fix(ble): reframe inbound L2CAP stream so packets survive non-SeqPacket backends

BLE delivery was unreliable on every backend except BlueZ. The receive
path assumed one `recv()` returned exactly one whole FIPS packet, which
only holds for BlueZ's SOCK_SEQPACKET (boundary-preserving). Android's
BluetoothSocket input stream and macOS CoreBluetooth are byte-stream
oriented: a read can return a fragment of a packet or several packets
coalesced. Under the old loop a fragment was shipped up as a runt packet
(rejected by FMP/Noise) and a coalesced tail was silently truncated and
dropped — so packets were lost and the transport thrashed.

Recover packet boundaries from the byte stream instead of trusting the
OS to preserve them. FIPS packets are self-delimiting via the 4-byte FMP
common prefix, so this reuses the exact length-prefixed framer TCP already
uses (`tcp::stream::read_fmp_packet`) — kept transport-agnostic for this
reason. A new `BleStreamRead` adapter turns the datagram-shaped `BleStream`
into the `AsyncRead` that framer expects, buffering leftover bytes across
reads. The recv future owns its scratch buffer and returns an owned Vec so
it is `'static` and storable across `poll_read` calls.

One reader is threaded through the pubkey exchange and the receive loop per
connection, so bytes a peer coalesces after the 33-byte pubkey stay buffered
rather than being lost at the handoff. The pubkey exchange now reads via
`read_exact` (reassembles a fragmented pubkey) instead of a brittle single
`recv()` with an exact-length check.

On BlueZ this is a transparent pass-through (one recv already equals one
packet); on stream backends it reassembles. Either way the layer above sees
one whole packet per read, identically on every platform.

Also bump the Android outbound SEND_QUEUE_CAP from 8 to 32. Swept against
the peer speedtest: 8 starved the radio's connection events (~half
throughput), 64 bufferbloated TCP, 32 was best (~200/500 kbps up/down). The
sweep was noisy and non-monotonic — run-to-run BLE variance (RF, 2M PHY /
connection-priority grants) rivals the knob's effect — so 32 is the
best-observed value pending re-validation with PHY/interval instrumentation,
not a proven optimum.

src/transport/ble/android_io.rs

@@ -59,12 +59,19 @@ const ANDROID_ADAPTER: &str = "ble0";
 /// tolerable since FMP/Noise above retransmits.
 const CHANNEL_CAP: usize = 256;
 
-/// Bound on the **outbound byte** queue (Rust → Kotlin writer). Kept shallow on
-/// purpose: a BLE link's bandwidth-delay product is ~1 packet, so a deep queue
-/// only bufferbloats it — RTT balloons to seconds and TCP above can't ramp. A
-/// small bound makes `BleStream::send` backpressure (the `SyncSender` blocks),
-/// which propagates up through FSP/MMP to the TUN and TCP's own flow control.
-const SEND_QUEUE_CAP: usize = 8;
+/// Bound on the **outbound byte** queue (Rust → Kotlin writer), fixed at channel
+/// creation. A bounded queue makes `BleStream::send` backpressure (the `SyncSender`
+/// blocks), propagating flow control up through FSP/MMP to the TUN and TCP rather
+/// than letting an unbounded queue bufferbloat the link.
+///
+/// 32 is empirical, swept against the peer speedtest: 8 starved the radio's
+/// connection events (~half throughput), 64 bufferbloated TCP (regressed), and 32
+/// was best (~200/500 kbps up/down). The sweep was noisy and non-monotonic across
+/// single runs, though — run-to-run BLE variance (RF, and whether the OS grants 2M
+/// PHY / high connection priority that session) rivals the effect of this knob — so
+/// 32 is the best-observed working value, to be re-validated with repeated runs +
+/// PHY/interval instrumentation, not a proven optimum.
+const SEND_QUEUE_CAP: usize = 32;
 
 /// Transport default MTU, used when the OS reports an unknown (0) channel MTU.
 /// Matches `DEFAULT_BLE_MTU` in `config/transport.rs`.

