fix: bound stale running turn recovery (rebased) - #85

Rebases and completes #83 (draft) onto current main, fixing #80.

What this carries over from #83

• Separates real ACP/ZCode protocol progress from projection-based liveness probes: a stale projection.status=running no longer refreshes the 120s no-progress deadline.
• At the deadline, probes the authoritative prompt lock via session/goal show:
  • lock held → defer the terminal decision by another 120s (protects legitimately long model/tool operations);
  • lock released/unknown → bounded max_turn_requests outcome as before.
• Consumes already-queued events before making a deadline decision.

Adaptations to current main (post #84)

• Resolved the session.ts conflicts: kept the steer-swallow guard, recordProtocolProgress() now replaces the bare lastProgress refresh.
• Dropped the deadline else if (turn.cancelled) branch — main's in-loop cancel check already covers it (the branch would misreport a user cancel as max_turn_requests).
• probePromptLock matches the lock-busy error by code 1308 first (same code the send-retry loop keys on), message-text matching kept only as a legacy fallback — guards against backend message drift killing a live turn.
• The two new stopBackendTurn calls pass turn.foregroundExecutionId so the v4 stop actually kills the generation.
• Regression tests' microtask pump budget raised 20 → 80 (main's prompt() has more async preamble).

Policy decision on #83's open question

No hard cap on lock-held deferral (backend-owned long work is legitimate; user cancel remains the escape hatch). A follow-up to surface deferral to the client as a session/update note is suggested instead.

Verification

• tsc --noEmit clean, eslint clean, prettier clean
• Full suite: 837/837 passed (includes the two stale-running regression tests from fix: bound stale running turn recovery #83)

Closes #80
Supersedes #83