Summary

• Scope end-of-build auto-drain to the current build run id (state.launch.runId, falling back to state.slug).
• Keep unrelated halt-event queue rows pending instead of draining or pair-collapsing them during another project's build.
• Add regression coverage for unrelated manual-recovery rows and unrelated DETECTED+RESOLVED pairs.
• Update the fork-local changelog entry for the halt-events drain behavior.

Tests

• PASS: PATH=\"$HOME/.bun/bin:$PATH\" bun test build/orchestrator/__tests__/auto-drain.test.ts build/orchestrator/__tests__/drain-halt-events.test.ts build/orchestrator/__tests__/drain-halt-events-audit-skip.test.ts build/orchestrator/__tests__/drain-halt-events-resolved-pairing.test.ts build/orchestrator/__tests__/manual-recovery-emit-site.test.ts — 23 pass, 0 fail.
• PASS: git diff --check.
• KNOWN BASELINE FAILURES: PATH=\"$HOME/.bun/bin:$PATH\" bun run test:build-skill — 2055 pass, 2 skip, 9 fail. The failures match the pre-existing baseline: Factory EXCLUDED_SKILL_DIRS, Kiro/OpenCode/Slate/Cursor/OpenClaw/Hermes/GBrain dry-run freshness checks, and release_queued without shippedAt/prNumber is detected as manual patch and reset timeout.

Notes

This is fork-local orchestrator work, so it follows the current top-level Unreleased - fork-local changelog convention without a VERSION/package.json bump.