merge master: reconcile the two direct install.sh hotfixes

Both were emergency live-fixes to master (since install.sh/install.ps1 are
curl'd straight from that branch's raw URL, bypassing the normal release
promotion): the `ostp setup` subcommand fix and the alpha/beta self-update
mechanism fix. alpha already has equivalent content for both via its own
separate commits, so this is a pure reconciliation.

.release-state.json

@@ -2,5 +2,5 @@
   "target_version": "0.4.2",
   "branch": "beta",
   "alpha_iteration": 0,
-  "beta_iteration": 2
+  "beta_iteration": 4
 }

Cargo.lock

@@ -1400,6 +1400,7 @@ dependencies = [
  "rlimit",
  "serde",
  "serde_json",
+ "sha2",
  "tokio",
  "tracing",
  "tracing-subscriber",

ostp-client/src/bridge.rs

@@ -231,7 +231,7 @@ impl Bridge {
                     self.handle_inbound_udp(udp_msg, &mut sessions_opt, &mut udp_rx_opt, &mut proxy_guard, &mut stream_map, &tx, &proxy_tx).await;
                 }
                 cmd = bridge_rx.recv() => {
-                    if !self.handle_bridge_cmd(cmd, &mut sessions_opt, &mut udp_rx_opt, &mut proxy_guard, &mut stream_map, &tx, &proxy_tx).await {
+                    if !self.handle_bridge_cmd(cmd, &mut bridge_rx, &mut sessions_opt, &mut udp_rx_opt, &mut proxy_guard, &mut stream_map, &tx, &proxy_tx).await {
                         break;
                     }
                 }
@@ -374,6 +374,7 @@ impl Bridge {
     async fn handle_bridge_cmd(
         &mut self,
         cmd: Option<BridgeCommand>,
+        bridge_rx: &mut mpsc::Receiver<BridgeCommand>,
         sessions_opt: &mut Option<Vec<SessionState>>,
         udp_rx_opt: &mut Option<mpsc::Receiver<(usize, Bytes)>>,
         proxy_guard: &mut Option<crate::sysproxy::SystemProxyGuard>,
@@ -465,6 +466,32 @@ impl Bridge {
                 tx.send(UiEvent::Log(format!("Obfuscation profile switched to {:?}", self.profile))).await.ok();
             }
             Some(BridgeCommand::NetworkChanged) => {
+                // A real network handoff (Wi-Fi <-> cellular) commonly fires
+                // onLost + onAvailable within milliseconds of each other on
+                // Android, queuing several NetworkChanged commands back to
+                // back. Each reconnect below is a full sequential handshake
+                // (up to ~1.2s x 4 attempts x mux_sessions) run synchronously
+                // in this select-loop iteration, so without coalescing, the
+                // first attempt often races the OS's own network switch and
+                // fails on the now-dead interface, then the SECOND queued
+                // NetworkChanged only starts its own full reconnect after
+                // that first one finishes - multiplying a sub-second handoff
+                // into many seconds of extra outage. Drain same-kind repeats
+                // so a burst collapses into one reconnect on the freshest
+                // signal; a different command found while draining is
+                // handled immediately rather than dropped.
+                while let Ok(next) = bridge_rx.try_recv() {
+                    if !matches!(next, BridgeCommand::NetworkChanged) {
+                        let more = Box::pin(self.handle_bridge_cmd(
+                            Some(next), bridge_rx, sessions_opt, udp_rx_opt, proxy_guard, stream_map, tx, proxy_tx,
+                        )).await;
+                        if !more {
+                            return false;
+                        }
+                        break;
+                    }
+                }
+
                 if self.running {
                     let _ = tx.send(UiEvent::Log("Network changed — starting immediate reconnect".to_string())).await;
                     self.metrics.connection_state.store(1, Ordering::Relaxed);

ostp-client/src/tunnel/proxy.rs

@@ -361,6 +361,10 @@ async fn handle_udp_associate(
 
     let mut direct_udp_v4: Option<Arc<UdpSocket>> = None;
     let mut direct_udp_v6: Option<Arc<UdpSocket>> = None;
+    // Held only to keep the direct-UDP readers' cancellation senders alive;
+    // dropping this (on every return path from this function) is what tells
+    // spawn_direct_udp_reader's tasks to stop. See its doc comment.
+    let mut direct_udp_cancel_txs: Vec<tokio::sync::oneshot::Sender<()>> = Vec::new();
 
