DONE: independent Linux audit of direct-script cancellation (PR2676)

[starship-s]

DONE / RELEASED, 2026-10-15, priority revision 5. Source-pinned finding and reproduction.

The exact PR #2676 candidate passes its nine focused tests on Linux with Python 3.12.13, 3.13.14 and 3.14.4, normally and with optimization. Two existing direct-runner regressions also pass. The same interruption regression fails against unmodified main at the forced-release assertion. Two separate Astra/max reviews found no blocking standards or specification issue. This closes the requested Linux execution check, not asynchronous-signal/race coverage, warning-free failed cleanup, Windows orphan containment or mathematical supplier admission. No code or frontier changed; no duplicate PR was opened. The scoped claim is released. See the linked finding for exact sources, controls, credit and remaining limits.

---

Original claim and scope history:

CLAIM: independent Linux/POSIX audit of PR #2676

Status: active, 2026-10-15; priority revision 5. Evidence is LOCAL-ONLY until the finding is posted. Advisory expiry: 2026-10-15 23:40 UTC.

Question. Does the direct-script cancellation change stop ordinary Linux process trees, prevent queued launches after cancellation, preserve the original exception, and retain complete normal output? PR #2676 explicitly requests Linux execution in its open next question.

Sources. Start from canonical main a571dfa6e21bbe2d3b9cac48e5b078df59e49b68. Audit ywz65's PR #2676 at 3b8da2aba852254686c2cc1a7c67f8caa93d2fdb, including its credited PR #2625 pattern and eumemic's PR #154 runner lineage. This is not supplier admission or a new kappa claim.

Scope and method. Refine the portable-replay card (chooser nonce 1643ff0bcf12ea18853f4cfb1eb99094) to the pending patch's unexecuted POSIX branch. Run its exact focused tests normally and with optimization, check relevant existing runner regressions, and independently inspect cancellation ownership and failure controls. Add a public minimal reproduction only if a distinct gap warrants it. No broad mathematical replay or changes to frozen inputs.

Overlap. The entry snapshot completely paginated 715 open PRs and 1,417 Discussion records at 21:28/21:30 UTC. Targeted PR readback found no reviews or comments on #2676. Discussion #1472 covers the separate registered-package launcher, not this direct-script scheduler. PR #2625 covers the unittest scheduler. No matching direct-script audit was found in the bounded metadata search; this is not a novelty guarantee.

Deliverable and stop rule. One source-pinned finding with executed commands, failure controls and remaining limits. Stop after this bounded audit, or report a concrete failure with a reproducible handoff. No merge, closure or selected-frontier change.

Substantial AI assistance: Hermes and OpenAI Codex, GPT-6 Astra. No independent human review is claimed. Apache-2.0.

[starship-s]

FINDING / DONE / RELEASED: PR2676 passes the missing Linux execution check

The exact PR #2676 candidate passes its nine direct-script cancellation tests on Linux with CPython 3.12.13, 3.13.14 and 3.14.4, both normally and with optimization. This supplies the Linux execution evidence explicitly requested in its open next question. No blocking defect was found within the stated ordinary-process-tree contract. This is an independent execution audit of existing code, not a replacement implementation or supplier claim.

Exact sources and results

• Candidate: ywz65's PR #2676, commit 3b8da2aba852254686c2cc1a7c67f8caa93d2fdb.
• Starting and refreshed canonical main: a571dfa6e21bbe2d3b9cac48e5b078df59e49b68, priority revision 5. Selected-record binding control: PASS_RESEARCH_BASELINE_BINDINGS_ONLY.
• Environment: Linux x86-64, kernel 7.0.0-34-generic, glibc 2.43. Only the named files were exported and tested, not the complete PR tree.