src/transport/ble/mod.rs

@@ -1,9 +1,11 @@
 //! BLE L2CAP Transport Implementation
 //!
 //! Provides BLE-based transport for FIPS peer communication using L2CAP
-//! Connection-Oriented Channels (CoC) in SeqPacket mode. L2CAP CoC
-//! preserves message boundaries (unlike TCP byte streams), so no FMP
-//! framing is needed — each send/recv is one FIPS packet.
+//! Connection-Oriented Channels (CoC). BlueZ (SeqPacket) preserves message
+//! boundaries, but stream-oriented backends (Android `BluetoothSocket`,
+//! CoreBluetooth) do not, so the receive path recovers FIPS packet boundaries
+//! with the shared FMP framer (`tcp::stream::read_fmp_packet`) over a
+//! [`stream_read::BleStreamRead`] adapter — reliable on every platform.
 //!
 //! ## Architecture
 //!
@@ -24,6 +26,7 @@ pub mod io;
 pub mod pool;
 pub mod psm;
 pub mod stats;
+pub mod stream_read;
 
 // The Android backend (radio in Kotlin, bytes over a bridge). Compiled on
 // Android, and under `cfg(test)` on any host so its channel logic is unit-tested
@@ -42,6 +45,9 @@ use discovery::DiscoveryBuffer;
 use io::{BleIo, BleScanner, BleStream};
 use pool::{BleConnection, ConnectionPool};
 use stats::BleStats;
+use stream_read::BleStreamRead;
+
+use crate::transport::tcp::stream::{StreamError, read_fmp_packet};
 
 use secp256k1::XOnlyPublicKey;
 use std::collections::HashMap;
@@ -376,9 +382,15 @@ impl<I: BleIo> BleTransport<I> {
             }
         };
 
+        // One buffering reader per connection, shared by the pubkey exchange
+        // and the receive loop so coalesced bytes are never dropped.
+        let recv_mtu = stream.recv_mtu();
+        let stream = Arc::new(stream);
+        let mut reader = BleStreamRead::new(Arc::clone(&stream), recv_mtu);
+
         // Pre-handshake pubkey exchange (temporary, pre-XX)
         if let Some(ref our_pubkey) = self.local_pubkey {
-            match pubkey_exchange(&stream, our_pubkey).await {
+            match pubkey_exchange(&mut reader, our_pubkey).await {
                 Ok(peer_pubkey) => {
                     debug!(addr = %addr, "BLE outbound pubkey exchange complete");
                     self.discovery_buffer
@@ -391,24 +403,26 @@ impl<I: BleIo> BleTransport<I> {
             }
         }
 
-        self.promote_connection(addr, &ble_addr, stream).await
+        self.promote_connection(addr, &ble_addr, stream, reader)
+            .await
     }
 
     /// Promote a newly established stream into the connection pool.
     ///
-    /// Spawns the receive loop and inserts into the pool with eviction.
+    /// Spawns the receive loop (driven by `reader`) and inserts into the pool
+    /// with eviction.
     async fn promote_connection(
         &self,
         addr: &TransportAddr,
         ble_addr: &BleAddr,
-        stream: I::Stream,
+        stream: Arc<I::Stream>,
+        reader: BleStreamRead<I::Stream>,
     ) -> Result<(), TransportError> {
         let send_mtu = stream.send_mtu();
         let recv_mtu = stream.recv_mtu();
-        let stream = Arc::new(stream);
 
         let recv_task = tokio::spawn(receive_loop(
-            Arc::clone(&stream),
+            reader,
             addr.clone(),
             Arc::clone(&self.pool),
             self.packet_tx.clone(),
@@ -496,9 +510,16 @@ impl<I: BleIo> BleTransport<I> {
 
