .cmp(&right_started)
+            left.draining_started_at_epoch_secs
+                .cmp(&right.draining_started_at_epoch_secs)
                 .then_with(|| left.created_at.cmp(&right.created_at))
                 .then_with(|| left.id.cmp(&right.id))
         });
-        let overflow = draining_writers.len().saturating_sub(drain_threshold as usize);
         warn!(
             draining_writers = draining_writers.len(),
             me_pool_drain_threshold = drain_threshold,
@@ -171,15 +186,10 @@ pub(super) async fn reap_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);
         if drain_ttl_secs > 0
-            && drain_started_at_epoch_secs != 0
-            && now_epoch_secs.saturating_sub(drain_started_at_epoch_secs) > drain_ttl_secs
+            && writer.draining_started_at_epoch_secs != 0
+            && now_epoch_secs.saturating_sub(writer.draining_started_at_epoch_secs) > drain_ttl_secs
             && should_emit_writer_warn(
                 warn_next_allowed,
                 writer.id,
@@ -194,22 +204,17 @@ pub(super) async fn reap_draining_writers(
                 generation = writer.generation,
                 drain_ttl_secs,
                 force_close_secs = pool.me_pool_force_close_secs.load(std::sync::atomic::Ordering::Relaxed),
-                allow_drain_fallback = writer.allow_drain_fallback.load(std::sync::atomic::Ordering::Relaxed),
+                allow_drain_fallback = writer.allow_drain_fallback,
                 "ME draining writer remains non-empty past drain TTL"
             );
         }
-        let deadline_epoch_secs = writer
-            .drain_deadline_epoch_secs
-            .load(std::sync::atomic::Ordering::Relaxed);
-        if deadline_epoch_secs != 0 && now_epoch_secs >= deadline_epoch_secs {
+        if writer.drain_deadline_epoch_secs != 0 && now_epoch_secs >= writer.drain_deadline_epoch_secs
+        {
             warn!(writer_id = writer.id, "Drain timeout, force-closing");
             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));
-
     let close_budget = health_drain_close_budget();
     let requested_force_close = force_close_writer_ids.len();
     let requested_empty_close = empty_writer_ids.len();
@@ -234,7 +239,9 @@ pub(super) async fn reap_draining_writers(
         if !closed_writer_ids.insert(writer_id) {
             continue;
         }
-        pool.remove_writer_and_close_clients(writer_id).await;
+        if !pool.remove_writer_if_empty(writer_id).await {
+            continue;
+        }
         closed_total = closed_total.saturating_add(1);
     }
 
@@ -247,6 +254,18 @@ pub(super) async fn reap_draining_writers(
             "ME draining close backlog deferred to next health cycle"
         );
     }
+
+    // Keep warn cooldown state for draining writers still present in the pool;
+    // drop state only once a writer is actually removed.
+    let active_draining_writer_ids = {
+        let writers = pool.writers.read().await;
+        writers
+            .iter()
+            .filter(|writer| writer.draining.load(std::sync::atomic::Ordering::Relaxed))
+            .map(|writer| writer.id)
+            .collect::<HashSet<u64>>()
+    };
+    warn_next_allowed.retain(|writer_id, _| active_draining_writer_ids.contains(writer_id));
 }
 
 pub(super) fn health_drain_close_budget() -> usize {
@@ -258,6 +277,18 @@ pub(super) fn health_drain_close_budget() -> usize {
         .clamp(HEALTH_DRAIN_CLOSE_BUDGET_MIN, HEALTH_DRAIN_CLOSE_BUDGET_MAX)
 }
 
+#[derive(Debug, Clone)]
+struct DrainingWriterSnapshot {
+    id: u64,
+    writer_dc: i32,
+    addr: SocketAddr,
+    generation: u64,
+    created_at: Instant,
+    draining_started_at_epoch_secs: u64,
+    drain_deadline_epoch_secs: u64,
+    allow_drain_fallback: bool,
+}
+
 fn should_emit_writer_warn(
     next_allowed: &mut HashMap<u64, Instant>,
     writer_id: u64,
@@ -1391,6 +1422,15 @@ mod tests {
             me_pool_drain_threshold,
             ..GeneralConfig::default()
         };
+        let mut proxy_map_v4 = HashMap::new();
+        proxy_map_v4.insert(
+            2,
+            vec![(IpAddr::V4(Ipv4Addr::new(203, 0, 113, 10)), 443)],
+        );
+        let decision = NetworkDecision {
+            ipv4_me: true,
+            ..NetworkDecision::default()
+        };
         MePool::new(
             None,
             vec![1u8; 32],
@@ -1402,10 +1442,10 @@ mod tests {
             None,
             12,
             1200,
-            HashMap::new(),
+            proxy_map_v4,
             HashMap::new(),
             None,
-            NetworkDecision::default(),
+            decision,
             None,
             Arc::new(SecureRandom::new()),
             Arc::new(Stats::default()),
@@ -1516,8 +1556,55 @@ mod tests {
         conn_id
     }
 
