impl(pm-stack): quota watchdog escalation — fleet queue task + deduplicated incident issue for a browser-enabled agent #6033

Implements rung 4 of the browser-dependent quota watchdog (design #6031; core in the sibling watchdog-core issue). Separable from the core: the core detects and repairs; this issue is how a failure the ladder cannot fix becomes work for a browser-enabled agent — and, when only Drew can unblock it, becomes visible to Drew without delay.

Mechanism (per the design)

On escalation for a source, deduplicated by (source, failure class) — one open incident per pair:

1. Fleet queue task enqueued (agent.tasks via fleet queue) with the full context packet, in the fleet's standing packet shape:
   • outcome wanted: "source <id> produces a fresh latest.json with real quota windows again"
   • failure context: class + class history, consecutive failures, the watchdog's persisted repairLog (every rung tried, with outcomes), producer history.jsonl tail reference, state-file path, adapter file reference
   • required capability, named so the claiming lane can actually do it: browser-enabled agent; laptop Helium via CDP (:19222); adapter origins only
   • acceptance check: a fresh producer document lands and the watchdog auto-closes the incident (the fixer does not close anything by hand)
   • stop conditions: a human-only challenge (2FA / verify-it's-you / passkey) → stop, mark human-blocked, do not work around it; anything outside the adapter's origin allowlist → stop and report on the incident issue
2. GitHub incident issue opened in the repo that owns the adapter (Mistral: developing-today-forks/openchamber), cross-linked to design(pm-stack): browser-dependent quota watchdog — CDP scrape producers that self-repair, escalate to a browser-enabled agent #6031, labeled as a watchdog incident. This is the durable record and the fixing agent's reporting surface (FILE THE ISSUE discipline) — the queue task is the trigger, the issue is the record.
3. Auto-close on recovery: the watchdog watches for the next producer success; on success it comments the resolution on the incident issue, closes it, and marks the queue task done. No human bookkeeping for recovery.
4. Human-blocked surfacing: incidents whose blocker is a human-only step carry the flag/label the orchestration pass and stewardship sweep already use for NEEDS DREW items, and they surface immediately (not at the 24 h threshold) — a 2FA prompt is a 30-second Drew unblock, and the 2026-10-25 lesson (a sign-in wait sat unreported ~30 min) is the failure this prevents. Non-human escalations surface to Drew only if still unresolved 24 h after escalation (design open question 1 in design(pm-stack): browser-dependent quota watchdog — CDP scrape producers that self-repair, escalate to a browser-enabled agent #6031; threshold lives in one named constant).

Tradeoffs recorded in the design (why not the alternatives)

• Issue-only escalation: durable but waits for a sweep — too slow as the sole trigger.
• A dedicated cron spawning a browser agent on a flag file: guaranteed attention, but a standing job burning quota to watch an almost-always-empty inbox — rejected as primary (efficiency rule); revisit only if queue claim latency for browser-class tasks proves bad in practice, and say so on this issue if that evidence appears.

Acceptance criteria

• Escalating a fixture incident creates exactly one queue task and one incident issue; a second escalation for the same (source, class) while the incident is open creates neither (dedup demonstrated).
• The queue task packet contains the full context listed above — a browser-enabled agent can start from the packet alone.
• A subsequent producer success auto-comments + auto-closes the incident issue and completes the queue task, with the watchdog's state incident cleared.
• A human-blocked fixture is flagged for NEEDS DREW surfacing in the same tick it is marked, with the reason stated in one line.
• An escalation unresolved at the threshold (24 h default) earns its NEEDS DREW line via the normal sweep path; before the threshold, Drew hears nothing (quiet-while-repairing rule).

Related: #6031 (design), #1260 (escalation split — Drew vs NEEDS ORCHESTRATOR), #2155 (orchestration residue in stewardship).