             match result {
                 Ok(Ok(stream)) => {
+                    let send_mtu = stream.send_mtu();
+                    let recv_mtu = stream.recv_mtu();
+                    let stream = Arc::new(stream);
+                    // Shared reader: pubkey exchange + receive loop read the
+                    // same buffer so coalesced bytes survive the handoff.
+                    let mut reader = BleStreamRead::new(Arc::clone(&stream), recv_mtu);
+
                     // Pre-handshake pubkey exchange (temporary, pre-XX)
                     if let Some(ref our_pubkey) = local_pubkey {
-                        match pubkey_exchange(&stream, our_pubkey).await {
+                        match pubkey_exchange(&mut reader, our_pubkey).await {
                             Ok(peer_pubkey) => {
                                 debug!(addr = %addr_clone, "BLE outbound pubkey exchange complete");
                                 discovery_buffer.add_peer_with_pubkey(&ble_addr, peer_pubkey);
@@ -513,12 +534,8 @@ impl<I: BleIo> BleTransport<I> {
                         }
                     }
 
-                    let send_mtu = stream.send_mtu();
-                    let recv_mtu = stream.recv_mtu();
-                    let stream = Arc::new(stream);
-
                     let recv_task = tokio::spawn(receive_loop(
-                        Arc::clone(&stream),
+                        reader,
                         addr_clone.clone(),
                         Arc::clone(&pool),
                         packet_tx,
@@ -689,31 +706,38 @@ const PUBKEY_EXCHANGE_TIMEOUT_SECS: u64 = 5;
 
 /// Exchange public keys over a newly established L2CAP connection.
 ///
-/// Both sides send `[0x00][our_pubkey:32]` and receive the peer's.
+/// Both sides send `[0x00][our_pubkey:32]` and receive the peer's. Reads go
+/// through the connection's [`BleStreamRead`] buffer so a peer that fragments
+/// the 33-byte message (stream-oriented backends) is read correctly, and any
+/// bytes it coalesced after the pubkey stay buffered for the receive loop.
 /// Returns the peer's XOnlyPublicKey on success.
 async fn pubkey_exchange<S: BleStream>(
-    stream: &S,
+    reader: &mut BleStreamRead<S>,
     local_pubkey: &[u8; 32],
 ) -> Result<XOnlyPublicKey, TransportError> {
+    use tokio::io::AsyncReadExt;
+
     // Send our pubkey
     let mut msg = [0u8; PUBKEY_EXCHANGE_SIZE];
     msg[0] = PUBKEY_EXCHANGE_PREFIX;
     msg[1..].copy_from_slice(local_pubkey);
-    stream.send(&msg).await?;
+    reader.stream().send(&msg).await?;
 
-    // Receive peer's pubkey (with timeout to prevent indefinite blocking)
+    // Receive peer's pubkey (with timeout to prevent indefinite blocking).
+    // read_exact reassembles across fragmented recvs and leaves any trailing
+    // bytes in the reader's buffer.
     let mut buf = [0u8; PUBKEY_EXCHANGE_SIZE];
     let timeout = std::time::Duration::from_secs(PUBKEY_EXCHANGE_TIMEOUT_SECS);
-    let n = match tokio::time::timeout(timeout, stream.recv(&mut buf)).await {
-        Ok(result) => result?,
+    match tokio::time::timeout(timeout, reader.read_exact(&mut buf)).await {
+        Ok(Ok(_)) => {}
+        Ok(Err(e)) => {
+            return Err(TransportError::RecvFailed(format!(
+                "pubkey exchange: {}",
+                e
+            )));
+        }
         Err(_) => return Err(TransportError::Timeout),
     };
-    if n != PUBKEY_EXCHANGE_SIZE {
-        return Err(TransportError::RecvFailed(format!(
-            "pubkey exchange: expected {} bytes, got {}",
-            PUBKEY_EXCHANGE_SIZE, n
-        )));
-    }
     if buf[0] != PUBKEY_EXCHANGE_PREFIX {
         return Err(TransportError::RecvFailed(format!(
             "pubkey exchange: bad prefix 0x{:02X}",
@@ -763,10 +787,13 @@ async fn accept_loop<A>(
 
                 let send_mtu = stream.send_mtu();
                 let recv_mtu = stream.recv_mtu();
+                let stream = Arc::new(stream);
+                // Shared reader for pubkey exchange + receive loop.
+                let mut reader = BleStreamRead::new(Arc::clone(&stream), recv_mtu);
 
