feat: event-driven thread wait coordinator (BM-52)

[jonolee-kr]

Summary

Fixes excessive memory and request load from bb thread wait (BM-52).

• Server: new reference-counted wait coordinator keyed by (threadId, target, cursor). Identical waits share one NotificationHub subscription and one coalesced DB re-check. Per-caller deadlines, abort cleanup, caller limits, pino observability with an unref()ed summary timer, and a stats accessor.
• Route: new GET /api/v1/threads/:id/wait?status=... long-poll (404 thread_not_found for missing threads). The existing events/wait route is refactored onto the coordinator with its response contract preserved.
• SDK: status waits long-poll the new route. Fallback to the legacy 250 ms polling loop triggers only on BbHttpError status 404 + code not_found (route absent on older servers). The wait sleep is now abort-aware. pollIntervalMs governs only the fallback path.
• CLI: bb thread wait --output (one blocking follow returning terminal status + final output), bb thread wait-many, and lazy-loaded thread subcommands. Existing wait messages, JSON shapes, and exit codes unchanged.
• Guidance/docs: builtin skills no longer teach log/wait alternation; changelog documents rollout compatibility; QA runbook adds PSS measurement.

Measured impact

• Old CLI waiter: 150–174 MiB RSS / 99–122 MiB PSS each, 4 status GETs/s per waiter (~6 DB ops each).
• New CLI waiter: 48 MiB RSS / 12 MiB PSS (standalone).
• Live demo against this branch's server: 20 concurrent CLI waiters → 1 coordinator entry, 20 total HTTP requests for the whole wait, one shared DB re-check on the status flip, all resolved within 77–133 ms, zero leaks. Old CLI in the same harness: 3.8 GETs/s per waiter.

Testing

• apps/server: 2083 passed / packages/sdk: 102 passed / apps/cli: 513 passed, 0 failed, no new skips (independently re-run by a second reviewer).
• Five deliberate mutations (broken dedup, shared abort, broad fallback, non-abortable sleep, wait-after-output) are each caught by the new tests.
• 100-duplicate-wait stress tests assert one entry, bounded checks, and full cleanup.

Plan, review artifacts, and demo report are attached to task BM-52.

🤖 Generated with Claude Code

[jonolee-kr]

Superseded by upstream PR get-bb#2688.