     let mut tcp_buf = [0u8; 1];
     loop {
@@ -432,7 +436,9 @@ async fn handle_udp_associate(
                                     match create_udp_socket_bypassing_tun(true, matcher.physical_if_index, &matcher.physical_if_name).await {
                                         Ok(s) => {
                                             let s_arc = Arc::new(s);
-                                            spawn_direct_udp_reader(s_arc.clone(), sock_tx.clone(), client_udp_addr.clone(), debug);
+                                            let (cancel_tx, cancel_rx) = tokio::sync::oneshot::channel();
+                                            spawn_direct_udp_reader(s_arc.clone(), sock_tx.clone(), client_udp_addr.clone(), debug, cancel_rx);
+                                            direct_udp_cancel_txs.push(cancel_tx);
                                             direct_udp_v6 = Some(s_arc);
                                         }
                                         Err(e) => {
@@ -446,7 +452,9 @@ async fn handle_udp_associate(
                                     match create_udp_socket_bypassing_tun(false, matcher.physical_if_index, &matcher.physical_if_name).await {
                                         Ok(s) => {
                                             let s_arc = Arc::new(s);
-                                            spawn_direct_udp_reader(s_arc.clone(), sock_tx.clone(), client_udp_addr.clone(), debug);
+                                            let (cancel_tx, cancel_rx) = tokio::sync::oneshot::channel();
+                                            spawn_direct_udp_reader(s_arc.clone(), sock_tx.clone(), client_udp_addr.clone(), debug, cancel_rx);
+                                            direct_udp_cancel_txs.push(cancel_tx);
                                             direct_udp_v4 = Some(s_arc);
                                         }
                                         Err(e) => {
@@ -520,11 +528,24 @@ fn spawn_direct_udp_reader(
     sock_tx: Arc<UdpSocket>,
     client_udp_addr: Arc<std::sync::Mutex<Option<std::net::SocketAddr>>>,
     _debug: bool,
+    mut cancel_rx: tokio::sync::oneshot::Receiver<()>,
 ) {
     tokio::spawn(async move {
         let mut buf = vec![0u8; 65536];
         loop {
-            match direct_socket.recv_from(&mut buf).await {
+            let recv_result = tokio::select! {
+                // Fires as soon as the sender half (held by handle_udp_associate
+                // for exactly this reason) is dropped - which happens the
+                // instant that function returns, on every exit path, with no
+                // explicit signaling needed. Without this, a UDP-associate
+                // session that ever bypassed traffic direct (excluded IP/
+                // domain) leaked this socket + task for the rest of the
+                // process's life once the session ended: nothing else ever
+                // stopped this loop.
+                _ = &mut cancel_rx => break,
+                res = direct_socket.recv_from(&mut buf) => res,
+            };
+            match recv_result {
                 Ok((len, target_addr)) => {
                     let client_addr = {
                         let guard = client_udp_addr.lock().unwrap();

ostp-core/src/congestion.rs

@@ -43,6 +43,11 @@ pub struct CongestionController {
     mtu: u64,
     /// Min RTT expiry: re-probe after 10 seconds
     min_rtt_stamp: Instant,
+    /// Loss events counted toward SLOW_START_LOSS_TOLERANCE within the
+    /// current SLOW_START_LOSS_WINDOW (see on_loss's SlowStart arm).
+    slow_start_losses: u32,
+    /// Start of the current loss-tolerance window.
+    slow_start_loss_window_start: Instant,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -67,6 +72,24 @@ const RTO_MAX: Duration = Duration::from_secs(16);
 /// Will be replaced by first real measurement within milliseconds.
 const INITIAL_RTT: Duration = Duration::from_millis(30);
 
