[iscekic]

Summary

The scheduler's settled-run sweep looked for running decider runs with no pending, queued or active chunk. Chunks of a dead-lettered lane can stay pending forever, so a run whose only remaining work belonged to dead lanes was never finalized: no live chunk was left to trigger per-completion finalization.

The sweep now uses the same unsettled-chunk rule as hasUnsettledChunks, which ignores chunks of lanes recorded in run_lane_failures. The rule is one shared SQL fragment, so the two checks cannot drift apart.

Verification

• Ran pnpm typecheck, pnpm lint and the full vitest suite in services/auto-routing-benchmark on a work VM. All passed.
• New regression test: a dispatch tick finalizes a running decider run whose only pending chunks belong to a dead lane. The dead entry's profile fails, the live entry's profile becomes ready, and the run completes. The test fails when the old sweep condition is restored.
• The existing query-plan test now also checks that the sweep scans neither scheduler_chunks nor run_lane_failures.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Reviewed the extracted shared unsettled-chunk SQL predicate and the new listSettledSchedulerRuns sweep. The predicate matches the previous hasUnsettledChunks semantics exactly, the sweep preserves the prior run selection (kind='decider', status='running', scheduler-initialized) while ignoring dead-lane chunks, and query-plan/test coverage for the new SQL looks sound. No security, correctness, or performance problems found in the changed lines.

Files Reviewed (4 files)

• services/auto-routing-benchmark/src/scheduler.ts
• services/auto-routing-benchmark/src/run.ts
• services/auto-routing-benchmark/src/scheduler.test.ts
• services/auto-routing-benchmark/src/scheduler-initialization.test.ts

---

_{Reviewed by deepseek-v4.1-flash-1017 · Input: 83.4K · Output: 19.5K · Cached: 1.6M}

_{Review guidance: REVIEW.md from base branch main}