ci: balance test shards and stabilize CI checks

[kmbroai]

Summary

Make CI faster and more reliable while retaining supported runtime coverage. Acceptance requires ten full passing PR attempts with p90 ≤214 seconds, versus the frozen 428-second original baseline, followed by verification of the exact post-merge main run.

Changes

• Replace separate CI runners with one small file-discovery runner and a timing table. Run four independent Bun processes on Unix and up to two inside each existing Windows shard; discover every eligible file and propagate failures. Add the large tracked-output fixture to the table at 11 seconds on Windows (median of five observed runs), while retaining its two-second Unix estimate. This changes only Windows file assignment; its real 6,000-file regression fixture, assertions, and worker counts remain unchanged.
• Split credential-home and ACL tests from the remaining runtime tests. Preserve all supported OS/Node lanes and protected check names. Keep the machine-policy check separate and serial, and keep local pnpm test serial.
• Run six Unix package-verification jobs alongside the six Unix test jobs, matching every OS/Node combination. The required aggregates depend on both matrices. This adds six jobs (33 total) and repeats setup, but does not duplicate tests or remove package checks.
• Cache the exact installed pnpm tool on Windows in a dedicated runner-temp prefix, keyed by OS, architecture, and the full packageManager pin. An exact hit skips reinstallation; a miss or cache error performs the real install, whose failures still propagate. Keep the resolved pnpm store cache. In full CI, only package-verification jobs restore the npm download cache.
• Use GitHub's existing RUNNER_TEMP for the fresh Windows npm consumer and a separate private directory for credential state. Local invocations without RUNNER_TEMP retain the normal temp directory; Unix consumer selection is unchanged. Both owned directories are cleaned up on success or failure. Preserve optional binaries, consumer install arguments, and timeouts.
• Use direct installed-Python lookup in the three scan-history fixtures, avoiding repeated production interpreter probes unrelated to their SQL, stdin, and path assertions. Raise those three child deadlines from 10 to 30 seconds after repeated Windows deadline failures. Keep real Python execution, missing-interpreter failures, functional assertions, and child cleanup; production discovery and its deadlines are unchanged. This is a fixture setup and time-budget change, not a proven host/root-cause fix.
• Publish the connected-publication descendant PID file only after its contents are written, using a sibling temporary file and rename. The existing readiness barrier then cannot signal and kill the descendant while its PID file is still empty. Real processes, signal escalation, exit checks, handoff assertions, and deadlines remain unchanged.
• Synchronize fixtures around actual readiness: signal handlers, login cancellation, React effects, and shared startup-lock release. Await subprocess output in the affected fixtures instead of using the pinned runtime's problematic synchronous pipe path.
• Force two real ambient-credential imports to overlap, checking distinct temporary filenames, successful results, exact credentials, and cleanup. This replaces eight unsynchronized imports and removes that eight-way ACL-churn stress scenario; its production behavior is not claimed fixed. Existing ACL churn and fail-closed tests remain.
• Remove an invalid export chunk-count assertion while retaining the real 2 MiB export and its other assertions. Neighboring tests still require forwarding before the exporter completes.
• Keep optional per-worker JUnit/LCOV diagnostics and distinguish parameterized case names. Preserve main's file-based CI classifier: Markdown-only pull requests run formatting and retain truthful required checks; full test and package jobs wait for successful classification. The earlier title-overlap and metadata-only optimization is superseded by main ci: use formatter-only checks for Markdown changes #650.
• Integrate main's native credential locking from fix(runtime): use process-owned credential-home locks #613 unchanged. Remove this PR's superseded rename/retry workaround and six operation-specific tests. Move upstream's replacement tests intact; update only their Unix timing estimate from 11 to the measured 18 seconds.

There is no production-source change relative to current main.

