Adopt validated benchmark checkpoint and resume without mixing incompatible evidence #325

Summary

Target v0.5.3. Integrate the public checkpoint/resume contract from acgetchell/research-repo-tools#168, preserving complete scientific evidence and explicit recovery semantics.

Implementation prerequisite: research-repo-tools#168 (completed), with the subsequent schema-6 fixes in the reviewed development pin. Historical supplier release: research-repo-tools#170. Final supplying publication and normal-package adoption belong to #333.

Implementation handoff — 2026-11-07

The maintainer has moved the published-tooling completion criteria to #333. The implementation and bounded local validation are complete in PR #335, commit a2debb424426b75a97e8a56a061b70985f53d950, using the explicitly authorized research-repo-tools development pin 055353c630604b1ae3438c1eaf534ba85e067a7e. This issue remains open until PR #335 merges after required review and CI. The PR's Closes reference owns automatic closure; published-package adoption and release verification remain separate in #333.

Release-only acceptance is retained, unchecked, in #333: verify the exact supplying PyPI release (target v0.2.7), replace every development pin through normal manifest/lock/isolated-invocation resolution, rerun the affected consumer and native platform/feature checks for that adoption revision, and record the final release evidence. No unpublished pin is accepted for an actual release.

Validation and contract evidence: workload review, candidate review, and #335. Full macOS just ci passed (367 Python tests, 1,143 Rust tests, default/exact doctests and the owning static/native checks); bounded rehearsals used synthetic timing samples, not a real campaign. Required PR CI is still separate from that local evidence.

Historical context

Release measurements span hours. Existing preparation and report/README retries are separate, but arbitrary partial acquisition reuse is not established by those boundaries. The complete current/baseline/peer inventories and required comparison context must remain authoritative.

Proposed Changes

• Integrate only the public installed capability, using the authorized exact development pin for rehearsal. release: plan, publish and verify la-stack v0.5.3 #333 owns final exact-version PyPI resolution. Keep source exports, process ownership, checkpoint schema and generic recovery machinery upstream.
• Declare consumer checkpoint locations and supported recovery boundaries, and expose the canonical documented recovery command without introducing compatibility aliases or hidden automatic retries.
• Resume only units accepted by the shared identity/session contract. Revalidate source/harness/plan, features, fixtures, toolchain/dependencies, sampling, original host/session, raw digests and independent completeness.
• Preserve compatible baseline/current/peer attribution and new-API availability. If a restart changes comparison eligibility, recollect the required comparison unit or retain absolute-only evidence according to the accepted policy.
• Document which edits/context changes invalidate a checkpoint, how a fresh run is requested, and how failed/partial work is distinguished from valid completed evidence.
• Keep rendering, README and artifact-publication retries separate; none may secretly resume or start measurement.

Benefits

Eligible completed work can survive failures while incompatible partial runs cannot be presented as one valid experiment.

Implementation Notes and Acceptance

• Merge feat: resume and reuse verified release benchmark acquisitions #335 after required review and CI; its closing reference then closes this issue automatically.

• Link Astra's upstream checkpoint/session/comparability decision before adoption; request a consumer Astra review if la-stack's paired revisions or baseline-only peers reveal an unsettled scientific contract.

• Transferred to release: plan, publish and verify la-stack v0.5.3 #333: verify supplying PyPI publication, update normal pins/locks and revalidate the released recovery contract in la-stack.

• Focused installed-package fixtures interrupt between units, preserve accepted raw bytes, resume only eligible missing work, and reject changed source/harness/policy, partial samples, incompatible sessions or duplicate ownership.

• Verify final independent inventories and native correctness gates remain required. Generic crash/recovery, atomicity and concurrency regressions belong upstream.

• Run applicable local consumer/native checks with bounded recovery fixtures; feat: resume and reuse verified release benchmark acquisitions #335 retains required PR CI, and release: plan, publish and verify la-stack v0.5.3 #333 owns checks for the final published-package adoption revision. Any long measurement remains a separate release operation.

• Preserve historical evidence and prior promoted outputs on failure. A safe fresh-run fallback is not a successful resume implementation.

Related: #320 (availability) and #324 (progress adoption). Progress is useful but not a technical prerequisite for safe resume. This issue does not authorize retrofitting or interrupting the active v0.5.2 run.

Workstream tracker and PR/dependency plan: #326.