+    async fn insert_live_writer(pool: &Arc<MePool>, writer_id: u64, writer_dc: i32) {
+        let (tx, _writer_rx) = mpsc::channel::<WriterCommand>(8);
+        let writer = MeWriter {
+            id: writer_id,
+            addr: SocketAddr::new(
+                IpAddr::V4(Ipv4Addr::new(203, 0, 113, (writer_id as u8).saturating_add(1))),
+                4000 + writer_id as u16,
+            ),
+            source_ip: IpAddr::V4(Ipv4Addr::LOCALHOST),
+            writer_dc,
+            generation: 2,
+            contour: Arc::new(AtomicU8::new(WriterContour::Active.as_u8())),
+            created_at: Instant::now(),
+            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(false)),
+            draining_started_at_epoch_secs: Arc::new(AtomicU64::new(0)),
+            drain_deadline_epoch_secs: Arc::new(AtomicU64::new(0)),
+            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);
+    }
+
     #[tokio::test]
     async fn reap_draining_writers_force_closes_oldest_over_threshold() {
+        let pool = make_pool(2).await;
+        insert_live_writer(&pool, 1, 2).await;
+        let now_epoch_secs = MePool::now_epoch_secs();
+        let conn_a = insert_draining_writer(&pool, 10, now_epoch_secs.saturating_sub(30)).await;
+        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();
+
+        reap_draining_writers(&pool, &mut warn_next_allowed).await;
+
+        let mut writer_ids: Vec<u64> = pool.writers.read().await.iter().map(|writer| writer.id).collect();
+        writer_ids.sort_unstable();
+        assert_eq!(writer_ids, vec![1, 20, 30]);
+        assert!(pool.registry.get_writer(conn_a).await.is_none());
+        assert_eq!(pool.registry.get_writer(conn_b).await.unwrap().writer_id, 20);
+        assert_eq!(pool.registry.get_writer(conn_c).await.unwrap().writer_id, 30);
+    }
+
+    #[tokio::test]
+    async fn reap_draining_writers_force_closes_overflow_without_replacement() {
         let pool = make_pool(2).await;
         let now_epoch_secs = MePool::now_epoch_secs();
         let conn_a = insert_draining_writer(&pool, 10, now_epoch_secs.saturating_sub(30)).await;
@@ -1527,7 +1614,8 @@ mod tests {
 
         reap_draining_writers(&pool, &mut warn_next_allowed).await;
 
-        let writer_ids: Vec<u64> = pool.writers.read().await.iter().map(|writer| writer.id).collect();
+        let mut writer_ids: Vec<u64> = pool.writers.read().await.iter().map(|writer| writer.id).collect();
+        writer_ids.sort_unstable();
         assert_eq!(writer_ids, vec![20, 30]);
         assert!(pool.registry.get_writer(conn_a).await.is_none());
         assert_eq!(pool.registry.get_writer(conn_b).await.unwrap().writer_id, 20);

src/transport/middle_proxy/health_adversarial_tests.rs

@@ -1,4 +1,5 @@
 use std::collections::HashMap;
+use std::collections::HashSet;
 use std::net::{IpAddr, Ipv4Addr, SocketAddr};
 use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, AtomicU8, AtomicU32, AtomicU64, Ordering};
@@ -181,6 +182,40 @@ async fn sorted_writer_ids(pool: &Arc<MePool>) -> Vec<u64> {
     ids
 }
 
+fn lcg_next(state: &mut u64) -> u64 {
+    *state = state.wrapping_mul(6364136223846793005).wrapping_add(1);
+    *state
+}
+
+async fn draining_writer_ids(pool: &Arc<MePool>) -> HashSet<u64> {
+    pool.writers
+        .read()
+        .await
+        .iter()
+        .filter(|writer| writer.draining.load(Ordering::Relaxed))
+        .map(|writer| writer.id)
+        .collect::<HashSet<u64>>()
+}
+
+async fn set_writer_runtime_state(
+    pool: &Arc<MePool>,
+    writer_id: u64,
+    draining: bool,
+    drain_started_at_epoch_secs: u64,
+    drain_deadline_epoch_secs: u64,
+) {
+    let writers = pool.writers.read().await;
+    if let Some(writer) = writers.iter().find(|writer| writer.id == writer_id) {
+        writer.draining.store(draining, Ordering::Relaxed);
+        writer
+            .draining_started_at_epoch_secs
+            .store(drain_started_at_epoch_secs, Ordering::Relaxed);
+        writer
+            .drain_deadline_epoch_secs
+            .store(drain_deadline_epoch_secs, Ordering::Relaxed);
+    }
+}
+
 #[tokio::test]
 async fn reap_draining_writers_clears_warn_state_when_pool_empty() {
     let (pool, _rng) = make_pool(128, 1, 1).await;
@@ -430,6 +465,149 @@ async fn me_health_monitor_eliminates_mixed_empty_and_deadline_backlog() {
     assert!(writer_count(&pool).await <= threshold as usize);
 }
 