Testing

• New 5f004f6: only the timing table changes. The existing report/shard suite passed seven tests / 48 assertions in 1.615 seconds; raw identities, outcomes, and assertions exactly match the prior hosted Ubuntu subset. Formatting and diff checks passed. Independent assignment checks found all 97 Windows files exactly once across 14 nonempty workers and all 98 Unix files exactly once across four workers, with Unix assignments identical to the prior head. The 6,000-file regression fixture and all other test source are unchanged. The existing 277-file archive was rehashed, not repacked; CI timing tables are excluded from the npm payload. No new full local suite or installed-consumer run was performed. Three fresh native reviews and a separate independent verifier cleared the exact a3b157c / 5f004f6 publication gate with no actionable findings. Optional native-review credential/API probes encountered filesystem-ownership and fixture failures and are not counted as passing tests; new-head hosted execution is unverified.

• Prior d201738: the one-line atomic PID fixture correction passed the 11-test file (326 assertions) and exact 367-case worker (358 passed, nine skipped / 2,346 assertions in 74.96 seconds). A disposable reproduction of the old empty-file race failed before the fix and passed all five assertions afterward. Types, formatting, actionlint, and diff checks passed. Three fresh native reviews and an independent verifier cleared the exact base/head; four optional credential probes blocked by sandbox ownership were not counted as passing tests. Seven complete hosted attempts all passed in 204, 202, 185, 221, 174, 196, and 237 seconds, but the declared cohort failed timing after two samples exceeded 214 seconds. Both slow samples remain included; seven-run p90 is 237 seconds. The collector did not request attempt 8. First and seventh reports preserve exact raw identities, outcomes, and assertion counts on Ubuntu and both Windows versions: 1,932 cases / 23,486 assertions on Ubuntu and 1,928 cases / 23,288 assertions per Windows version. The PID case also passed on the previously failing Ubuntu Node 24 lane and macOS. Independent Node/Bun setup delays and unrelated runtime spikes show host variance remains; this change does not claim to fix that cause.

• Prior d3e1514: the exact failed seven-file Python worker passed 89 cases / 581 assertions in 34.08 seconds; types, formatting, actionlint, and diff checks passed. Three fresh native reviews and independent verification found no actionable issues. Fifteen full hosted attempts produced 12 passes, three failures, combined p90 204 seconds; all three declared cohorts failed. First-run Ubuntu/Windows inventories exactly matched the prior passing reports. The modified Python fixtures passed on both Windows versions; the first run did not exercise their larger deadline. Failures were a Windows export exit, a macOS diff-preview timeout, and the Ubuntu Node 24 empty-descendant-PID race addressed by the current correction. The first two causes remain unproven. All reports from attempts 1, 8, 10, and 15 were retained; uploaded reports exclude the macOS and Ubuntu Node 24 failure lanes, whose logs were preserved separately.

• Prior 25ff9db integrates main a3b157c and retains the workflow split and pnpm tool cache. 302 focused checks passed in 28.07 seconds; types, formatting, actionlint, and diff checks passed. Raw comparison with the previous focused run found exactly nine superseded identities removed and eight upstream replacements added, including one renamed stable-context test. No other identities were omitted. All six pnpm-cache and independent package-verification cases remain; upstream file-based CI classification and Markdown formatting replace the superseded event-based checks. The workflow expands to 33 jobs, 20 test lanes, and eight package-verification lanes. Subsequent hosted runs verified exact Windows pnpm cache misses and hits and real Unix package/test overlap. Four complete attempts measured 228 seconds passed, 214 failed, 167 passed, 195 failed; both declared cohorts failed on Python history fixtures. The combined four-attempt p90 was 228 seconds. All report artifacts from attempts 1, 2, and 4 were saved and independently checked; no test identities disappeared. The only failure deltas were the two history cases in attempt 2 and the alias case in attempt 4.

• Reused the unchanged 2125c0a full local four-worker coverage run: 103.912 seconds; 1,927 cases, 1,899 passed, 28 skipped. The first hosted 25ff9db run subsequently verified Ubuntu 1,932 = 1,904 passed + 28 skipped and each Windows version 1,928 = 1,866 passed + 62 skipped, with only the declared CI-test replacements and additions. The earlier installed-consumer result and unchanged 277-file archive are reused explicitly; no repack or consumer rerun was performed for these excluded files.

