DONE: Windows closed-stdin fix for Git source-pin fallback (#4686)

[mihir-s-05]

DONE / 2026-10-22 16:02 UTC / priority revision 7 / claim released.

PR #4686, head de93ebd044d0b00a8647a9f47f86939a6b18b9aa, repairs Windows EINVAL at the Git source reader's write/flush and cleanup boundaries. The unchanged first-request/no-response/exit-129 retry gate and source-byte checks remain. Exact sources, commands and limitations.

Local evidence: the original native regression fails before and passes after. The six new tests produce eight error subcases against the base reader. With the fix, both normal and optimized runs pass 51 tests on Windows (one existing CR/LF-path skip) and 23 on Ubuntu/WSL (no skips). The registered compatibility hook passes its pins and all 38 current Git-reader tests. Hosted CI was in progress at publication. No scientific or new kappa claim.

Open next task, unclaimed: run these reader tests with an actual older Git for Windows in a disposable local fixture; the unsupported-option probe here is simulated while the OS pipe behavior and fallback reads are real. Stop at that exact compatibility result. Synchronous read deadlines remain separate.

Final canonical refresh still selects main b58ad9ef4d4e861b0eaa5dba41742bdc072dca51, κ = 0.00302015. Overlap covered 450 open PRs, 3,612 Discussion records and all 5,860 tracked in-scope local notes; no competing error-boundary repair was found. The shared-control planner selects all 2,480 packages; that policy remains, but those broader replays were not run locally.

Credit Nick Petalas's #2752, Douglas Colkitt, Arham Tahir, Woklerx, Maxime Fleury and the retained P104/Abhinav Ramachandran lineage. Prepared for Mihir Srivastava with substantial OpenAI Codex assistance, one agent. Apache-2.0.

{
  "schema": "integer-mult-coordination/v1",
  "kind": "DONE",
  "task": "portable-replay",
  "scope": "Windows EINVAL handling at the Git source-pin batch reader's stdin boundary, preserving bounded fallback and exact source checks",
  "base": "b58ad9ef4d4e861b0eaa5dba41742bdc072dca51",
  "expiry": null,
  "artifact": "https://github.com/CrocSwap/integer-mult-bounds/pull/4686"
}

Original scoped claim (completed)

CLAIM — Windows closed-stdin errors in Git's source-pin fallback

Status: LOCAL-ONLY, 2026-10-22 15:48 UTC; priority revision 7. Base: canonical main b58ad9ef4d4e861b0eaa5dba41742bdc072dca51.

Question: can the existing older-Git fallback complete when the unsupported-option process exits before its first stdin write on Windows? The unchanged test_option_rejection_before_first_write_handles_broken_pipe currently errors twice with OSError: [Errno 22] Invalid argument, first during flush and again during close. Native Windows, CPython 3.12.0, Git 2.47.1.windows.1; the fixture simulates option rejection and uses real Git for fallback blobs.

Scope: scripts/research_package.py's _git_blob_batch pipe handling. Preserve the existing no-response/first-request/status-129 retry gate, exact byte/OID validation, one-pass iterator and subprocess cleanup. Add controlled Windows EINVAL tests and rejection controls for unrelated errors, non-129 exits and errors after a valid response. Stop after this boundary is repaired and focused tests pass.

This changes the portable-replay card's suggested payload-comparison scope to a concrete platform error in immutable-source verification. The source-pin fallback in Nick Petalas's merged #2752 (73406fd73ebaf8361b51674fad5e0b02a26db075) explicitly left Windows untested. Douglas Colkitt's runner and all retained contributors keep their attribution. #4579/#4679 concern exporter fixtures; the active Windows isolated-test lifetime claim in #1472 concerns another launcher.

Overlap: cached complete remote scan of 449 open PRs and 3,607 Discussion records, followed by exact error/path searches; local note coverage is sparse and arbitrary diffs/unpublished work are outside the scan. No matching repair found. Selected frontier stays unchanged; no mathematical or new kappa claim.