+#[tokio::test]
+async fn reap_draining_writers_deterministic_mixed_state_churn_preserves_invariants() {
+    let threshold = 9u64;
+    let (pool, _rng) = make_pool(threshold, 1, 1).await;
+    let mut warn_next_allowed = HashMap::new();
+    let mut seed = 0x9E37_79B9_7F4A_7C15u64;
+    let mut next_writer_id = 20_000u64;
+    let now_epoch_secs = MePool::now_epoch_secs();
+
+    for writer_id in 1..=72u64 {
+        let bound_clients = if writer_id % 4 == 0 { 0 } else { 1 };
+        let deadline = if writer_id % 5 == 0 {
+            now_epoch_secs.saturating_sub(1)
+        } else {
+            0
+        };
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(500).saturating_add(writer_id),
+            bound_clients,
+            deadline,
+        )
+        .await;
+    }
+
+    for _round in 0..90 {
+        reap_draining_writers(&pool, &mut warn_next_allowed).await;
+
+        let draining_ids = draining_writer_ids(&pool).await;
+        assert!(
+            warn_next_allowed.keys().all(|id| draining_ids.contains(id)),
+            "warn-state keys must always be a subset of live draining writers"
+        );
+
+        let writer_ids = sorted_writer_ids(&pool).await;
+        if writer_ids.is_empty() {
+            continue;
+        }
+
+        let remove_n = (lcg_next(&mut seed) % 3) as usize;
+        for writer_id in writer_ids.iter().copied().take(remove_n) {
+            let _ = pool.remove_writer_and_close_clients(writer_id).await;
+        }
+
+        let survivors = sorted_writer_ids(&pool).await;
+        if !survivors.is_empty() {
+            let idx = (lcg_next(&mut seed) as usize) % survivors.len();
+            let target = survivors[idx];
+            set_writer_runtime_state(&pool, target, false, 0, 0).await;
+        }
+
+        let survivors = sorted_writer_ids(&pool).await;
+        if survivors.len() > 1 {
+            let idx = (lcg_next(&mut seed) as usize) % survivors.len();
+            let target = survivors[idx];
+            let expired_deadline = if lcg_next(&mut seed) & 1 == 0 {
+                now_epoch_secs.saturating_sub(1)
+            } else {
+                0
+            };
+            set_writer_runtime_state(
+                &pool,
+                target,
+                true,
+                now_epoch_secs.saturating_sub(120),
+                expired_deadline,
+            )
+            .await;
+        }
+
+        let inject_n = (lcg_next(&mut seed) % 4) as usize;
+        for _ in 0..inject_n {
+            let bound_clients = if lcg_next(&mut seed) & 1 == 0 { 0 } else { 1 };
+            let deadline = if lcg_next(&mut seed) & 1 == 0 {
+                now_epoch_secs.saturating_sub(1)
+            } else {
+                0
+            };
+            insert_draining_writer(
+                &pool,
+                next_writer_id,
+                now_epoch_secs.saturating_sub(240),
+                bound_clients,
+                deadline,
+            )
+            .await;
+            next_writer_id = next_writer_id.saturating_add(1);
+        }
+    }
+
+    for _ in 0..64 {
+        reap_draining_writers(&pool, &mut warn_next_allowed).await;
+        if writer_count(&pool).await <= threshold as usize {
+            break;
+        }
+    }
+
+    assert!(writer_count(&pool).await <= threshold as usize);
+    let draining_ids = draining_writer_ids(&pool).await;
+    assert!(warn_next_allowed.keys().all(|id| draining_ids.contains(id)));
+}
+
+#[tokio::test]
+async fn reap_draining_writers_repeated_draining_flips_never_leave_stale_warn_state() {
+    let (pool, _rng) = make_pool(64, 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(240),
+            1,
+            0,
+        )
+        .await;
+    }
+
+    let mut warn_next_allowed = HashMap::new();
+    for _round in 0..48u64 {
+        for writer_id in 1..=24u64 {
+            let draining = (writer_id + _round) % 3 != 0;
+            set_writer_runtime_state(
+                &pool,
+                writer_id,
+                draining,
+                now_epoch_secs.saturating_sub(120),
+                0,
+            )
+            .await;
+        }
+
+        reap_draining_writers(&pool, &mut warn_next_allowed).await;
+
+        let draining_ids = draining_writer_ids(&pool).await;
+        assert!(
+            warn_next_allowed.keys().all(|id| draining_ids.contains(id)),
+            "warn-state map must not retain entries for writers outside draining set"
+        );
+    }
+}
+
 #[test]
 fn health_drain_close_budget_is_within_expected_bounds() {
     let budget = health_drain_close_budget();

src/transport/middle_proxy/health_integration_tests.rs

@@ -161,6 +161,20 @@ async fn insert_draining_writer(
     }
 }
 
+async fn wait_for_pool_empty(pool: &Arc<MePool>, timeout: Duration) {
+    let start = Instant::now();
+    loop {
+        if pool.writers.read().await.is_empty() {
+            return;
+        }
+        assert!(
+            start.elapsed() < timeout,
+            "timed out waiting for pool.writers to become empty"
+        );
+        tokio::time::sleep(Duration::from_millis(5)).await;
+    }
+}
+
 #[tokio::test]
 async fn me_health_monitor_drains_expired_backlog_over_multiple_cycles() {
     let (pool, rng) = make_pool(128, 1, 1).await;
@@ -178,7 +192,7 @@ async fn me_health_monitor_drains_expired_backlog_over_multiple_cycles() {
     }
 
     let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
-    tokio::time::sleep(Duration::from_millis(60)).await;
+    wait_for_pool_empty(&pool, Duration::from_secs(1)).await;
     monitor.abort();
     let _ = monitor.await;
 
