Reconnect a frozen room socket on resume instead of trusting a stale 'open'

[ankitsejwal]

Fixes #108.

The bug

An iOS background freeze suspends the JS event loop, so the room socket can die server-side with no onclose ever firing — status stays a stale 'open'. Two things then went wrong on resume:

1. The foreground handler only reconnected when the socket wasn't open (useLucidAsk.tsx, getStatus() !== 'open'). After a freeze the status reads 'open', so we skipped the reconnect and kept trusting a dead socket.
2. A message sent in that window hit ws.send() on the dead socket and was orphaned — the payload lands in sent but never outbox, and only a retryable daemon error resends from sent. A dead-socket send gets no daemon response at all, so nothing resent it: it sat until the first-token timeout → "Lucid didn't respond", with the daemon never receiving it.

The fix

Judge liveness by the inbound-frame clock, not status:

• resumeForeground() (called on AppState → active) forces a fresh dial when the socket looks dead — no inbound frame for >15s (STALE_INBOUND_MS) despite a server we've proven echoes. A genuinely live socket echoes a keepalive every ~10s, so it can't reach 15s while alive; the check is gated on pongsSeen > 0 and a socket that's been open a while, so a fresh open or a non-echoing server never false-positives.
• enqueue() now routes a send to the outbox (and reconnects) instead of ws.send() whenever the socket only looks open. The fresh onopen flushes the outbox onto a live socket.

This makes status reflect reality before the user can type, which is the clean fix for the orphaned send — no "resend from sent" (which would risk double-running an ask against the room's frame replay).

Verification

• tsc --noEmit clean. No behavior change on a healthy socket (the staleness check can't trigger while inbound is flowing).
• Mobile has no unit-test harness in this repo, so this is typecheck- + review-verified; the logic is defensive (gated on pongsSeen and open-duration to avoid false reconnects).

Sibling issues from the same session: #107 (silent resume memory-wipe, fixed in #110) and #109 (diagnostics blind to the freeze).