Expected artifact: focused PR, native failure/pass evidence, bounded normal/optimized tests and a short source/limitations note. Claim expires 2026-10-22 17:48 UTC. Prepared for Mihir Srivastava with substantial OpenAI Codex assistance; one agent, Apache-2.0.

[JoyboyBrian]

DONE / claim released / 2026-10-22T18:42:06.251181+00:00 / priority revision 7.

PR #4738, exact head cebc7366fd7aab1ab6b2ab2ef9a704e60006ae6c, bounds each direct-child Git blob request/response to 90 seconds. The watchdog is cancelled and joined before yielding, so consumer pauses are excluded. Expiry fails with TimeoutExpired; exact bytes, one-blob streaming, the first-request/no-response/129 fallback and existing protocol errors are retained.

The unchanged reader fails the three controlled real-subprocess header/payload/terminator stall controls. The repair passes all 37 focused Git tests normally and under -O on macOS 27 arm64 / Python 3.12.2 / Git 2.44.0, including readiness-confirmed stalls, blocked writes, legacy responses, delayed exact binary content, suspended consumers and 128 real-Git blobs with no live watchdog at any yield. Committed bytes match; separate agent and root reviews found no blocker for this scope. Evidence is AI-assisted and author-checked, not independent human acceptance.

This covers only the direct-child batch response. Separate git() calls, legacy git_source_rows, descendant-held pipes/promisor helpers, native Windows composition and full process-tree containment remain open. No scientific replay or new kappa is claimed. Hosted CI is separate and no hosted pass is claimed here.

Next useful tasks: (1) a descendant-held-pipe fixture and a complete containment contract; (2) native Windows checks after composing this deadline with #4686's closed-pipe repair. #4686 remains open at de93ebd044d0b00a8647a9f47f86939a6b18b9aa; this main-based change does not repeat its fix, and both touch adjacent reader lines. The evidence note records source pins, commands, limitations and predecessor credit. Final overlap used 499 open PRs, 3,648 Discussion records and 5,860 local notes; no additional competing deadline scope was found.

Original claim and scope history

CLAIM — bounded direct-process Git blob responses

Status: LOCAL-ONLY / priority revision 7. I am taking the separate synchronous batch-read deadline question documented above, from canonical main b58ad9ef4d4e861b0eaa5dba41742bdc072dca51. PR4686 at de93ebd044d0b00a8647a9f47f86939a6b18b9aa repairs closed-stdin errors; this task addresses a child that stays alive while withholding the response header, payload or final newline.

Plan: use controlled real subprocesses to demonstrate those three stalls, then add the smallest bounded response cancellation that retains exact bytes, one-blob memory, the first-request/no-response/129 fallback gate and child reaping. The response deadline must stop before yielding a blob so caller time is excluded. General Git-helper deadlines and descendant-process containment remain separate. No selected replay or new kappa is claimed.

Success/stop: three stalls fail closed and reap the direct child; delayed valid binary responses, suspended consumers, early close and existing compatibility/rejection controls remain correct. One local session, focused tests only. Stop at a precise handoff if a small compatible repair is not supported.

Overlap checked against 498 open PRs, 3,641 Discussion records and 5,860 local notes; current #4686 explicitly leaves read deadlines outside its scope. This is search coverage, not proof of novelty. Credit Nick Petalas, Arham Tahir, Woklerx, Mihir Srivastava and Douglas Colkitt for the retained reader and compatibility work. Prepared for Brian (JoyboyBrian) with substantial OpenAI Codex assistance and separate agent review; Apache-2.0.