@@ -194,7 +208,7 @@ async fn me_health_monitor_cleans_empty_draining_writers_without_force_close() {
     }
 
     let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
-    tokio::time::sleep(Duration::from_millis(30)).await;
+    wait_for_pool_empty(&pool, Duration::from_secs(1)).await;
     monitor.abort();
     let _ = monitor.await;
 
@@ -219,7 +233,7 @@ async fn me_health_monitor_converges_retry_like_threshold_backlog_to_empty() {
     }
 
     let monitor = tokio::spawn(me_health_monitor(pool.clone(), rng, 0));
-    tokio::time::sleep(Duration::from_millis(60)).await;
+    wait_for_pool_empty(&pool, Duration::from_secs(1)).await;
     monitor.abort();
     let _ = monitor.await;
 

src/transport/middle_proxy/health_regression_tests.rs

@@ -168,6 +168,21 @@ async fn current_writer_ids(pool: &Arc<MePool>) -> Vec<u64> {
     writer_ids
 }
 
+async fn writer_exists(pool: &Arc<MePool>, writer_id: u64) -> bool {
+    pool.writers
+        .read()
+        .await
+        .iter()
+        .any(|writer| writer.id == writer_id)
+}
+
+async fn set_writer_draining(pool: &Arc<MePool>, writer_id: u64, draining: bool) {
+    let writers = pool.writers.read().await;
+    if let Some(writer) = writers.iter().find(|writer| writer.id == writer_id) {
+        writer.draining.store(draining, Ordering::Relaxed);
+    }
+}
+
 #[tokio::test]
 async fn reap_draining_writers_drops_warn_state_for_removed_writer() {
     let pool = make_pool(128).await;
@@ -257,6 +272,123 @@ async fn reap_draining_writers_limits_closes_per_health_tick() {
     assert_eq!(pool.writers.read().await.len(), writer_total - close_budget);
 }
 
+#[tokio::test]
+async fn reap_draining_writers_keeps_warn_state_for_deadline_backlog_writers() {
+    let pool = make_pool(0).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let close_budget = health_drain_close_budget();
+    let writer_total = close_budget.saturating_add(5);
+    for writer_id in 1..=writer_total as u64 {
+        insert_draining_writer(
+            &pool,
+            writer_id,
+            now_epoch_secs.saturating_sub(60),
+            1,
+            now_epoch_secs.saturating_sub(1),
+        )
+        .await;
+    }
+    let target_writer_id = writer_total as u64;
+    let mut warn_next_allowed = HashMap::new();
+    warn_next_allowed.insert(
+        target_writer_id,
+        Instant::now() + Duration::from_secs(300),
+    );
+
+    reap_draining_writers(&pool, &mut warn_next_allowed).await;
+
+    assert!(writer_exists(&pool, target_writer_id).await);
+    assert!(warn_next_allowed.contains_key(&target_writer_id));
+}
+
+#[tokio::test]
+async fn reap_draining_writers_keeps_warn_state_for_overflow_backlog_writers() {
+    let pool = make_pool(1).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let close_budget = health_drain_close_budget();
+    let writer_total = close_budget.saturating_add(6);
+    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 target_writer_id = writer_total.saturating_sub(1) as u64;
+    let mut warn_next_allowed = HashMap::new();
+    warn_next_allowed.insert(
+        target_writer_id,
+        Instant::now() + Duration::from_secs(300),
+    );
+
+    reap_draining_writers(&pool, &mut warn_next_allowed).await;
+
+    assert!(writer_exists(&pool, target_writer_id).await);
+    assert!(warn_next_allowed.contains_key(&target_writer_id));
+}
+
+#[tokio::test]
+async fn reap_draining_writers_drops_warn_state_when_writer_exits_draining_state() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    insert_draining_writer(&pool, 71, now_epoch_secs.saturating_sub(60), 1, 0).await;
+
+    let mut warn_next_allowed = HashMap::new();
+    warn_next_allowed.insert(71, Instant::now() + Duration::from_secs(300));
+
+    set_writer_draining(&pool, 71, false).await;
+    reap_draining_writers(&pool, &mut warn_next_allowed).await;
+
+    assert!(writer_exists(&pool, 71).await);
+    assert!(
+        !warn_next_allowed.contains_key(&71),
+        "warn cooldown state must be dropped after writer leaves draining state"
+    );
+}
+
+#[tokio::test]
+async fn reap_draining_writers_preserves_warn_state_across_multiple_budget_deferrals() {
+    let pool = make_pool(0).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(1);
+    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 tail_writer_id = writer_total as u64;
+    let mut warn_next_allowed = HashMap::new();
+    warn_next_allowed.insert(
+        tail_writer_id,
+        Instant::now() + Duration::from_secs(300),
+    );
+
+    reap_draining_writers(&pool, &mut warn_next_allowed).await;
+    assert!(writer_exists(&pool, tail_writer_id).await);
+    assert!(warn_next_allowed.contains_key(&tail_writer_id));
+
+    reap_draining_writers(&pool, &mut warn_next_allowed).await;
+    assert!(writer_exists(&pool, tail_writer_id).await);
+    assert!(warn_next_allowed.contains_key(&tail_writer_id));
+
+    reap_draining_writers(&pool, &mut warn_next_allowed).await;
+    assert!(!writer_exists(&pool, tail_writer_id).await);
+    assert!(
+        !warn_next_allowed.contains_key(&tail_writer_id),
+        "warn cooldown state must clear once writer is actually removed"
+    );
+}
+
 #[tokio::test]
 async fn reap_draining_writers_backlog_drains_across_ticks() {
     let pool = make_pool(128).await;
@@ -460,3 +592,67 @@ async fn reap_draining_writers_mixed_backlog_converges_without_leaking_warn_stat
 fn general_config_default_drain_threshold_remains_enabled() {
     assert_eq!(GeneralConfig::default().me_pool_drain_threshold, 128);
 }
+
+#[tokio::test]
+async fn reap_draining_writers_does_not_close_writer_that_became_non_empty_after_snapshot() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+
+    let empty_writer_id = 700u64;
+    insert_draining_writer(
+        &pool,
+        empty_writer_id,
+        now_epoch_secs.saturating_sub(60),
+        0,
+        0,
+    )
+    .await;
+
+    let stale_empty_snapshot = vec![empty_writer_id];
+    let (rebound_conn_id, _rx) = pool.registry.register().await;
+    assert!(
+        pool.registry
+            .bind_writer(
+                rebound_conn_id,
+                empty_writer_id,
+                ConnMeta {
+                    target_dc: 2,
+                    client_addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 9050),
+                    our_addr: SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443),
+                    proto_flags: 0,
+                },
+            )
+            .await,
+        "writer should accept a new bind after stale empty snapshot"
+    );
+
+    for writer_id in stale_empty_snapshot {
+        assert!(
+            !pool.remove_writer_if_empty(writer_id).await,
+            "atomic empty cleanup must reject writers that gained bound clients"
+        );
+    }
+
+    assert!(
+        writer_exists(&pool, empty_writer_id).await,
+        "empty-path cleanup must not remove a writer that gained a bound client"
+    );
+    assert_eq!(
+        pool.registry.get_writer(rebound_conn_id).await.map(|w| w.writer_id),
+        Some(empty_writer_id)
+    );
+
+    let _ = pool.registry.unregister(rebound_conn_id).await;
+}
+
+#[tokio::test]
+async fn prune_closed_writers_closes_bound_clients_when_writer_is_non_empty() {
+    let pool = make_pool(128).await;
+    let now_epoch_secs = MePool::now_epoch_secs();
+    let conn_ids = insert_draining_writer(&pool, 910, now_epoch_secs.saturating_sub(60), 1, 0).await;
+
+    pool.prune_closed_writers().await;
+
+    assert!(!writer_exists(&pool, 910).await);
+    assert!(pool.registry.get_writer(conn_ids[0]).await.is_none());
+}

