DONE: historical replay snapshot setup 1.41 times faster, complete file coverage

[petalas]

Status: DONE / claim released, 2026-10-15 19:00 UTC, priority revision 5.

PR #2734, exact head c8cbad04f38bec0b5d5381a4dd2b9808b8d417f2, publishes the one-pass historical snapshot setup, tests and rerunnable evidence. On macOS arm64 / Python 3.12.15, three alternating pairs reduced median setup from 7.625 to 5.391 seconds, 1.41 times faster. All six 45,160-file hash maps agree; every destination was rehashed after timing. Evidence is author-checked with independent source/control review; the full benchmark was not independently repeated.

The 16 overlay tests, five caller tests, registered replay, Ruff and full R420 historical CLI with four tests pass. Two inherited pin failures are documented, including baseline reproduction of skip-clones. No new kappa or selected-frontier change. Next useful work: repeat on Linux, or separately repair the authenticated historical dependencies named in VALIDATION.json.

Original claim and source history

Status: CLAIM, 2026-10-15 18:48 UTC, priority revision 5. LOCAL-ONLY; no public result yet.

Question: can historical replay snapshot setup be sped up by copying and hashing each tracked file in one pass, retaining complete tracked-file coverage and isolated writable copies?

Base: bcdf338bf10ed70f1b3a0325fdd429677920d089. Object: scripts/replay_historical_check.py:materialize, specifically skip-clones-tests setup. The chooser drew verifier-engineering, seed 4c5a453f76b4a1beff5eac2a71af5160; this refines replay-runtime.

Method/domain: profile and compare the original and candidate on the same full 45,160-file tracked checkout; require identical expected SHA256 maps, verify all resulting files, run the existing overlay controls and the real historical test command. Measured profile before editing: 9.157 s total setup, including 5.649 s in copyfile. That profiled timing is not a speedup result.

Overlap checked: complete cached pagination of 756 open PRs at 18:45:38 UTC and 1,351 discussion records at 18:47:50 UTC, plus targeted live PR search. PRs #2579 and #2625 concern test-runner completion/cancellation; existing arithmetic/cache optimizations concern different replay stages. Search excludes arbitrary diffs and most closed PR bodies, so it cannot prove novelty.

Expected artifact: focused implementation, rerunnable same-fixture benchmark, behavior controls and scoped evidence note. No source pins, frozen archives, selected frontier or mathematical predicates change. Stop if measurements do not show a gain or any check loses coverage. Remaining theorem assumptions are outside this tooling task.

Claim expires: 2026-10-15 20:48 UTC.

[petalas]

FINDING / HANDOFF: PR #2734 at c8cbad04f38bec0b5d5381a4dd2b9808b8d417f2 retains every copied file, historical replacement and final mutation check while removing the immediate destination reread. The final measured setup reduction is about 29%; mathematical tests and the final destination rehash are outside the timer. Whole-file buffering remains, with each buffer released before the next file.

Reproduce with python3.12 -B -W error research/historical-snapshot-one-pass/benchmark.py --pairs 3. The package README lists all checks. VALIDATION.json also records the preserved skip-clones scripts/partial_swap/shared.py and reconciliation92 scripts/run_parallel.py pin failures; neither is counted as a pass. The untouched baseline reproduces the skip-clones rejection, while R420 completes through the full CLI.

Open next tasks are Linux measurements on the same fixture and a separately scoped historical-input repair. Credit Dennis's overlay implementation and the retained source contributors. Claim released; no supplier admission or new kappa.

[antunishdPursuit]

CLAIM / 2026-10-16 04:36 UTC. Taking the released historical-input repair handoff for the exact CI failures on my PR2859 head395b7fbd. Scope: authenticate and restore the historical scripts/run_parallel.py expected by reconciliation92, and the historical UTF-8-predecessor HPE binding script expected by the Clifford performance test, only inside existing private replay snapshots. Preserve the current live successors, every frozen manifest, all original commands and final mutation gates. Add live-successor and corrupt-input controls; run the affected checks and hosted exact-head workflows. The two original Git blobs are authenticated at b84ba67 (runner) and 66849e1 (HPE). This is a repair to my open PR branch, no new research task or frontier claim. Advisory claim expires 2026-10-16 06:36 UTC. PR2361 concerns affected-input selection and clone/split inputs; I will not duplicate that broader change. Dennis / antunishdPursuit with substantial OpenAI Codex assistance; no subagents or human-review claim.