                 // Pre-handshake pubkey exchange (temporary, pre-XX)
                 if let Some(ref our_pubkey) = local_pubkey {
-                    match pubkey_exchange(&stream, our_pubkey).await {
+                    match pubkey_exchange(&mut reader, our_pubkey).await {
                         Ok(peer_pubkey) => {
                             debug!(addr = %ta, "BLE inbound pubkey exchange complete");
                             discovery_buffer.add_peer_with_pubkey(&addr, peer_pubkey);
@@ -792,11 +819,9 @@ async fn accept_loop<A>(
                     }
                 }
 
-                let stream = Arc::new(stream);
-
                 // Spawn receive loop
                 let recv_task = tokio::spawn(receive_loop(
-                    Arc::clone(&stream),
+                    reader,
                     ta.clone(),
                     Arc::clone(&pool),
                     packet_tx.clone(),
@@ -841,8 +866,14 @@ async fn accept_loop<A>(
 }
 
 /// Receive loop: reads packets from a BLE stream and delivers to node.
-async fn receive_loop<S: BleStream>(
-    stream: Arc<S>,
+///
+/// Recovers FIPS packet boundaries from the byte stream via the shared FMP
+/// framer ([`read_fmp_packet`]) rather than assuming one `recv` is one packet.
+/// L2CAP only preserves SDU boundaries on BlueZ (SeqPacket); stream-oriented
+/// backends (Android, CoreBluetooth) fragment and coalesce, so the framer's
+/// length-prefixed reads are what make delivery reliable across platforms.
+async fn receive_loop<S: BleStream + 'static>(
+    mut reader: BleStreamRead<S>,
     addr: TransportAddr,
     pool: Arc<Mutex<ConnectionPool<Arc<S>>>>,
     packet_tx: PacketTx,
@@ -850,21 +881,21 @@ async fn receive_loop<S: BleStream>(
     stats: Arc<BleStats>,
     recv_mtu: u16,
 ) {
-    let mut buf = vec![0u8; recv_mtu as usize];
     loop {
-        match stream.recv(&mut buf).await {
-            Ok(0) => {
-                debug!(addr = %addr, "BLE connection closed by peer");
-                break;
-            }
-            Ok(n) => {
-                stats.record_recv(n);
-                let packet = ReceivedPacket::new(transport_id, addr.clone(), buf[..n].to_vec());
-                if packet_tx.send(packet).await.is_err() {
+        match read_fmp_packet(&mut reader, recv_mtu).await {
+            Ok(packet) => {
+                stats.record_recv(packet.len());
+                let received = ReceivedPacket::new(transport_id, addr.clone(), packet);
+                if packet_tx.send(received).await.is_err() {
                     trace!("BLE packet_tx closed, stopping receive loop");
                     break;
                 }
             }
+            // Clean peer close (EOF at a packet boundary) is expected, not an error.
+            Err(StreamError::Io(e)) if e.kind() == std::io::ErrorKind::UnexpectedEof => {
+                debug!(addr = %addr, "BLE connection closed by peer");
+                break;
+            }
             Err(e) => {
                 debug!(addr = %addr, error = %e, "BLE receive error");
                 stats.record_recv_error();
@@ -998,7 +1029,12 @@ async fn scan_probe_loop<I: io::BleIo>(
 
         // Pubkey exchange, then promote connection to pool
         let ta = addr.to_transport_addr();
-        match pubkey_exchange(&stream, &our_pubkey).await {
+        let send_mtu = stream.send_mtu();
+        let recv_mtu = stream.recv_mtu();
+        let stream = Arc::new(stream);
+        // Shared reader for pubkey exchange + receive loop.
+        let mut reader = BleStreamRead::new(Arc::clone(&stream), recv_mtu);
+        match pubkey_exchange(&mut reader, &our_pubkey).await {
             Ok(peer_pubkey) => {
                 debug!(addr = %addr, "BLE probe complete");
 
