v0.21.5

Added

• feat(listen): single-instance flock guard for sac listen (#266,
  operator task #26 sub (1)). A second sac listen while one already
  holds the port used to crash uvicorn with bare EADDRINUSE + a
  Python traceback — loud but with no diagnostic about which process
  held the port. New _listen/_single_instance.py (acquire_listen_lock,
  release_listen_lock, ListenAlreadyRunningError, default_lock_dir)
  takes a port-scoped flock at <lock_dir>/listen-<port>.pid BEFORE
  uvicorn binds, stamps the current PID in the file body, and on
  conflict fails loud with the holding PID and lock-file path so
  kill <pid> is actionable without lsof / netstat. The flock is
  kernel-released on process exit (even SIGKILL / OOM) so a crashed
  listen never permanently jams the port. Acquired before
  _register_self_comms_node / _maybe_sync_on_start so a duplicate
  launch never touches the federated registry.
• feat(systemd): sac-listen.service hand-maintained user unit
  (#268, operator task #26 sub (3)). New scripts/systemd/sac-listen. service (Type=simple + Restart=on-failure + RestartSec=5s +
  StandardOutput=journal). The companion flock guard (sub (1))
  ensures Restart=on-failure cannot double-bind. The README is
  split into "federated scheduled jobs" (sac.accounts-refresh
  pattern, materialised from scitex_dev.jobs) vs "hand-maintained
  long-running services" (sac-listen.service lives here) so a
  future operator does not move this into the federated path by
  mistake. Install: copy to ~/.config/systemd/user/,
  daemon-reload, enable --now.

Tests

• test(channel): pin SSE auto-reconnect across listen restart
  (#267, operator task #26 sub (2)). Three no-mocks regression tests
  in tests/scitex_agent_container/_mcp/test_channel_reconnect.py
  using a real asyncio TCP server pin the invariant that the
  in-container SSE consumer (_mcp/channel.py::_consume_sse)
  reconnects after the listen drops the stream mid-flight (operator-
  restart scenario), records ≥ 2 actual TCP connection attempts, and
  recovers when the server starts AFTER the consumer is already
  trying. The exponential-backoff loop (0.5s → 30s cap) was already
  implemented; this PR pins it against regression. A flake-guard
  using a held-socket port reservation replaces a brittle "start →
  stop → restart on same port" approach.