Repository navigation
RES-644: Persistent research workspace - #3
Conversation
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
|
RES-644 independent engineering review — 2026-11-12Reviewed the eight-commit stack against Confirmed defect: append created derived Parquet directories before the SQLite reference and left them behind after a validation, materialization or transaction exception. The repair removes only the generated directory for an uncommitted attempt. Existing committed revisions and admitted inputs remain byte-identical. Injected Parquet and audit/transaction failures both roll back; retry succeeds. Abrupt process termination can still leave an unreferenced directory; it cannot be selected as a revision. Automatic deletion on startup is deliberately avoided because another process may be materializing there. Reviewed contracts: additive transactional schema v2; versioned entity documents; explicit participant-ID confirmation; source/rights/schema/instrument/calibration/timebase compatibility; separate attachments without pooling; captured revisions and frozen cohort/plan versions; computation outside the preparation transaction; stale completion rejection; transactional report freezing and rendering from frozen context. Historical runs and reports are not rewritten. Validation: nine targeted workspace tests passed. Full Python/scientific/Qt regression on repaired source: 292 passed in 347.60 seconds; Disposition: engineering repair verified by full regression. No new portable/native RES-644 build is claimed; original evidence remains historical. RES-647 will exercise the inherited repaired source through current native/portable workflows. Owner merge authorization remains required. Operational database was read only by existing copy/verification tests. The owner manifest remains untouched, SHA-256 |
|
Final independent review disposition: the append-failure repair PR #3 remains a ten-commit branch against main, with no RES-647 changes introduced into it. Original historical evidence is retained. No statistical algorithm or threshold changed. Handled failures now remove only their own uncommitted derived directory; abrupt OS termination can still leave an unreferenced directory that never becomes selected evidence. Protected content hashes are unchanged. The operational SQLite SHM modification timestamp was refreshed by read-only source-copy tests; its content, DB/WAL content and owner manifest remain checksum-identical. Neither PR was merged. See the updated PR description and linked independent review for the exact evidence and limits. |
Persistent research workspace
One stable workspace retains immutable data revisions, exact scientific run bindings, frozen report versions and revision/run history. This remains the base of stacked draft PR #4.
Independent review confirmed and repaired derived-data directories left by handled append failures. Parquet and transaction failure injection preserve committed revisions and allow a clean retry. Additional checks exercise a genuinely absent v2 table in a v1 store and prove transactional DDL rollback without changing legacy documents. No statistical algorithm or threshold changed.
Validation: 292 passing full Python/scientific/Qt regressions, plus two passing real-v1 migration cases added afterward; all inherited checks also pass in the final 314-test RES-647 suite. Repair commits:
2bfa41ef82b7460a7f0133df942d537047a8d85b,2bdf638305e62137b10fc42520de5701b2c90495.Independent review and full log. Original qualification records remain historical evidence. Abrupt OS termination can still leave a harmless unreferenced directory; no automatic deletion races concurrent writers.
The owner's manifest is byte-preserved. Operational SQLite DB/WAL/SHM contents retain their original hashes; read-only source-copy tests refreshed the SHM modification timestamp only. No installed application was replaced.
Unmerged; owner authorization is required to merge.