fix: end the one-time phone's pre-outcome waits at the room's deadline

[dormouse-bot]

OneTimeClient.connectOnce computed the room's hard deadline (expiry * 1000 + ONE_TIME_EXPIRY_GRACE_MS) but only handed it to the session core's exchange and exchangeControl. The waits before those — generateNoiseKeyPair, createNoiseInitiator, writeMessage, and the rendezvous socket's open — run through #race, which a failure or close() could interrupt but no timer could. A socket that stayed CONNECTING therefore left the page on its code screen indefinitely, instead of showing ONE_TIME_LINK_EXPIRED_MESSAGE as docs/specs/one-time.md's failure table promises.

connectOnce now arms one timer at the hard deadline that calls #fail(ONE_TIME_LINK_EXPIRED_MESSAGE), which wakes #race and the core's waiters alike. It is cleared once an ok outcome is read (from there DIRECT_ONLY_DEADLINE_MS bounds the direct-path wait, so the "between an ok outcome and the switch" copy is unchanged) and in #teardown, so close() still leaves nothing armed.

The new case in one-time-client.test.ts opens a socket that never opens, advances the test clock to the hard deadline, and expects the expired copy, a normally closed socket, and no armed timers. Before the change it hung until the 5 s test timeout; it passes now, along with the rest of one-time-client.test.ts and one-time-e2e.test.ts. The phone-client section of one-time.md gains the rule, and its word budget is ratcheted for it.

Closes #914 — automated triage

[cloudflare-workers-and-pages]

Deploying mouseterm with    Cloudflare Pages

Latest commit:	f086d51
Status:	⚡️  Build in progress...

View logs