fix(store): prune run checkpoints and preserve SQLite delta baselines

[DeliciousBuding]

Summary

Closes #2335
Refs #2333 / #2256.

Two production-line fixes at existing ownership boundaries:

• Remove a run's checkpoint from the shared in-memory evidence-pruning path used by terminal cleanup and thread deletion.
• Include a deep-cloned checkpoint map in SQLite's committed delta baseline, so unchanged checkpoints are not rewritten and removals become durable row deletes.

No pool/PRAGMA/schema/API/timer changes. Update the Edge owner document for checkpoint lifetime and the actual finished terminal run status.

Regression evidence

Corrected red output on mainline, followed by green behavior.

• Real SQLite audit trigger: unrelated run mutation rewrote two unchanged fixture checkpoints before the fix; now zero. A legitimate checkpoint content edit updates only its own row and is durable.
• Terminal-run cleanup and thread deletion: old code retained the checkpoint in memory, in the database and after reopen; the fix removes it while preserving an unrelated queued run and checkpoint.
• Readback happens before a final Close/Flush could hide a missing durable write.
• Existing clone test now covers nested checkpoint Files ownership.

Validation

• New behavioral regressions pass on Windows with pure-Go SQLite.
• Full store-package Linux race: PASS (4.996 s).
• Full store-package Windows tests (CGO_ENABLED=0): PASS (1.569 s).
• go vet ./internal/store, staticcheck 2026.1, verify-doc-ssot.py, and diff check pass.
• The test(store): benchmark snapshot persistence and SQLite facade contention #2334 benchmark fixture is unchanged. All 45 matching non-race Linux samples pass: unchanged-checkpoint row amplification falls from 13 / 103 / 1003 to 3. At 1,000 runs without readers, the local median mean write time is 62.315 → 39.663 ms; writer-only allocation slightly increases, not decreases. Full raw before/after evidence is on fix(store): prune run checkpoints and persist their delta state #2335. All seven required CI gates passed.

Boundaries

Only isolated generated fixtures. No production database, deployment or migration. This prevents future orphan retention through the corrected removal paths; it does not automatically repair checkpoints already orphaned in an existing live store. Any existing-data repair is a separate scope. Allocation/timing comparisons are local workload results, not production guarantees.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Team

Run ID: f1efd785-46cf-4366-ae29-bc377378185f

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[DeliciousBuding]

Final-head review at 350ee4fba3bc938e0fe472fe059218724fcbed40: live strict branch protection still requires exactly validate, go-edge, go-hub, windows-go, windows-frontend, backend-required, frontend-required; all seven are present and SUCCESS. Merge state CLEAN. No submitted reviews or unresolved threads; a skipped automated review is not counted as approval.

The two production lines, real SQLite negative/positive regressions and owner-document change were reviewed together. Full local package/Windows/race/static/doc validation and 45 matching-fixture post-fix samples pass. No migration, existing orphan repair or runtime deployment.