Stop silently wiping Ask Lucid memory on a slow resume

[ankitsejwal]

Fixes #107.

The bug

On the Code tab, Ask Lucid would answer a follow-up with zero memory of the earlier turns — no error, nothing in the diagnostics dump. It wasn't the socket: there was no reconnect between the message that still had memory and the one that didn't.

Every turn runs claude -p --resume <sessionId>. The daemon bounded the resumed run with a 15s first-output deadline (RESUME_FIRST_OUTPUT_MS); if no first token arrived in 15s it assumed the --resume handle was dead and silently re-ran with resume: undefined — a fresh, memoryless session in the same worktree (runCodeTurn.ts, same shape for the pod ask in room/client.ts). The fresh turn answers normally and hands back a new session id that the phone adopts, so the thread is re-anchored to a session that never saw the earlier messages.

The mismatch: the phone waits 90s for a coding turn's first token (CODE_FIRST_TOKEN_TIMEOUT) because coding turns legitimately plan/read before answering — but the daemon gave up at 15s. Any resumed coding turn slow past 15s silently lost its memory. (The executor only yields on the first assistant token, so this "first-output" window is really time-to-first-token.)

The fix

Align the daemon's resume first-output deadline with the phone's per-surface first-token budget:

• pod ask: RESUME_FIRST_OUTPUT_MS 15s → 30s (= phone DEFAULT_FIRST_TOKEN_TIMEOUT)
• coding: new RESUME_CODE_FIRST_OUTPUT_MS = 90s (= phone CODE_FIRST_TOKEN_TIMEOUT)

We no longer abandon a resume the phone is still awaiting. A genuinely dead/pruned handle still falls back to fresh — just after this window, where the phone's own timeout + lost/recover grace already cover the wait. An ended (empty) stream still self-heals immediately as before; only a hanging handle waits out the deadline.

Tests

• A guard test pinning the deadlines ≥ the phone budgets (prevents a silent revert to 15s).
• A behavioral test: a resumed turn slow to its first token but within budget runs once and keeps its resume id (memory preserved) instead of retrying fresh.
• Existing self-heal / no-loop / long-turn / abort tests still pass (48/48).

Follow-ups (filed separately)

• A message sent right after unlocking the phone can silently get no reply #108 — a message sent right after unlocking the phone can silently get no reply (stale 'open' socket + orphaned send).
• Room diagnostics is blind to background freezes and silent session resets #109 — room diagnostics is blind to the background freeze and to the silent session swap; includes surfacing a fresh-session fallback so the residual dead-handle case is never silent.