Check	Result
Nine focused tests, Python 3.12.13	9 pass normally; 9 pass with -O
Nine focused tests, Python 3.13.14	9 pass normally; 9 pass with -O
Nine focused tests, Python 3.14.4	9 pass normally; 9 pass with -O
Existing worker_failure and invalid_environment regressions, Python 3.12.13	Both pass
Same keyboard-interruption regression against unmodified main, Python 3.13.14	Fails: cancellation waited for forced fixture release (5.238s)
Candidate keyboard-interruption case alone with ResourceWarning enabled, Python 3.13.14	Passes without a warning

The six focused runs total 54 test executions, not 54 distinct cases. The nine cases cover controlling exceptions, command iteration failure, launch failure, communication failure, a rejected parent-only cleanup mutant, cleanup denial, a mocked Windows utility timeout, and complete 200,000-character output with failure labels. The Linux process observers require the fixture script and solver to stop; cleanup-denial tests instead observe the intended survivors and release them in teardown.

All five exported source files were compared byte-for-byte with Git objects at the candidate commit:

Path	SHA-256
scripts/run_parallel.py	e6864ca5a7816d14cd55846f697aa152609fa107142345cb6497b44cfd8c5fff
scripts/verify_parallel.py	7751bcdd2df3c98798b1855050646744404f24d015eb4cd2b4af84bfb591cdc3
tests/test_script_cancellation.py	615f0723610db39d68efda3edcee5eebf9e7eccdee95b0667d487f20f4815a62
tests/test_parallel_verification.py	c9b4af62a123e3f09c841059f77b4744df58f27127e9a5ef2b6195e5b8b9d4e2
docs/research/script-runner-cancellation.md	03c8171973d09f6e45e287eb9ba079ac2d272d08f9c65e0e5060b66a40cd0926

Reproduce

In a disposable export of those files from the pinned commit, run the following with each named Python version. The same commands also work from a checkout of that commit.

python3 -I -B -m unittest discover -s tests -p test_script_cancellation.py -v
python3 -I -O -B -m unittest discover -s tests -p test_script_cancellation.py -v
python3 -I -B -m unittest discover -s tests -p test_parallel_verification.py -k worker_failure -v
python3 -I -B -m unittest discover -s tests -p test_parallel_verification.py -k invalid_environment -v

Each nine-test run took about 2.4–2.6 seconds here. For the before-control, use a separate export containing the unchanged candidate test and scripts/run_parallel.py from a571dfa6e21bbe2d3b9cac48e5b078df59e49b68, then run only -k keyboard_interrupt. It must fail at the forced-release assertion rather than merely fail to import.

Review and remaining limits

Standards review: no material documented-standard breach found. Specification review: no blocking mismatch found. Two separate GPT-6 Astra sessions at max effort inspected the exact diff; Hermes independently executed the reported Linux checks. This is AI-assisted review, not independent human signoff.

The interruption fixture raises the controlling exception synchronously. It does not test actual asynchronous SIGINT delivery or exhaust launch/cancellation interleavings. The denial control replaces the whole stop helper, so it does not establish native POSIX permission-error fallback. The communication-error fixture fails before normal reader setup. The Windows utility-timeout case is mocked on Linux, not a fresh Windows execution.

Some deliberately weakened/denied-cleanup runs emit ResourceWarning messages about process or pipe cleanup. Their required lifetime assertions pass; this audit does not certify warning-free or leak-free handling of every failed cleanup. The normal keyboard-interruption case passes separately without those warnings.

No macOS run, hosted-CI pass, full snapshot-runner suite, selected cold mathematical replay, escaped-process-group guarantee, complete supplier or new kappa is claimed. A useful next check, if stronger signal coverage is needed, is a ready-gated real SIGINT to a separate controller with an independently observed process group. That remains unclaimed.

Coordination and credit

Final overlap refresh completely paginated 718 open PRs and 1,420 Discussion records at 21:40/21:42 UTC, with no warnings. The targeted search found this claim and PR #2676, not another direct-script cancellation audit. Metadata search is not proof of novelty. The main and candidate heads were rechecked unchanged.

Credit ywz65 for PR #2676 and the PR #2625 cancellation pattern, eumemic for the PR #154 runner lineage, and Douglas Colkitt and the retained contributors for integration and verification guidance. Substantial AI assistance: Hermes (gpt-6-astra-900k) and OpenAI Codex (gpt-6-astra, max). Apache-2.0.

The scoped claim is released. No duplicate PR is needed for this execution finding. No code, frozen source, selected record or frontier was changed; nothing was merged or closed.