+/// Isolated packet loss during slow start (a single dropped frame from
+/// wireless noise, a brief LTE handover blip, etc.) is normal on real
+/// mobile/Wi-Fi links and does NOT mean the link is congested. The previous
+/// behavior exited slow start and halved cwnd on the very FIRST loss, which
+/// on any link with a non-zero background loss rate permanently downgrades
+/// the session from exponential growth to linear (+1 MTU/RTT) ProbeBandwidth
+/// growth within the first few RTTs - turning what should be a sub-second
+/// ramp-up into tens of seconds to minutes before throughput opens up
+/// (observed as: a trickle of KB/s, then a sudden jump once cwnd finally
+/// claws back up). Only treat loss as a real congestion signal - and pay
+/// the full slow-start-exit + halving cost - once this many losses land
+/// within SLOW_START_LOSS_WINDOW.
+const SLOW_START_LOSS_TOLERANCE: u32 = 3;
+/// Window within which SLOW_START_LOSS_TOLERANCE losses must land to count
+/// as sustained (rather than isolated) loss. Roughly a few RTTs on a
+/// well-connected link, generous on a slow one.
+const SLOW_START_LOSS_WINDOW: Duration = Duration::from_millis(500);
+
 impl CongestionController {
     pub fn new(mtu: u64) -> Self {
         let now = Instant::now();
@@ -88,6 +111,8 @@ impl CongestionController {
             pacing_rate: initial_pacing,
             mtu,
             min_rtt_stamp: now,
+            slow_start_losses: 0,
+            slow_start_loss_window_start: now,
         }
     }
 
@@ -197,11 +222,28 @@ impl CongestionController {
 
         match self.phase {
             Phase::SlowStart => {
-                // Exit slow start, set ssthresh to half of cwnd
-                self.ssthresh = self.cwnd / 2;
-                self.cwnd = self.ssthresh.max(MIN_CWND_PACKETS * self.mtu);
-                self.phase = Phase::ProbeBandwidth;
-                tracing::debug!(cwnd = self.cwnd, ssthresh = self.ssthresh, "congestion: loss during slow start");
+                let now = Instant::now();
+                if now.duration_since(self.slow_start_loss_window_start) > SLOW_START_LOSS_WINDOW {
+                    // Previous window's losses have aged out - this loss starts a fresh count.
+                    self.slow_start_losses = 0;
+                    self.slow_start_loss_window_start = now;
+                }
+                self.slow_start_losses += 1;
+
+                if self.slow_start_losses >= SLOW_START_LOSS_TOLERANCE {
+                    // Sustained loss within the window: treat as real congestion.
+                    // Exit slow start, set ssthresh to half of cwnd.
+                    self.ssthresh = self.cwnd / 2;
+                    self.cwnd = self.ssthresh.max(MIN_CWND_PACKETS * self.mtu);
+                    self.phase = Phase::ProbeBandwidth;
+                    tracing::debug!(cwnd = self.cwnd, ssthresh = self.ssthresh, "congestion: sustained loss during slow start, exiting");
+                } else {
+                    // Isolated loss: likely non-congestive noise. Take a mild,
+                    // temporary haircut but keep exponential growth going -
+                    // don't throw away slow start over a single dropped frame.
+                    self.cwnd = (self.cwnd * 8 / 10).max(MIN_CWND_PACKETS * self.mtu);
+                    tracing::debug!(cwnd = self.cwnd, count = self.slow_start_losses, "congestion: isolated loss during slow start, staying in slow start");
+                }
             }
             Phase::ProbeBandwidth => {
                 // Multiplicative decrease: cwnd *= 0.7 (BBR-style, less aggressive than Cubic's 0.5)
@@ -290,6 +332,50 @@ mod tests {
         assert!(cc.cwnd() < initial);
     }
 