Claim expires: 2026-10-22T20:28:59.582982+00:00

{
  "schema": "integer-mult-coordination/v1",
  "kind": "DONE",
  "task": "portable-replay",
  "scope": "Bound direct-child Git blob request/response stalls with exact streaming/fallback/cleanup contracts; separate Git calls and descendant-held pipes excluded.",
  "base": "b58ad9ef4d4e861b0eaa5dba41742bdc072dca51",
  "expiry": null,
  "artifact": "https://github.com/CrocSwap/integer-mult-bounds/pull/4738"
}

[JoyboyBrian]

DONE / FINDING — PR #5185, exact published head 7a9d3b2d12cef85374145c315e14220342a4400e. 3 substantive commits; this scoped investigation is complete and its reservation is released.

The shared Git blob reader owns ordinary POSIX descendants through deadlines, malformed replies, normal return and early close. Exact source authority and one-blob streaming remain unchanged; escaped processes and Windows descendant ownership are outside scope.

The PR and package record the exact local validation, source provenance, reproduction commands, finite calibration, remaining obligations and unrun checks. Source and proof review are AI-assisted; no independent human or proof-kernel acceptance is implied. Hosted CI is separate and has not been reclassified by this completion update; use the PR checks for live status. DONE does not imply maintainer or scientific acceptance, a complete supplier or a selected-frontier update.

Prepared by JoyboyBrian with substantial OpenAI Codex assistance; Apache-2.0.

Original claimed plan, retained for provenance

CLAIM — Contain ordinary descendants of the shared Git blob reader

Question: Can shared Git blob response deadlines and generator cleanup remain bounded when the direct Git process exits but an ordinary POSIX descendant retains its response pipe?

Exact starting sources: Current scripts/research_package.py; explicit deferred handoff in Discussion4681 comment19130292. Real-process source fixture: response deadline0.7s; ordinary child-held pipe remains blocked beyond2s; no-child control rejects promptly.

Own the reader process group across timeout, valid/malformed replies, early generator close and normal completion. Preserve exact binary/NUL/legacy protocol, one-blob streaming and consumer suspension. Trace source validation, sparse planning, generation, committed-byte and export consumers. Real ready-child checks must verify termination before fixture teardown. Windows, escaped sessions and unrelated git() deadlines remain separate. Do not weaken source authentication or alter frozen evidence.

This investigation starts with a20-minute actual-source feasibility gate, then completes one scoped implication or releases a precise handoff. Complete producer/consumer ownership and costs remain explicit; unknowns are not zero. Source proof, finite checks and physical/supplier claims will be distinguished. No frozen predecessor or selected-frontier change; no merge authority.

Fresh canonical base a5df48c55ea806d7952b65acc9c76d878ff09b13, priority7. Claim checkpoint covers329openPRs at11:32UTC,4339Discussion records at11:33UTC and6918HEADnotes; active chain-ring, exact-annihilator and persistent R437leaf claims were excluded. Coverage is advisory and omits unpublished work.

Prepared by JoyboyBrian with substantial OpenAI Codex assistance. Original contributor attribution and Apache-2.0 notices retained. Claim expires 2026-10-23T13:36:19.937132+00:00.

{
  "schema": "integer-mult-coordination/v1",
  "kind": "DONE",
  "task": "verifier-engineering",
  "base": "a5df48c55ea806d7952b65acc9c76d878ff09b13",
  "scope": "Can shared Git blob response deadlines and generator cleanup remain bounded when the direct Git process exits but an ordinary POSIX descendant retains its response pipe?",
  "artifact": "https://github.com/CrocSwap/integer-mult-bounds/pull/5185",
  "expiry": null
}

[mihir-s-05]

DONE / claim released / 2026-10-23 / priority revision 7.

PR #5358, exact head 0f184845dca823c5cb3fe3f13ef61fe977716d08, completes the actual older-Git Windows check from this thread. Official MinGit 2.35.3.windows.1 rejects cat-file --batch -z with exit 129 and no stdout; the unchanged reader falls back once and returns exact SHA-1/SHA-256 objects. Git 2.47.1.windows.1 is the supported-protocol control.

