todo.md

Remaining Priorities

1. spsc_wrapping cancellation/lifetime The review item about recv_proxy_loop holding shared_ptr<stream> forever is still real. If the underlying stream never returns and nobody drives set_closed(), the wrapper can self-retain indefinitely. This needs an explicit cancellation/stop path, not just atomic flags.

2. Reply/control priority in stream_transport Replies and control traffic should not compete fairly with outbound requests. Split outbound queues into:

   • high priority: call_receive, try_cast_receive, addref_receive, init responses, close/control replies
   • normal priority: call_send, posts, ordinary outbound work

3. Listener shutdown hardening beyond the late-accept fix The obvious post-stop spawn race is fixed, but the listener still needs a tighter shutdown model around accept loop ownership and readiness/stop sequencing.

4. spsc_wrapping error/close semantics audit Timeout handling and send-failure propagation are fixed, but the wrapper still needs a consistency pass for:

   • underlying receive native errors
   • whether close should wake blocked paths more explicitly
   • whether receive-side failure should also be latched

5. Streaming wire validation cleanup One bad assert(prefix.direction) was replaced, but the transport receive path should be reviewed for release-safe validation of malformed wire input instead of debug-only assumptions.

6. Regression coverage for discovered bugs Add targeted tests for:

   • io_uring timed receive timeout/cancel association
   • transport queued-send wakeup/progress
   • pending_transmits_ lifetime on disconnect/cancel
   • spsc_wrapping send failure propagation

7. Diagnostic cleanup review Ensure only useful low-overhead diagnostics remain and no debug-only fields are accidentally shaping behavior.

8. Lower-priority performance cleanup Still worth revisiting later:

   • queued_send_item heap-copy overhead in streaming transport
   • any remaining avoidable buffer churn on large payload paths