@@ -1017,12 +1053,8 @@ async fn scan_probe_loop<I: io::BleIo>(
                 }
 
                 // Promote connection to pool — no second L2CAP connect needed
-                let send_mtu = stream.send_mtu();
-                let recv_mtu = stream.recv_mtu();
-                let stream = Arc::new(stream);
-
                 let recv_task = tokio::spawn(receive_loop(
-                    Arc::clone(&stream),
+                    reader,
                     ta.clone(),
                     Arc::clone(&pool),
                     packet_tx.clone(),

src/transport/ble/stream_read.rs

@@ -0,0 +1,175 @@
+//! `AsyncRead` adapter over a datagram-shaped [`BleStream`].
+//!
+//! L2CAP delivers different boundary guarantees per platform:
+//!
+//! - **BlueZ** (`SOCK_SEQPACKET`) preserves SDU boundaries — one `recv` is
+//!   exactly one FIPS packet.
+//! - **Android** (`BluetoothSocket` input stream) and **CoreBluetooth** are
+//!   byte-stream oriented — a `recv` may return a fragment of a packet, or
+//!   several packets coalesced.
+//!
+//! FIPS packets are self-delimiting via the 4-byte FMP common prefix, so the
+//! shared framer [`crate::transport::tcp::stream::read_fmp_packet`] can recover
+//! boundaries from any byte stream. This adapter turns a [`BleStream`] (whose
+//! `recv` fills a `&mut [u8]`) into the [`AsyncRead`] that framer expects. On a
+//! SeqPacket backend it's a no-op pass-through; on a stream backend it
+//! reassembles. Either way the layer above sees one whole packet per read.
+
+use std::future::Future;
+use std::pin::Pin;
+use std::sync::Arc;
+use std::task::{Context, Poll};
+
+use tokio::io::{AsyncRead, ReadBuf};
+
+use super::io::BleStream;
+
+/// An in-flight `recv`: owns its scratch buffer and yields an owned `Vec`, so
+/// the future is `'static` and can live across `poll_read` calls.
+type PendingRecv = Pin<Box<dyn Future<Output = std::io::Result<Vec<u8>>> + Send>>;
+
+/// Buffers a [`BleStream`] into an [`AsyncRead`] byte stream.
+pub struct BleStreamRead<S: BleStream + 'static> {
+    stream: Arc<S>,
+    /// Per-`recv` scratch size; also the framer's MTU bound.
+    mtu: u16,
+    /// Bytes from the last `recv` not yet consumed by the framer.
+    leftover: Vec<u8>,
+    /// Read cursor into `leftover`.
+    pos: usize,
+    /// In-flight `recv`, if one is underway.
+    pending: Option<PendingRecv>,
+}
+
+impl<S: BleStream + 'static> BleStreamRead<S> {
+    /// Wrap a stream. `mtu` is the per-`recv` scratch size (use the channel's
+    /// recv MTU).
+    pub fn new(stream: Arc<S>, mtu: u16) -> Self {
+        Self {
+            stream,
+            mtu: mtu.max(1),
+            leftover: Vec::new(),
+            pos: 0,
+            pending: None,
+        }
+    }
+
+    /// The wrapped stream, for sending on the same channel (e.g. the pubkey
+    /// exchange writes here while reads come through the buffer).
+    pub fn stream(&self) -> &S {
+        &self.stream
+    }
+}
+
+impl<S: BleStream + 'static> AsyncRead for BleStreamRead<S> {
+    fn poll_read(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        dst: &mut ReadBuf<'_>,
+    ) -> Poll<std::io::Result<()>> {
+        // BleStreamRead is Unpin (all fields are), so get_mut is sound.
+        let this = self.get_mut();
+        loop {
+            // Serve buffered bytes first.
+            if this.pos < this.leftover.len() {
+                let avail = &this.leftover[this.pos..];
+                let n = avail.len().min(dst.remaining());
+                dst.put_slice(&avail[..n]);
+                this.pos += n;