src/transport/middle_proxy/mod.rs

@@ -27,6 +27,8 @@ mod health_regression_tests;
 mod health_integration_tests;
 #[cfg(test)]
 mod health_adversarial_tests;
+#[cfg(test)]
+mod send_adversarial_tests;
 
 use bytes::Bytes;
 

src/transport/middle_proxy/pool.rs

@@ -160,6 +160,7 @@ pub struct MePool {
     pub(super) refill_inflight: Arc<Mutex<HashSet<RefillEndpointKey>>>,
     pub(super) refill_inflight_dc: Arc<Mutex<HashSet<RefillDcKey>>>,
     pub(super) conn_count: AtomicUsize,
+    pub(super) draining_active_runtime: AtomicU64,
     pub(super) stats: Arc<crate::stats::Stats>,
     pub(super) generation: AtomicU64,
     pub(super) active_generation: AtomicU64,
@@ -438,6 +439,7 @@ impl MePool {
             refill_inflight: Arc::new(Mutex::new(HashSet::new())),
             refill_inflight_dc: Arc::new(Mutex::new(HashSet::new())),
             conn_count: AtomicUsize::new(0),
+            draining_active_runtime: AtomicU64::new(0),
             generation: AtomicU64::new(1),
             active_generation: AtomicU64::new(1),
             warm_generation: AtomicU64::new(0),
@@ -690,6 +692,33 @@ impl MePool {
         }
     }
 
+    #[allow(dead_code)]
+    pub(super) fn draining_active_runtime(&self) -> u64 {
+        self.draining_active_runtime.load(Ordering::Relaxed)
+    }
+
+    pub(super) fn increment_draining_active_runtime(&self) {
+        self.draining_active_runtime.fetch_add(1, Ordering::Relaxed);
+    }
+
+    pub(super) fn decrement_draining_active_runtime(&self) {
+        let mut current = self.draining_active_runtime.load(Ordering::Relaxed);
+        loop {
+            if current == 0 {
+                break;
+            }
+            match self.draining_active_runtime.compare_exchange_weak(
+                current,
+                current - 1,
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => break,
+                Err(actual) => current = actual,
+            }
+        }
+    }
+
     pub(super) async fn key_selector(&self) -> u32 {
         self.proxy_secret.read().await.key_selector
     }

