feat(backend): implement graceful shutdown with drain, flush, and connection teardown (#349)

[prissca]

Summary

setupGracefulShutdown() already existed and was unit-tested, but src/index.ts's actual production entry point (main()) never called it — it had its own, much simpler inline SIGTERM/SIGINT handler that only closed the HTTP server, with no draining, no DB closes, no task/agent state cleanup. This PR wires the real function in and fixes the gaps that made it unsafe to rely on:

• api/app.ts's close() now awaits jobWorker.stop(jobWorkerStopTimeoutMs) (new AppOptions field, default 10s) before closing the HTTP/WS server — previously it called jobWorker.stop() without awaiting it, so "drain" did nothing: the server closed immediately regardless of in-flight work.
• Removed the Phase 3 taskDb.failRunningTasks() call from setupGracefulShutdown, which force-marked every running task's DAG nodes as failed on shutdown. That directly worked against the job queue's own resumption path: JobWorker.start() already calls recoverIncompleteJobs(), which resets any job still "active" from a previous run back to "pending" for retry — a job stuck mid-drain is meant to resume, not be declared dead. (failRunningTasks() itself is left in db/tasks.ts in case it's useful elsewhere; it's just no longer called from the graceful-shutdown path.)
• Added Phase 4: eventBus.store.close() — the event store was never closed on shutdown even though every other DB was.
• src/index.ts's main() now calls setupGracefulShutdown(httpServer, close, config, { cleanupService, reconciliationService, globalAgentRegistry }) instead of its own inline handler; setupGracefulShutdown's signature gained an optional 4th extras argument (backward compatible — the existing 3-argument test call still passes) to stop those three services, which main() previously stopped inline but setupGracefulShutdown had no way to reach before.
• New test in src/queue/worker.test.ts: a job still "active" when a JobWorker.stop() call times out is picked up and completed by a fresh JobWorker instance over the same store — the restart-mid-stream scenario the acceptance criteria asks for, exercised at the layer that actually owns resumption (recoverIncompleteJobs()).
• tests/shutdown.test.ts updated: asserts failRunningTasks/createTaskDb are no longer called, asserts the three extra services are stopped, asserts the event store and job DB are closed, and adds a backward-compatibility test for the 3-argument call.

Necessary prerequisite: several backend files were corrupted by a bad merge

package.json, jest.config.js, tsconfig.json, config/index.ts, api/app.ts, api/routes/stream.ts, api/routes/agents.ts, api/routes/stats.ts, and api/routes/health.ts each contained two full, conflicting versions of their own content concatenated together, which blocked npm install/npm test outright. Same root cause and fix already documented in Epta-Node/ai-net#443 (a different fork, issue #359) and this repo's #460 (issue #353) — kept the newer half matching actual codebase usage in each file, discarded the stale duplicate. This PR's health.ts fix is the minimal corruption fix only (no /live or /ready extensions — those are #353's PR, #460); api/app.ts here additionally includes the close()/job-worker-drain change described above, which #460's app.ts deliberately does not.

Acceptance Criteria

• In-flight tasks complete or resume on restart
• E2E test validates restart mid-stream

Test plan

• npx jest tests/shutdown.test.ts — 5/5 passing (full phase sequence, extras stopped, no more failRunningTasks, event store + job DB closed, 3-arg backward compatibility, forced-exit-on-timeout).
• npx jest src/queue/worker.test.ts — 10/10 passing, including the new restart-mid-stream test.

Closes #349

[drips-wave]

@prissca Great news! 🎉 Based on an automated assessment of this PR, the linked Wave issue(s) no longer count against your application limits.

You can now already apply to more issues while waiting for a review of this PR. Keep up the great work! 🚀

Learn more about application limits

[vercel]

@Obiajulu-gif is attempting to deploy a commit to the Jaja's projects Team on Vercel.

A member of the Team first needs to authorize it.