feat(background_tasks): worker status admin page

[antosubash]

Summary

Adds a read-only /admin/background-tasks/workers sub-page so operators can see whether a Celery worker is connected to the broker and what each worker is doing — closes the visibility gap where "tasks stuck in pending" was indistinguishable from "no tasks enqueued yet."

• New WorkerInspector wraps celery.control.inspect() with a probe-then-inspect flow that disambiguates broker-down from broker-up-with-no-workers. Catches OperationalError, RedisError, ConnectionError, OSError and surfaces them as WorkerSnapshot(broker_reachable=False, error=...) instead of 5xx-ing the page.
• Two endpoints: Inertia view at /admin/background-tasks/workers (initial render) and JSON at /api/background_tasks/admin/workers (Refresh button uses plain fetch). Both run the inspector via asyncio.to_thread so the broker round-trip never blocks the event loop.
• Workers.tsx page with three states — broker unreachable, no workers connected, list of worker cards (hostname, online dot, queue badges, active/pool/processed counts, software). Refresh button shows a sonner toast on failure, mirroring the existing retry pattern.
• Discoverability: a "Workers" button in the existing Index toolbar (no new sidebar entry per the design).

Design choices worth flagging

• Why a separate page (not a banner on Index): keeps the executions table focused on tasks; lets us show per-worker detail without crowding the toolbar.
• Why manual Refresh, not auto-poll: inspect.ping() is a broadcast over the broker; polling every few seconds adds steady chatter for a page operators only glance at. One-click verify is intentional.
• Why probe + inspect (extra round-trip): without the probe, inspect.*() returns None for both "broker down" and "no workers replied" — operators couldn't tell which.
• Why redis.RedisError is caught explicitly: redis.exceptions.TimeoutError / ConnectionError don't inherit from built-in OSError or ConnectionError, so they'd otherwise escape and become 500s. Verified by an authentication-error test case.

Test plan

• make test — 1054 Python + 8 JS tests pass (33 in modules/background_tasks/tests/ including 4 new inspector unit tests + 4 new endpoint tests covering auth, snapshot payload, broker-unreachable response, and Inertia render).
• make lint — ruff format/check, ty, biome, tsc, 300-line cap, doctor checks (SM003/SM004/SM018/SM019), metadata, READMEs all clean.
• make doctor — no new warnings.
• Manual smoke (deferred to reviewer): with make dev running, exercise the three UI states by stopping the broker / starting it without workers / starting uv run python scripts/run_worker.py. Confirm the Index toolbar "Workers" link round-trips.

Notes for the reviewer

• The plan and spec are committed at docs/superpowers/{specs,plans}/2026-05-01-background-tasks-worker-status-page-*.md for context.
• A small refactor commit at the end consolidates the _stub_celery test fixture into a module conftest (opt-in via pytestmark.usefixtures so signal tests still get an unstarted app), and lifts WorkerInfo/WorkerSnapshot/formatTs into pages/constants.ts so Detail.tsx and Workers.tsx share them.