fix(research): preserve source-version review checkpoints

[harlan-zw]

🔗 Linked issue

Related to #598

❓ Type of change

• 📖 Documentation
• 🐞 Bug fix
• 👌 Enhancement
• ✨ New feature
• 🧹 Chore
• ⚠️ Breaking change

📚 Description

Source refreshes and failed checkpoint writes can discard completed description evidence.

Match author descriptions and hashes against current renders or retained snapshots. Save the original blob SHA and replace checkpoint files atomically.

🤖 AI disclosure: Harlan Agent Kit modified this description. My AI open-source policy.

[harlan-github-agent]

🤖 MERGED

Harlan Agent Kit posted this automated review. It is not Harlan's personal review or approval. AI open source policy. Last updated: 2026-10-09 05:29 UTC.

GitHub merged this pull request.

• Logged (28/100): A malformed rendered_frontmatter value matching the identity can error the whole INSERT..SELECT statement because SQLite AND is not short-circuit guaranteed. View code Optional follow-up: Wrap extraction as CASE WHEN json_valid(rendered_frontmatter) THEN json_extract(rendered_frontmatter,'$.description') END (or use a json_extract that tolerates invalid JSON) in the current predicate.
  • Ask an agent to verify and fix this finding
• Logged (12/100): The PR's only attached evidence image is unavailable (HTTP 400), so its depiction of the checkpoint flow is unverified documentation. View code Optional follow-up: Re-attach the diagram to the PR or add it under docs/ops/ so the checkpoint flow has reviewable evidence.
  • Ask an agent to verify and fix this finding

c778d82e-7f37-41f0-800b-c27019207034

Selected findings run after merge and open separate pull requests.

[harlan-github-agent]

Harlan Agent Kit posted this automated review. AI open source policy.

Automated Review: READY.

Read the Review results

[harlan-zw]

🤖 Harlan Agent Kit Agent verification of head f082b01.

A cached-only CLI run with a 1 KiB file-size limit reproduced a damaged checkpoint before the fix. The same failed write preserved the previous complete checkpoint after atomic replacement. Both runs made zero inference requests.

Three source versions refreshed during the research run. D1 retains their original hashes and descriptions. The equivalent current-or-history guard saved all 12,141 frozen size observations; the current-only guard saved 12,138.