[BUG] UDS Connection::recv is not cancel-safe — both session loops drop it mid-frame

Verified still reproducing on f1fea6c (2026-09-16). Nothing has been
fixed. The server-side loss is now reproduced over a real socket, the client
loop turns out to have the same shape, and the fix originally proposed here
would not have worked — see Fix. References are refreshed after
#248/#250 moved TCP onto core's FramedConnection. Claim-by-claim check in
Verification.

Summary

Both session loops rebuild Connection::recv() on every select_biased!
iteration and drop it whenever another arm wins, so recv must be cancel-safe.
TCP, serial and WebSocket are; UDS is not. A UDS frame that needs more than one
read, arriving while another arm becomes ready, is cut in half and lost without
a trace.

The mechanism

Server — aimdb-core/src/session/server.rs:130-153:

let step = {
    let mut recv = conn.recv().fuse();
    let mut event = core::pin::pin!(event_rx.recv().fuse());
    select_biased! {
        r  = recv  => ...,
        ev = event => ...,
        sub_id = subs.select_next_some() => ...,
    }
};

Client — drive_connection in aimdb-core/src/session/client.rs:627-661 has the
same shape, with keepalive, prune and caller cmd arms beside recv.

select_biased! only prioritizes among futures that are ready. If recv is
Pending and another arm is Ready, the select resolves on that arm and recv — a
local — is dropped at the end of the block. The next iteration builds a fresh
one. The Connection trait (aimdb-core/src/session/mod.rs:376-378) does not
state the requirement.

Transport-by-transport

Transport	recv impl	Cancel-safe?
TCP, serial	core's FramedConnection (aimdb-core/src/session/io.rs:281) — partial bytes live in the connection's Framer; ByteStream::read is cancel-safe on the shipped adapters (io.rs:66-72)	yes
WebSocket server (aimdb-websocket-connector/src/transport.rs:73)	pending queue on self, then self.ws.recv()	yes
WebSocket client (transport.rs:144)	self.ws.next()	yes
UDS (aimdb-uds-connector/src/transport.rs:41)	read_line into a local String	no

The UDS impl:

fn recv(&mut self) -> BoxFut<'_, TransportResult<Option<Vec<u8>>>> {
    Box::pin(async move {
        let mut line = String::new();
        match self.reader.read_line(&mut line).await { ... }
    })
}

read_line consumes bytes from the BufReader into the future's own buffer
before the line is complete. Dropping the future discards them; the BufReader
has already advanced.

Impact

Server — reproduced. The request's head is lost. Its tail arrives as a frame
of its own, fails to decode (it carries the outer object's unmatched }), and is
skipped by the malformed-frame arm (server.rs:183-186). The session survives;
the request gets no reply and nothing is logged.

Client — by inspection. aimdb-client's AimxConnection dials unix://
endpoints through UdsDialer (aimdb-client/src/endpoint.rs:124), so the CLI
and anything else on that client are exposed the same way. A reply or event
arriving in pieces while the caller issues a command, the keepalive timer fires,
or a prune sweep runs is lost; the client also skips the malformed tail
(client.rs:737), so the call waits for a reply that never comes.

The trigger is a frame larger than one read — large record.set params, a
record.list over a big database, a large event payload — while the other
direction is active on the same connection. Rare, and hard to diagnose when it
does happen.

Fix

Either option stays inside aimdb-uds-connector.

Option A — move UDS onto FramedConnection. Design 052 already names this
route for the crate (docs/design/052-runtime-neutral-connectors.md:258). Add
an NDJSON Framer (split on \n, keep today's trailing-\r tolerance; a bad
line is FrameFault::Recoverable, since the next newline resyncs) and build
connections through it, as TCP does with FramingListener/FramingDialer.
Partial lines then live in the framer, and UDS stops being the one hand-rolled
transport. TokioByteStream<UnixStream> already satisfies ByteStream, but
aimdb-tokio-adapter is only a dev-dependency here — add it, or wrap
UnixStream in a local newtype.

Option B — minimal. Keep the hand-rolled connection and hoist the partial
line into it:

pub struct UdsConnection {
    reader: BufReader<OwnedReadHalf>,
    writer: OwnedWriteHalf,
    pending: Vec<u8>, // survives a dropped recv future
    peer: PeerInfo,
}

Read with read_until(b'\n', &mut self.pending) and take the buffer only once
it ends in a newline (or at EOF).

It must be read_until into a Vec<u8>, not read_line into a String.
This issue originally proposed a pending: String resumed with read_line.
That would not fix it: tokio moves the String's contents into the future and,
if the future is dropped, leaves the String empty. Tokio documents read_line
as not cancel-safe for exactly this reason and read_until as resumable.

Also worth doing

State the requirement on Connection::recv (aimdb-core/src/session/mod.rs:376-378):
both loops drop the future whenever another arm wins. ByteStream already
carries equivalent wording (io.rs:66-72) to mirror.

Test

Deterministic, in aimdb-uds-connector/tests/, against the real UdsServer:

1. Subscribe to a SpmcRing record with with_remote_access().
2. Write the first bytes of a record.list request, no newline; pause so the
   server reads them.
3. Produce one value — the event arm wins the select.
4. Read the event, write the rest of the request, and assert the reply arrives.

Keep the control case alongside: the same split request with no event in between
must be answered, which shows it is the event, not the split, that loses it.
A client-side counterpart needs a raw UnixListener peer that writes half a
reply while the caller sends a second command.

---

Verification

Checked against f1fea6c on 2026-09-16. The test above was written as a scratch
integration test in aimdb-uds-connector/tests/ and run five times, with the
same result each time:

control  (split, no event)   -> {"t":"reply","id":2,"ok":[…]}
event while head is buffered -> {"t":"event","seq":1,"sub":"1","data":{"n":1}}
split request after event    -> (no reply within 2 s)
whole request afterwards     -> {"t":"reply","id":4,"ok":[…]}

Claim by claim:

Claim as filed	Status
run_session drops recv when another arm wins	holds — server.rs:130-153, unchanged
UDS recv uses read_line into a local String	holds — transport.rs:41-57, unchanged
Truncated frame is skipped, the session survives	holds — reproduced; arm now at server.rs:183-186
The remainder "can mis-decode too"	dropped — the tail always fails to decode and is skipped
TCP is cancel-safe via self.acc in tokio_transport.rs:59	outdated — file removed in #248/#250; TCP and serial use FramedConnection, still cancel-safe
WebSocket recv at transport.rs:140	moved — server :73, client :144; both still cancel-safe
Fix: resume read_line into a pending: String	would not fix it — see Fix
No mention of cancel safety anywhere in session/	outdated — ByteStream documents it (io.rs:66-72); Connection still does not
Connection::recv at session/mod.rs:345	moved — :376-378
Only the server loop is affected	incomplete — client drive_connection (client.rs:627-661) has the same shape
Found on feat/retire-aimdb-ws-protocol	merged as #201; the bug survived the merge

---

Verification section generated by Claude Code