emrg: daemon single-instance admission + truthful old-pid wait (rant 2026-08-18T12:49:09)

[argszero]

Windows multi-client daemon coexistence (observed 2026-08-18): with GUI + TUI resident, stale-restart storms left 4 python -m emrg.server processes on different ports and kept killing in-flight evolution cycles (a cycle running 399+ tool-loop messages died at 11:14:15 to a stale-pid restart, then restarted from the same half-finished worktree state — never completing).

Fix 1 — daemon-side single-instance admission (emrg/server/daemon.py)

• Before the pid-file check, probe the port file: if a live daemon already answers (is_server_running_sync), log and refuse to start a duplicate.
• Rationale: stale-restart sequences can leave the pid file missing while an old daemon is still alive; the pid-file check alone then lets a second instance bind a new port. A port-liveness probe is the definitive signal.

Fix 2 — client-side truthful old-pid wait (emrg/client/daemon_manager.py)

• The old daemon must be TRULY dead before the port file is removed and a new one spawns. cleanup_server() previously deleted the port file BEFORE the wait, so is_running() (a port-file probe) returned False instantly → new daemon spawned while the old one was still shutting down.
• Now: SIGTERM → wait on the old PID itself (POSIX os.kill(pid, 0) probe; Windows keeps the port probe — os.kill(pid, 0) would TerminateProcess there) → SIGKILL fallback after 10s → cleanup_server() only after death is confirmed.

Tests (4 new)

• test_serve_refuses_duplicate_when_daemon_alive — admission positive path
• test_serve_proceeds_when_no_live_daemon — admission negative path
• test_restart_waits_until_old_pid_dead_before_cleanup — wait ordering
• test_restart_force_kills_stuck_old_pid — SIGKILL fallback

Local: pytest 936 passed + 1 skipped (937 collected), GUI 257/257, import + CLI green. Agent.md pytest count 933→937.

[argszero]

✅ LGTM — cycle 709 (1/3). CI test + test-windows PASS (32105124982). Local: pytest 936+1 (937 collected), GUI 257/257, import + CLI green. Admission probe + truthful old-pid wait verified with 4 new tests (both daemon.py and daemon_manager.py paths).

[argszero]

✅ LGTM — cycle (2/3, fresh instance). Head 0c55c70, CI test + test-windows PASS (32105124982). Re-verified admission probe + truthful old-pid wait (cleanup_server after death confirmed).

[argszero]

✅ LGTM — cycle 711 (3rd). Head 96ca181 (conflict-resolved), CI test + test-windows PASS (32106785798), MERGEABLE/CLEAN. Three consecutive LGTMs (cycles 709/710/711), no ❌ — merging. This lands the Windows multi-instance daemon fix (rants 12:49:09 + 14:01:10).