src/transport/middle_proxy/pool_reinit.rs

@@ -141,6 +141,38 @@ impl MePool {
         out
     }
 
+    pub(super) async fn has_non_draining_writer_per_desired_dc_group(&self) -> bool {
+        let desired_by_dc = self.desired_dc_endpoints().await;
+        let required_dcs: HashSet<i32> = desired_by_dc
+            .iter()
+            .filter_map(|(dc, endpoints)| {
+                if endpoints.is_empty() {
+                    None
+                } else {
+                    Some(*dc)
+                }
+            })
+            .collect();
+        if required_dcs.is_empty() {
+            return true;
+        }
+
+        let ws = self.writers.read().await;
+        let mut covered_dcs = HashSet::<i32>::with_capacity(required_dcs.len());
+        for writer in ws.iter() {
+            if writer.draining.load(Ordering::Relaxed) {
+                continue;
+            }
+            if required_dcs.contains(&writer.writer_dc) {
+                covered_dcs.insert(writer.writer_dc);
+                if covered_dcs.len() == required_dcs.len() {
+                    return true;
+                }
+            }
+        }
+        false
+    }
+
     fn hardswap_warmup_connect_delay_ms(&self) -> u64 {
         let min_ms = self.me_hardswap_warmup_delay_min_ms.load(Ordering::Relaxed);
         let max_ms = self.me_hardswap_warmup_delay_max_ms.load(Ordering::Relaxed);
@@ -475,12 +507,30 @@ impl MePool {
             coverage_ratio = format_args!("{coverage_ratio:.3}"),
             min_ratio = format_args!("{min_ratio:.3}"),
             drain_timeout_secs,
-            "ME reinit cycle covered; draining stale writers"
+            "ME reinit cycle covered; processing stale writers"
         );
         self.stats.increment_pool_swap_total();
+        let can_drop_with_replacement = self
+            .has_non_draining_writer_per_desired_dc_group()
+            .await;
+        if can_drop_with_replacement {
+            info!(
+                stale_writers = stale_writer_ids.len(),
+                "ME reinit stale writers: replacement coverage ready, force-closing clients for fast rebind"
+            );
+        } else {
+            warn!(
+                stale_writers = stale_writer_ids.len(),
+                "ME reinit stale writers: replacement coverage incomplete, keeping draining fallback"
+            );
+        }
         for writer_id in stale_writer_ids {
             self.mark_writer_draining_with_timeout(writer_id, drain_timeout, !hardswap)
                 .await;
+            if can_drop_with_replacement {
+                self.stats.increment_pool_force_close_total();
+                self.remove_writer_and_close_clients(writer_id).await;
+            }
         }
         if hardswap {
             self.clear_pending_hardswap_state();

src/transport/middle_proxy/pool_writer.rs

@@ -42,11 +42,10 @@ impl MePool {
         }
 
         for writer_id in closed_writer_ids {
-            if self.registry.is_writer_empty(writer_id).await {
-                let _ = self.remove_writer_only(writer_id).await;
-            } else {
-                let _ = self.remove_writer_and_close_clients(writer_id).await;
+            if self.remove_writer_if_empty(writer_id).await {
+                continue;
             }
+            let _ = self.remove_writer_and_close_clients(writer_id).await;
         }
     }
 
@@ -501,6 +500,17 @@ impl MePool {
         }
     }
 
+    pub(crate) async fn remove_writer_if_empty(self: &Arc<Self>, writer_id: u64) -> bool {
+        if !self.registry.unregister_writer_if_empty(writer_id).await {
+            return false;
+        }
+
+        // The registry empty-check and unregister are atomic with respect to binds,
+        // so remove_writer_only cannot return active bound sessions here.
+        let _ = self.remove_writer_only(writer_id).await;
+        true
+    }
+
     async fn remove_writer_only(self: &Arc<Self>, writer_id: u64) -> Vec<BoundConn> {
         let mut close_tx: Option<mpsc::Sender<WriterCommand>> = None;
         let mut removed_addr: Option<SocketAddr> = None;
@@ -514,6 +524,7 @@ impl MePool {
                 let was_draining = w.draining.load(Ordering::Relaxed);
                 if was_draining {
                     self.stats.decrement_pool_drain_active();
+                    self.decrement_draining_active_runtime();
                 }
                 self.stats.increment_me_writer_removed_total();
                 w.cancel.cancel();
@@ -572,6 +583,7 @@ impl MePool {
                     .store(drain_deadline_epoch_secs, Ordering::Relaxed);
                 if !already_draining {
                     self.stats.increment_pool_drain_active();
+                    self.increment_draining_active_runtime();
                 }
                 w.contour
                     .store(WriterContour::Draining.as_u8(), Ordering::Relaxed);

src/transport/middle_proxy/registry.rs

@@ -436,6 +436,37 @@ impl ConnRegistry {
             .map(|s| s.is_empty())
             .unwrap_or(true)
     }
+
+    pub async fn unregister_writer_if_empty(&self, writer_id: u64) -> bool {
+        let mut inner = self.inner.write().await;
+        let Some(conn_ids) = inner.conns_for_writer.get(&writer_id) else {
+            // Writer is already absent from the registry.
+            return true;
+        };
+        if !conn_ids.is_empty() {
+            return false;
+        }
+
+        inner.writers.remove(&writer_id);
+        inner.last_meta_for_writer.remove(&writer_id);
+        inner.writer_idle_since_epoch_secs.remove(&writer_id);
+        inner.conns_for_writer.remove(&writer_id);
+        true
+    }
+
+    #[allow(dead_code)]
+    pub(super) async fn non_empty_writer_ids(&self, writer_ids: &[u64]) -> HashSet<u64> {
+        let inner = self.inner.read().await;
+        let mut out = HashSet::<u64>::with_capacity(writer_ids.len());
+        for writer_id in writer_ids {
+            if let Some(conns) = inner.conns_for_writer.get(writer_id)
+                && !conns.is_empty()
+            {
+                out.insert(*writer_id);
+            }
+        }
+        out
+    }
 }
 