• Prior 2125c0a changed only the excluded package-smoke harness and testing documentation. The actual installed-consumer check passed in 32.50 seconds, including the native lock fixture and nested worker. A separate injected child failure propagated and removed both owned temporary directories; the success path also removed both. All 22 focused checks, typecheck, formatting, actionlint, and diff checks passed. The first full hosted run also passed in 202 seconds. Windows npm installation took 20.6 and 25.6 seconds, down from 143.4 and 141.2 seconds on the previous revision; both native credential-lock fixtures and installed consumers passed.

• Prior ef4aa31 correction changes only the Windows verification workflow, its documentation, and the diff-preview Git fixture. 299 focused tests passed; the exact failed 341-case worker passed in 74.57 seconds (337 passed, four skipped), with unchanged raw identities, outcomes, and assertions. Types, formatting, actionlint, and diff checks passed. That correction subsequently passed the full hosted matrix, including both Windows package checks, in 282 seconds.

• Prior 7952cf6 local four-worker coverage run on four CPUs: 104.19 seconds; 1,927 cases, 1,899 passed, 28 expected skips. The preceding run with the old timing estimate passed in 110.14 seconds; every raw identity/outcome is identical between those runs.

• For 7952cf6, compared with the previous 1,931-case revision, the only inventory changes were six superseded PR cases removed, four upstream cases replaced, and six upstream cases added. No identity normalization or unexpected omission. All 136 upstream runtime test declarations remain across the split files; only the previously reviewed ambient-import fixture differs. Three upstream cases retain their names with updated upstream assertions.

• Prior 7952cf6 fresh package build and actual installed-consumer checks passed, including public imports, NodeNext types, CLI, the new Node SQLite credential-lock fixture, bundled executable, and nested worker without a global install. The archive contains 277 files, including 113 bundled plugin files. Tests, CI runners, timing table, and the checkout-only native fixture are excluded. The archive is unchanged and was not repacked for the excluded harness/documentation change; the new consumer-script execution is reported above.

• Runtime/API source, README, and native fixture are byte-identical to main a3b157c. Bun remains 1.3.14; the attempted 1.4.0 upgrade failed unrelated compatibility checks and was not adopted.

• The exact a3b157c / 25ff9db publication gate passed: three fresh native reviews and a separate independent verifier found no actionable issues. Earlier review receipts were not reused.

• All 15 first-run report artifacts from 2125c0a were captured before retry. Raw identities and outcomes exactly match ef4aa31: Ubuntu 1,927 = 1,899 passed + 28 skipped; each Windows version 1,923 = 1,862 passed + 61 skipped. The tested merge has parents 84d0717 and 2125c0a, and its tree equals that revision. Prior reports and trials remain evidence for their respective revisions.

Hosted acceptance protocol

Measure each complete workflow attempt from run_started_at to its last completed job, including queues, setup, dependencies, tests, packaging, and uploads. Nearest-rank p90 is the ninth of ten sorted durations. The frozen original baseline is the latest ten qualifying full PR attempts before implementation; it includes five successes and five failures.

Declare each non-overlapping ten-run cohort before it starts, on one unchanged head. All ten must pass. A code revision starts a new declared cohort; the prior incomplete cohort remains recorded. Stop and diagnose a failure; a cohort may stop for timing futility after two samples exceed 214 seconds. Retain every failed cohort and earlier trial. The retry-cohort rule was amended after the first 6fa83bc timing failure; no slow sample is dropped within a cohort. Require both the fixed-cohort p90 and the combined p90 of every complete current-head attempt to be ≤214 seconds. Report both; ten runs do not guarantee long-run p90.

Validation checkpoint before publication: 5f004f6 follows d201738 on base a3b157c and changes only the large tracked-output fixture's Windows timing estimate. New-head hosted behavior and ten-run acceptance remain unverified. All 70 completed earlier trials remain below. Current hosted results are maintained in the benchmark comment.

