fix(rpc): record connection edges before Windows named-pipe idle exit

[code-yeongyu]

Summary

Windows named-pipe RPC hosts no longer drop the public endpoint between ensureHost() readiness and the first real client, and the remaining turn-active lifecycle waiter no longer uses a nested 15s clock inside a 60s test.

Root causes (product)

1. Authenticated connection edges (including short readiness probes) can fall entirely between idle ticks. The supervisor now records attach/detach and turn start/settle immediately so the idle window measures continuous inactivity.
2. Post-readiness lock release and state writes could outlast a transient host's idle window on a loaded Windows runner, so the named pipe was gone before ensureHost() returned (connect ENOENT). The authenticated readiness connection now stays attached through the endpoint lock's commit and close.

Residual CI failure (test waiter, not ENOENT)

On head 3fbd2e673, run 34823696518 attempts 1–2 passed. Attempt 3 (job 103917291025) failed:

does not exit while a turn is active even with no connections; exits after the turn settles → Timed out waiting for RPC record at JsonlPeer.waitFor (default 15s). 30 passed / 1 failed / 3 skipped.

That stack is the record waiter, not connect ENOENT. The test never reached the no-connection idle assertion. Same SHA passed on attempts 1, 2, and 4, so this is a nested 15s budget vs loaded-runner session/prompt/agent_start latency, not a third product idle-race.

Fix (test-only): arm JsonlPeer.waitFor before scanning buffered records; mark sibling waiter timeouts handled; include seen {type,id,command} on timeout; bind open/prompt/agent_start waits to the case's existing 60s test budget; restore the mock API key for the held-model turn. No product idle-window change, no retries, no skips.

Head

231546b4a on origin/main 5fb5d638d (includes the waiter fix f04ecdc1a).

Proof still required on this head

• Three consecutive green RPC named pipes (Windows) executions on 231546b4a (earlier-head streaks do not count).
• Combined lifecycle/ensure/transport/daemon suites on the macOS proof host (pending; not claimed here).

Refs #1656

Plan: .omo/plans/omo-dependency-diet.md

[code-yeongyu]

Handoff caveat: main advanced again to 5d8c339 (PR #1665) after the requested rebase onto 7924335 and the three successful CI executions on head 4b7ba6b. GitHub now reports this PR CONFLICTING / DIRTY. All recorded acceptance results apply to 4b7ba6b against 7924335; they do not claim compatibility with the subsequently advanced main. No merge was performed.