 #[cfg(test)]
@@ -634,4 +665,35 @@ mod tests {
         );
         assert!(registry.get_writer(conn_id).await.is_none());
     }
+
+    #[tokio::test]
+    async fn non_empty_writer_ids_returns_only_writers_with_bound_clients() {
+        let registry = ConnRegistry::new();
+        let (conn_id, _rx) = registry.register().await;
+        let (writer_tx_a, _writer_rx_a) = tokio::sync::mpsc::channel(8);
+        let (writer_tx_b, _writer_rx_b) = tokio::sync::mpsc::channel(8);
+        registry.register_writer(10, writer_tx_a).await;
+        registry.register_writer(20, writer_tx_b).await;
+
+        let addr = SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443);
+        assert!(
+            registry
+                .bind_writer(
+                    conn_id,
+                    10,
+                    ConnMeta {
+                        target_dc: 2,
+                        client_addr: addr,
+                        our_addr: addr,
+                        proto_flags: 0,
+                    },
+                )
+                .await
+        );
+
+        let non_empty = registry.non_empty_writer_ids(&[10, 20, 30]).await;
+        assert!(non_empty.contains(&10));
+        assert!(!non_empty.contains(&20));
+        assert!(!non_empty.contains(&30));
+    }
 }

src/transport/middle_proxy/send.rs

@@ -372,17 +372,20 @@ impl MePool {
                 }
                 let effective_our_addr = SocketAddr::new(w.source_ip, our_addr.port());
                 let (payload, meta) = build_routed_payload(effective_our_addr);
-                match w.tx.try_send(WriterCommand::Data(payload.clone())) {
-                    Ok(()) => {
-                        self.stats.increment_me_writer_pick_success_try_total(pick_mode);
+                match w.tx.clone().try_reserve_owned() {
+                    Ok(permit) => {
                         if !self.registry.bind_writer(conn_id, w.id, meta).await {
                             debug!(
                                 conn_id,
                                 writer_id = w.id,
-                                "ME writer disappeared before bind commit, retrying"
+                                "ME writer disappeared before bind commit, pruning stale writer"
                             );
+                            drop(permit);
+                            self.remove_writer_and_close_clients(w.id).await;
                             continue;
                         }
+                        permit.send(WriterCommand::Data(payload.clone()));
+                        self.stats.increment_me_writer_pick_success_try_total(pick_mode);
                         if w.generation < self.current_generation() {
                             self.stats.increment_pool_stale_pick_total();
                             debug!(
@@ -422,18 +425,21 @@ impl MePool {
             self.stats.increment_me_writer_pick_blocking_fallback_total();
             let effective_our_addr = SocketAddr::new(w.source_ip, our_addr.port());
             let (payload, meta) = build_routed_payload(effective_our_addr);
-            match w.tx.send(WriterCommand::Data(payload.clone())).await {
-                Ok(()) => {
-                    self.stats
-                        .increment_me_writer_pick_success_fallback_total(pick_mode);
+            match w.tx.clone().reserve_owned().await {
+                Ok(permit) => {
                     if !self.registry.bind_writer(conn_id, w.id, meta).await {
                         debug!(
                             conn_id,
                             writer_id = w.id,
-                            "ME writer disappeared before fallback bind commit, retrying"
+                            "ME writer disappeared before fallback bind commit, pruning stale writer"
                         );
+                        drop(permit);
+                        self.remove_writer_and_close_clients(w.id).await;
                         continue;
                     }
+                    permit.send(WriterCommand::Data(payload.clone()));
+                    self.stats
+                        .increment_me_writer_pick_success_fallback_total(pick_mode);
                     if w.generation < self.current_generation() {
                         self.stats.increment_pool_stale_pick_total();
                     }

