Summary

• Fixes the minion worker poll loop so retryable promoteDelayed() connection failures trigger engine.reconnect() instead of only logging Promotion error: No database connection... forever.
• Extracts the existing claim reconnect path into a shared worker helper, preserving claim's no-inline-retry safety rule.
• Adds a regression test proving reconnect happens before the worker continues to claim work after a promotion-loop connection loss.

Closes #1491.

Why this is not a duplicate

I checked the closest open PRs before editing:

• fix(autopilot): use engine.reconnect() so DB health check doesn't wipe saved config #465 fixes src/commands/autopilot.ts health-check recovery so autopilot calls engine.reconnect() instead of connect() with no config.
• fix(postgres-engine): build-then-swap reconnect() so a failed rebuild can't brick the engine #1906 hardens PostgresEngine.reconnect() itself with build-then-swap semantics.

This PR targets the remaining worker-loop site in src/core/minions/worker.ts: an escaped retryable error from queue.promoteDelayed() was logged and ignored, leaving the worker in the repeated Promotion error: No database connection: connect() has not been called loop described in #1491.

Test plan

• RED: bun test test/worker-promote-reconnect.test.ts failed before the fix with Expected: 1 reconnect call / Received: 0.
• GREEN: bun test test/worker-promote-reconnect.test.ts
  • 1 pass, 0 fail
• Focused regression/adjacent suite:
  • bun test test/worker-promote-reconnect.test.ts test/worker-shutdown-disconnect.test.ts test/minions.test.ts test/retry-matcher.test.ts test/core/retry.test.ts
  • 239 pass, 0 fail
• Post-review focused suite:
  • bun test test/worker-promote-reconnect.test.ts test/queue-lock-retry.test.ts
  • 3 pass, 0 fail
• bun run typecheck
  • pass
• bun run build
  • pass
• bun run verify
  • 29/30 checks passed
  • blocked by existing untouched test isolation lint in test/db-lock-heartbeat-takeover.test.ts for direct process.env.GBRAIN_LOCK_STEAL_GRACE_SECONDS mutation

Autoreview

• Hermes subagent review: clean enough for draft PR. Non-blocking suggestion to assert reconnect-before-claim ordering was applied.
• Codex review against exact diff: clean. It confirmed reconnect-on-promote is safe because promoteDelayed() is idempotent and claim still avoids inline retry.