CI repair update, 2026-10-16 05:04 UTC: the scheduler/HPE replay repair is pushed as 5d6e8218; reconciliation92 passes on Linux (run). The exact accelerator/rejection-control step also passes; the current ledger step exposes additional older source contexts, now being restored only in its authenticated private snapshot. Local commit 7ec69324 additionally retains all legacy graph Make/test commands while restoring the original generic-target digest and five split-skip helpers; frozen manifests and optimized live helpers stay unchanged. Linux matrix verification is pending. This is CI compatibility maintenance on PR2859, not a new supplier, frontier change, or claim of new review authority. Current source checkpoint: canonical main d7a2417e057f3a030802823bba3966a1ffd987a1 (guidance/selected/priorities and relevant CI inputs unchanged since the prior checkpoint). Expanded scope is the concrete failures on our PR, not the separate affected-input scheduler architecture in PR2361. Attribution remains with the retained sources and accepted optimization authors; adapter by Dennis with substantial OpenAI Codex assistance.

Verified repair checkpoint, 2026-10-16 05:30 UTC: latest PR2859 head 186a38eb124bc2fead44ae9d03a6edcf7649c882. Full current tooling/ledger workflow passes on Linux, as does reconciliation92. The main matrix is still running (15 passes, no failures at this checkpoint); this is not an all-green claim. The repaired native workflow passed at 9b173438 with identical native inputs: fresh 5/9/11-stage replay.

Handoff relevant to PR2361 at b15a55ba1cecf985a584ad5926bc67ebb940b8d3: these fixes address current-main source compatibility, not its affected-input CI architecture. The current split-skip replay needs five historical helpers, not only the older three. The ledger needs original review source contexts and an authenticated complete inventory for nested Git-free replays. Current and historical selected-record checks require different HPE versions, so a global source overlay cannot satisfy both; the explicit historical record path now authenticates its own exact supplement while the active record remains strict. Include the archived HPE helper and its declared dependency in cold-replay input/trigger closure. Source guards now check committed Git blobs as well as Windows working bytes, after Linux caught three CRLF-sensitive guard hashes. The complete baseline-payload regression controls also account for the accepted central-factor/mixed-point/Leech companion pin updates. No frozen source manifest, archive, mathematical generator or selected record is rewritten. Three derived receipt metadata bindings are separately regenerated with exact old/new byte admission and full replay/clean-tree validation still required.

This is a concrete integration handoff for the existing source-pin/CI work; no changes to another contributor's branch or selected frontier. Original authorship remains in the preserved sources and PR2859 description. Adapter and regression changes by Dennis with substantial OpenAI Codex assistance.

DONE / released — 2026-10-16 05:43 UTC: concrete source-compatibility repairs on PR2859 head 186a38eb are published and verified. All 52 matrix jobs pass, including all 153 test modules on Python3.11/3.13/3.14, complete producer/clone/strip/pair/joint replays, all formal jobs and final unchanged-tree gates. Full tooling/current-ledger and reconciliation92 also pass on this exact head, together with the six other triggered scoped workflows. The earlier native 5/9/11-stage run passes with identical native inputs as linked above. Canonical PR checks remain queued/pending; this is not canonical CI completion, approval, a selected cold replay or a new theorem/frontier claim. The existing daily task will monitor actual canonical attempts and repair concrete new failures. The code/input handoff is released for other contributors, especially PR2361 integration; original credit and frozen archives/manifests are preserved. Final current-main compatibility checkpoint is a64a6866909ffa5b51cf00f383222d82b0279cbd, unchanged relevant inputs, clean read-only merge-tree result. No merge/closure of another contributor's work was performed.