Retain invalid integration IDs as unresolved findings

[Vilin97]

A large review aborted in its fourth rubric when the integration model cited an unknown obligation ID. Preserve that diagnostic and any unresolved real obligations in a discussion verdict instead of aborting the entire run. Unknown IDs cannot support approval, including an ID that tries to name its own diagnostic.

Also checkpoint each completed rubric so a later failure retains its integrated evidence, coverage, and ordered source reports.

Validation: all 381 Python tests pass; Ruff check/format and diff checks pass. Regressions cover unknown IDs, attempted self-resolution, and retention of a completed rubric after a later failure.

[greptile-apps]

Advisory cross-file review against Lean Pool's supervisory rubric. This is not by itself a significance, source-verification, or merge verdict.

The PR appears safe to merge, with a non-blocking reliability improvement recommended for atomic evidence checkpoint replacement.

Findings

1.  Checkpoint Can Be Corrupted ▶

Summary

• Unknown IDs now force a discussion verdict and remain visible as diagnostics.
• Completed rubric results are persisted before subsequent rubric execution.
• Regression tests cover unknown IDs, self-resolution attempts, and later rubric failures.

_{Reviews (1) · Last reviewed commit: "Retain unknown integration IDs as unreso..."}

[greptile-apps]

Checkpoint Can Be Corrupted

Each completed rubric rewrites the evidence file in place. If a later write is interrupted or fails after truncating the file, the valid earlier checkpoint can become malformed or incomplete, defeating the new recovery guarantee. Write to a temporary sibling and atomically replace the artifact so the last complete checkpoint survives.

Knowledge Base Used:

• Python maintenance tooling
• Review and advisory automation