+    #[test]
+    fn test_isolated_slow_start_loss_does_not_exit_slow_start() {
+        // A single dropped packet (wireless noise, a brief handover blip) is
+        // normal on real links and must not permanently downgrade the
+        // session from exponential to linear growth.
+        let mut cc = CongestionController::new(1200);
+        cc.on_loss(1200);
+        assert_eq!(cc.phase, Phase::SlowStart, "one isolated loss must not exit slow start");
+
+        // It should still shrink the window somewhat (not ignored entirely),
+        // just far less punishing than the sustained-congestion case.
+        let after_one = cc.cwnd();
+        assert!(after_one < INITIAL_CWND_PACKETS * 1200);
+    }
+
+    #[test]
+    fn test_sustained_slow_start_loss_exits_slow_start() {
+        // Losses landing close together (within SLOW_START_LOSS_WINDOW) are
+        // a real congestion signal and must still trigger the harsher
+        // exit-slow-start + halve response.
+        let mut cc = CongestionController::new(1200);
+        for _ in 0..SLOW_START_LOSS_TOLERANCE {
+            cc.on_loss(1200);
+        }
+        assert_eq!(cc.phase, Phase::ProbeBandwidth, "sustained loss must exit slow start");
+    }
+
+    #[test]
+    fn test_slow_start_loss_window_resets_after_expiry() {
+        // Two losses far enough apart (window expired between them) must
+        // each be treated as isolated, not accumulated toward the sustained-
+        // loss threshold.
+        let mut cc = CongestionController::new(1200);
+        cc.on_loss(1200);
+        assert_eq!(cc.phase, Phase::SlowStart);
+
+        // Simulate the window having expired by resetting its start
+        // directly (std::thread::sleep in a unit test would be flaky/slow).
+        cc.slow_start_loss_window_start = Instant::now() - SLOW_START_LOSS_WINDOW - Duration::from_millis(1);
+        cc.on_loss(1200);
+        assert_eq!(cc.phase, Phase::SlowStart, "a loss after the window expired must restart the count, not accumulate");
+        assert_eq!(cc.slow_start_losses, 1);
+    }
+
     #[test]
     fn test_can_send_limits() {
         let mut cc = CongestionController::new(1200);

ostp-flutter/pubspec.yaml

@@ -16,7 +16,7 @@ publish_to: 'none' # Remove this line if you wish to publish to pub.dev
 # https://developer.apple.com/library/archive/documentation/General/Reference/InfoPlistKeyReference/Articles/CoreFoundationKeys.html
 # In Windows, build-name is used as the major, minor, and patch parts
 # of the product and file versions while build-number is used as the build suffix.
-version: 0.4.2+21
+version: 0.4.2+23
 
 environment:
   sdk: ^3.11.4

ostp/src/main.rs

@@ -3,6 +3,7 @@ use clap::Parser;
 use std::fs;
 use std::path::PathBuf;
 use colored::Colorize;
+use sha2::Digest;
 
 #[derive(Parser, Debug)]
 #[command(author, version, about = "OSTP Core - Ospab Stealth Transport Protocol", long_about = None)]

scripts/install.sh

@@ -115,12 +115,18 @@ if [ -n "$TARGET_VERSION" ]; then
     fi
     echo "Fetching requested release $LATEST_RELEASE..."
 else
-    if [ "$TARGET_BRANCH" == "alpha" ]; then
-        echo "Fetching alpha release..."
-        LATEST_RELEASE="alpha"
-    elif [ "$TARGET_BRANCH" == "beta" ]; then
-        echo "Fetching beta release..."
-        LATEST_RELEASE="beta"
+    if [ "$TARGET_BRANCH" == "alpha" ] || [ "$TARGET_BRANCH" == "beta" ]; then
+        # There is no floating "alpha"/"beta" GitHub Release - gha.ps1 cuts a
+        # fresh versioned tag every time (v0.4.2-beta.4, v0.4.2-alpha.7, ...).
+        # /releases/latest only ever returns the newest NON-prerelease
+        # (stable) tag, so it can't find these. Query the full releases list
+        # (newest first) and take the first tag_name containing "-$TARGET_BRANCH".
+        echo "Fetching latest ${TARGET_BRANCH} release..."
+        LATEST_RELEASE=$(curl -s "https://api.github.com/repos/${GITHUB_REPO}/releases" \
+            | grep '"tag_name":' \
+            | grep -- "-${TARGET_BRANCH}" \
+            | head -1 \
+            | sed -E 's/.*"tag_name": *"([^"]+)".*/\1/')
     else
         echo "Fetching latest stable release..."
         LATEST_RELEASE=$(curl -s "https://api.github.com/repos/${GITHUB_REPO}/releases/latest" | grep '"tag_name":' | sed -E 's/.*"([^"]+)".*/\1/')