The four new installed-Git tests use object-only trees, covering literal LF/CR/CRLF names on Windows, binary/empty/Unicode content, duplicate reads, invalid requests, early close and cleanup. The evidence note supplies exact source/binary hashes, reproduction commands, timings and scope.

Local Windows evidence for each Git version, normally and under -O: 53 passed, 8 existing platform skips out of 61 focused tests. All four new tests run without skips; their final source also passes on WSL/Linux in both modes. Disabling fallback or literal-name resolution in separate temporary source copies makes the new exact-object test fail. Native direct-child stall controls pass with the composed #4686/#4738/#5185 reader. The note distinguishes the full older-Git run from the final test-only teardown cleanup and rerun.

Limits: the broader suite explicitly used PYTHONUTF8=1; an initial default-encoding diagnostic assertion failed and remains documented. Windows descendant-held pipes, escaped processes, selected cold replay and mathematical/proof-kernel acceptance are not established. Hosted CI was in progress when this update was posted. Production code, frozen sources and selected frontier are unchanged. DONE reports author-checked tooling evidence, not maintainer acceptance.

Unclaimed next question: native Windows descendant-held-pipe ownership, beginning with readiness and proof that the pipe survives the leader. The existing diagnostic-encoding assertion is a separate small portability question. Refreshed overlap covered 305 open PR bodies and 4,637 Discussion records with complete remote pagination; local sparse-note and historical-diff coverage remain incomplete.

Credit Nick Petalas, Mihir Srivastava, Brian / JoyboyBrian, Woklerx, Arham Tahir, Douglas Colkitt and Tekalign's object-tree fixture lineage. Prepared for Mihir Srivastava with substantial OpenAI Codex assistance, one agent. Apache-2.0.

Original claim and scope correction, retained for provenance

CLAIM — actual older Git for Windows compatibility and composed reader deadlines

I am taking the open native older-Git check in this thread and the limited Windows composition request in the PR4738 reply. Base: canonical main 7c47256, priority revision 7. The selected source-bound orientation passes; no mathematical replay is implied.

Scope: use the official portable MinGit 2.35.3 x64 release in an isolated local fixture, verify its published SHA256, confirm that the actual executable rejects cat-file --batch -z, then run current reader compatibility, closed-pipe and direct-child deadline controls normally and under optimized Python on native Windows. Add a real-object witness for exact binary/empty/Unicode/CR-LF-name transport and observe the actual first-request/empty-response/exit-129 fallback. Current Git is the positive control. No system Git replacement, network operation by the old binary, production source change, Windows descendant-ownership claim or new kappa is planned.

Success/stop: exact reads and bounded direct-child failures retain existing rejection semantics; stop with a concrete failure or reproducible evidence report. Expected duration under 40 minutes; expiry 2026-10-23T21:05:00Z. Evidence is LOCAL-ONLY at claim time. Live #4681, 304 open PR bodies and 4,625 Discussion records were checked; the sparse local-note coverage is incomplete, and no novelty claim follows from search.

Credit Nick Petalas (#2752), Mihir Srivastava (#4686), Brian / JoyboyBrian (#4738/#5185), Woklerx, Arham Tahir and Douglas Colkitt for the existing reader, source contracts and controls. Prepared for Mihir Srivastava with substantial OpenAI Codex assistance; Apache-2.0.

Scope correction, 19:03 UTC: the initial MinGit 2.40.1 probe returned status 0 because that release already supports -z. It is retained only as a supported-protocol control; MinGit 2.35.3 is the actual unsupported-option target. No fallback result is attributed to 2.40.1.

{
  "schema": "integer-mult-coordination/v1",
  "kind": "DONE",
  "task": "portable-replay",
  "scope": "Actual older Git for Windows fallback, exact Git-only path transport, and composed direct-child reader controls",
  "base": "7c472563474e60185e6ca101a17622133d047e56",
  "expiry": null,
  "artifact": "https://github.com/CrocSwap/integer-mult-bounds/pull/5358"
}