src/transport/middle_proxy/send_adversarial_tests.rs

@@ -0,0 +1,263 @@
+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::pool::{MePool, MeWriter, WriterContour};
+use crate::config::{GeneralConfig, MeRouteNoWriterMode, MeSocksKdfPolicy, MeWriterPickMode};
+use crate::crypto::SecureRandom;
+use crate::network::probe::NetworkDecision;
+use crate::stats::Stats;
+
+async fn make_pool() -> (Arc<MePool>, Arc<SecureRandom>) {
+    let general = GeneralConfig {
+        me_route_no_writer_mode: MeRouteNoWriterMode::AsyncRecoveryFailfast,
+        me_route_no_writer_wait_ms: 50,
+        me_writer_pick_mode: MeWriterPickMode::SortedRr,
+        me_deterministic_writer_sort: true,
+        ..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.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,
+        general.me_writer_pick_mode,
+        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,
+        general.me_route_no_writer_mode,
+        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_writer(
+    pool: &Arc<MePool>,
+    writer_id: u64,
+    writer_dc: i32,
+    addr: SocketAddr,
+    register_in_registry: bool,
+) -> mpsc::Receiver<WriterCommand> {
+    let (tx, rx) = mpsc::channel::<WriterCommand>(8);
+    let writer = MeWriter {
+        id: writer_id,
+        addr,
+        source_ip: addr.ip(),
+        writer_dc,
+        generation: pool.current_generation(),
+        contour: Arc::new(AtomicU8::new(WriterContour::Active.as_u8())),
+        created_at: Instant::now(),
+        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(false)),
+        draining_started_at_epoch_secs: Arc::new(AtomicU64::new(0)),
+        drain_deadline_epoch_secs: Arc::new(AtomicU64::new(0)),
+        allow_drain_fallback: Arc::new(AtomicBool::new(false)),
+    };
+
+    pool.writers.write().await.push(writer);
+    {
+        let mut map = pool.proxy_map_v4.write().await;
+        map.entry(writer_dc)
+            .or_insert_with(Vec::new)
+            .push((addr.ip(), addr.port()));
+    }
+    pool.rebuild_endpoint_dc_map().await;
+    if register_in_registry {
+        pool.registry.register_writer(writer_id, tx).await;
+    }
+    rx
+}
+
+async fn recv_data_count(rx: &mut mpsc::Receiver<WriterCommand>, budget: Duration) -> usize {
+    let start = Instant::now();
+    let mut data_count = 0usize;
+    while Instant::now().duration_since(start) < budget {
+        let remaining = budget.saturating_sub(Instant::now().duration_since(start));
+        match tokio::time::timeout(remaining.min(Duration::from_millis(10)), rx.recv()).await {
+            Ok(Some(WriterCommand::Data(_))) => data_count += 1,
+            Ok(Some(WriterCommand::DataAndFlush(_))) => data_count += 1,
+            Ok(Some(WriterCommand::Close)) => {}
+            Ok(None) => break,
+            Err(_) => break,
+        }
+    }
+    data_count
+}
+
+#[tokio::test]
+async fn send_proxy_req_does_not_replay_when_first_bind_commit_fails() {
+    let (pool, _rng) = make_pool().await;
+    pool.rr.store(0, Ordering::Relaxed);
+
+    let (conn_id, _rx) = pool.registry.register().await;
+    let mut stale_rx = insert_writer(
+        &pool,
+        10,
+        2,
+        SocketAddr::new(IpAddr::V4(Ipv4Addr::new(127, 0, 0, 10)), 443),
+        false,
+    )
+    .await;
+    let mut live_rx = insert_writer(
+        &pool,
+        11,
+        2,
+        SocketAddr::new(IpAddr::V4(Ipv4Addr::new(127, 0, 0, 11)), 443),
+        true,
+    )
+    .await;
+
+    let result = pool
+        .send_proxy_req(
+            conn_id,
+            2,
+            SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 30000),
+            SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443),
+            b"hello",
+            0,
+            None,
+        )
+        .await;
+
+    assert!(result.is_ok());
+    assert_eq!(recv_data_count(&mut stale_rx, Duration::from_millis(50)).await, 0);
+    assert_eq!(recv_data_count(&mut live_rx, Duration::from_millis(50)).await, 1);
+
+    let bound = pool.registry.get_writer(conn_id).await;
+    assert!(bound.is_some());
+    assert_eq!(bound.expect("writer should be bound").writer_id, 11);
+}
+
+#[tokio::test]
+async fn send_proxy_req_prunes_iterative_stale_bind_failures_without_data_replay() {
+    let (pool, _rng) = make_pool().await;
+    pool.rr.store(0, Ordering::Relaxed);
+
+    let (conn_id, _rx) = pool.registry.register().await;
+
+    let mut stale_rx_1 = insert_writer(
+        &pool,
+        21,
+        2,
+        SocketAddr::new(IpAddr::V4(Ipv4Addr::new(127, 0, 1, 21)), 443),
+        false,
+    )
+    .await;
+    let mut stale_rx_2 = insert_writer(
+        &pool,
+        22,
+        2,
+        SocketAddr::new(IpAddr::V4(Ipv4Addr::new(127, 0, 1, 22)), 443),
+        false,
+    )
+    .await;
+    let mut live_rx = insert_writer(
+        &pool,
+        23,
+        2,
+        SocketAddr::new(IpAddr::V4(Ipv4Addr::new(127, 0, 1, 23)), 443),
+        true,
+    )
+    .await;
+
+    let result = pool
+        .send_proxy_req(
+            conn_id,
+            2,
+            SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 30001),
+            SocketAddr::new(IpAddr::V4(Ipv4Addr::LOCALHOST), 443),
+            b"storm",
+            0,
+            None,
+        )
+        .await;
+
+    assert!(result.is_ok());
+    assert_eq!(recv_data_count(&mut stale_rx_1, Duration::from_millis(50)).await, 0);
+    assert_eq!(recv_data_count(&mut stale_rx_2, Duration::from_millis(50)).await, 0);
+    assert_eq!(recv_data_count(&mut live_rx, Duration::from_millis(50)).await, 1);
+
+    let writers = pool.writers.read().await;
+    let writer_ids = writers.iter().map(|w| w.id).collect::<Vec<_>>();
+    drop(writers);
+    assert_eq!(writer_ids, vec![23]);
+}