fix(server): stop one wedged provider stop from stalling every queued prompt

[tusharbhardwaj-bk]

Summary

bkmain outage 2026-08-20 10:41–12:02 UTC: every new prompt stayed Queued. Root cause and fix:

• An unanswered Claude AskUserQuestion outlived its turn. ClaudeAdapter.stopSessionInternal cancels pendingApprovals but not pendingUserInputs, then waits on Fiber.interrupt(streamFiber) with no bound; the stream fiber sits in the uninterruptible Stream.fromAsyncIterable finalizer behind the SDK's pending canUseTool promise.
• The session reaper wedged first (holding the per-thread lifecycle lock). The owner's later Stop request wedged the ProviderCommandReactor's single sequential worker lane, and every turn-start-requested wake / session-set runDue queued behind it.
• DurableExecutionCoordinator only learned about work through that lane → parked in Queue.take with runnable rows in the table.

Changes

• ClaudeAdapter: resolve pending user inputs on stop; bound the stream-fiber interrupt (5s) — query.close() still ends the runtime.
• ProviderCommandReactor: bound providerService.stopSession (15s) in processSessionStopRequested so one provider cannot hold the lane.
• DurableExecutionCoordinator: every idle wait is bounded by a 10s safety poll; a lost wake costs one poll, not an outage.

Test plan

• New regression test: work accepted with no wake is claimed by the safety poll (drives only the TestClock). Times out on the previous coordinator.
• DurableExecutionCoordinator.test.ts 9/9 green; apps/server typecheck green.
• CI (deploy-bkt3.yml) green → auto-deploy to bkmain → queue probe stays GREEN and Stop on a session with a pending question returns within seconds.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}