+                return Poll::Ready(Ok(()));
+            }
+
+            // Buffer drained: pull the next datagram. The future owns its
+            // scratch and returns it truncated, so it captures only `Arc<S>`.
+            if this.pending.is_none() {
+                let stream = Arc::clone(&this.stream);
+                let mtu = this.mtu as usize;
+                this.pending = Some(Box::pin(async move {
+                    let mut scratch = vec![0u8; mtu];
+                    let n = stream
+                        .recv(&mut scratch)
+                        .await
+                        .map_err(|e| std::io::Error::other(e.to_string()))?;
+                    scratch.truncate(n);
+                    Ok(scratch)
+                }));
+            }
+
+            match this.pending.as_mut().unwrap().as_mut().poll(cx) {
+                Poll::Ready(Ok(buf)) => {
+                    this.pending = None;
+                    // A zero-length recv is the BleStream peer-closed signal;
+                    // leaving `dst` unfilled surfaces as EOF to the framer.
+                    if buf.is_empty() {
+                        return Poll::Ready(Ok(()));
+                    }
+                    this.leftover = buf;
+                    this.pos = 0;
+                    // loop to copy out
+                }
+                Poll::Ready(Err(e)) => {
+                    this.pending = None;
+                    return Poll::Ready(Err(e));
+                }
+                Poll::Pending => return Poll::Pending,
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::transport::ble::addr::BleAddr;
+    use crate::transport::ble::io::MockBleStream;
+    use tokio::io::AsyncReadExt;
+
+    fn addr(n: u8) -> BleAddr {
+        BleAddr {
+            adapter: "hci0".to_string(),
+            device: [0xAA, 0xBB, 0xCC, 0xDD, 0xEE, n],
+        }
+    }
+
+    /// Several small recvs reassemble into one read_exact.
+    #[tokio::test]
+    async fn reassembles_fragmented_recv() {
+        let (a, b) = MockBleStream::pair(addr(1), addr(2), 2048);
+        // Peer sends three fragments that together form one 10-byte message.
+        a.send(b"abc").await.unwrap();
+        a.send(b"defg").await.unwrap();
+        a.send(b"hij").await.unwrap();
+
+        let mut reader = BleStreamRead::new(Arc::new(b), 2048);
+        let mut out = [0u8; 10];
+        reader.read_exact(&mut out).await.unwrap();
+        assert_eq!(&out, b"abcdefghij");
+    }
+
+    /// A recv carrying several packets is served across multiple reads without
+    /// dropping the tail (the coalescing case).
+    #[tokio::test]
+    async fn serves_coalesced_recv_in_pieces() {
+        let (a, b) = MockBleStream::pair(addr(1), addr(2), 2048);
+        a.send(b"0123456789").await.unwrap();
+
+        let mut reader = BleStreamRead::new(Arc::new(b), 2048);
+        let mut first = [0u8; 4];
+        reader.read_exact(&mut first).await.unwrap();
+        assert_eq!(&first, b"0123");
+        let mut rest = [0u8; 6];
+        reader.read_exact(&mut rest).await.unwrap();
+        assert_eq!(&rest, b"456789");
+    }
+
+    /// A closed stream surfaces as EOF (read_exact errors with UnexpectedEof).
+    #[tokio::test]
+    async fn closed_stream_is_eof() {
+        let (a, b) = MockBleStream::pair(addr(1), addr(2), 2048);
+        drop(a); // peer closes
+        let mut reader = BleStreamRead::new(Arc::new(b), 2048);
+        let mut out = [0u8; 4];
+        let err = reader.read_exact(&mut out).await.unwrap_err();
+        assert_eq!(err.kind(), std::io::ErrorKind::UnexpectedEof);
+    }
+}