[austinjiann]

Uploads could be saved while no supervisor would ever pick them up, and the run page still said processing was automatic. The dashboard now checks for a live worker that accepts new uploads and shows when a saved run is waiting for a matching worker.

• Register each supervisor independently, refreshing its heartbeat during long agent turns and expiring stale records after 45 seconds. Preserve the legacy heartbeat timestamp.
• Match the actual runs directory and selected run. A single-run watcher or one-pass worker cannot claim readiness for new uploads.
• Offer the existing Queue for later path when a worker is unavailable. Report disconnects without failing runs, hiding completed outcomes, or resubmitting provider work.
• Keep worker liveness separate from provider credentials, spending authorization, and reconstruction quality. No worker is started or broadened by checking status.

#68 is merged. This PR now targets main and contains only the worker-availability changes. Reconciling the squash merge did not change the reviewed source tree.

Validation: 1,390 Python tests and 13 admin tests passed, including upload/API/progress/supervisor integration, and the isolated admin production build, Ruff, repository format check and diff check passed. Browser verification on the existing local dashboard correctly reports no worker accepting new uploads. All four CI jobs passed on aa20e08. Independent review found no actionable defects. No paid provider calls were made.

The current local worker remains restricted to the existing bottle run; this does not enable an unattended paid upload service or certify the bottle reconstruction. That service still needs a generally available worker with its provider access and spending controls configured.