Make failure cleanup reliable by design

What changed

The runner now has a separate bounded cleanup phase. Previously, placing cleanup at the end of the fail-fast verification list meant an earlier failed test could skip cleanup and leave ignored build artifacts behind. The existing safety gate then rejected the workspace as mutated.

The fix always runs the explicitly configured cleanup list after verification, even if a check fails or times out. It does not turn a failed check into success; it preserves the failure and leaves cleanup failures visible to the gate.

Verified checks

• Server-loop tests: 38 passed, including real generated-artifact cleanup after failure.
• Configuration migration: 3 passed; agent registry: 4 passed; free-provider worker: 13 passed; verified proposal: 3 passed; launcher: 4 passed.
• Node tests and JavaScript syntax checks passed.
• Shell syntax and JSON configuration validation passed.
• Patch apply check and diff whitespace validation passed.

Download the tested patch · Read full engineering report · Machine-readable evidence

Limits and next gate

This is a tested patch candidate, not a live deployment. The current checkout is on a feature branch although the launcher requires main; the restricted service user cannot write the shared Git metadata needed for worktrees; and upstream authentication has not been established. The loop's last recorded state remains cycle 669.

After source synchronization is repaired, deployment must first pass an exact service-identity fetch/worktree smoke test. Then a new live cycle must produce a complete evidence bundle, leave the disposable workspace clean, and preserve production HEAD. Until then, no recovery or model-improvement claim is warranted.

This release contains no raw operational traces, private prompts, credentials or private keys. No production configuration, model weights or production graph were modified.