Frozen original runs

Run / attempt	Outcome	Seconds
32906238652/1	success	323
32906206529/1	failure	337
32905921906/1	success	361
32886246469/1	success	338
32882775986/1	failure	428
32882083859/1	success	340
32881019494/1	success	314
32880328045/1	failure	318
32879636235/1	failure	341
32878596186/1	failure	431

All recorded candidate attempts

Revision	Run / attempt	Outcome	Seconds
410ab2b	32928314748/1	failure	304
410ab2b	32928314748/2	failure	229
e59b593	32929828556/1	success	258
e59b593	32929828556/2	success	265
cbc9733	32930853349/1	success	275
cbc9733	32930853349/2	success	203
5f29cb4	32931630001/1	success	209
5f29cb4	32931630001/2	success	226
4c32cba	32932492767/1	failure	179
fa39fe7	32933787767/1	failure	310
8db6213	32935106865/1	success	178
8db6213	32935106865/2	success	182
8db6213	32935106865/3	failure	188
27c9128	32937189245/1	failure	219
27c9128	32937189245/2	failure	213
6fa83bc	32938698465/1	success	479
6fa83bc	32938698465/2	success	443
6fa83bc	32938698465/3	success	160
6fa83bc	32938698465/4	failure	209
6fa83bc	32938698465/5	success	187
6fa83bc	32938698465/6	success	183
6fa83bc	32938698465/7	failure	184
7be8379	32943455320/1	success	169
7be8379	32943455320/2	failure	235
617db52	32945287144/1	success	250
617db52	32945287144/2	failure	173
384d242	32947470193/1	success	203
384d242	32947470193/2	success	204
384d242	32947470193/3	success	188
384d242	32947470193/4	success	217
384d242	32947470193/5	failure	192
91625fe	32950104898/1	success	206
91625fe	32950104898/2	success	186
91625fe	32950104898/3	success	208
91625fe	32950104898/4	success	404
91625fe	32950104898/5	success	199
91625fe	32950104898/6	success	226
7952cf6	32953762811/1	failure	195
ef4aa31	32954968767/1	success	282
2125c0a	32956533234/1	success	202
2125c0a	32956533234/2	success	212
2125c0a	32956533234/3	failure	231
2125c0a	32956533234/4	success	237
2125c0a	32956533234/5	success	228
25ff9db	32961380321/1	success	228
25ff9db	32961380321/2	failure	214
25ff9db	32961380321/3	success	167
25ff9db	32961380321/4	failure	195
d3e1514	32963948331/1	success	187
d3e1514	32963948331/2	success	199
d3e1514	32963948331/3	success	185
d3e1514	32963948331/4	success	180
d3e1514	32963948331/5	success	168
d3e1514	32963948331/6	success	172
d3e1514	32963948331/7	success	195
d3e1514	32963948331/8	failure	229
d3e1514	32963948331/9	success	179
d3e1514	32963948331/10	failure	190
d3e1514	32963948331/11	success	165
d3e1514	32963948331/12	success	175
d3e1514	32963948331/13	success	196
d3e1514	32963948331/14	success	198
d3e1514	32963948331/15	failure	204
d201738	32970204874/1	success	204
d201738	32970204874/2	success	202
d201738	32970204874/3	success	185
d201738	32970204874/4	success	221
d201738	32970204874/5	success	174
d201738	32970204874/6	success	196
d201738	32970204874/7	success	237

Risk and rollout

No public CLI surface, runtime dependency, supported runtime lane, or protected check name changes. The three scan-history child deadlines increase from 10 to 30 seconds; the existing whole-test and workflow limits remain unchanged. This permits slower functional-fixture completion and may delay failure of a stuck child by up to 20 seconds. Whole-workflow measurements still include that time. The six additional Unix verification jobs repeat setup and use more runner resources; any queue or setup cost remains part of the whole-workflow benchmark. Timing estimates affect scheduling only; per-worker LCOV is diagnostic, not a merged line-coverage guarantee. The upstream classifier precedes heavy jobs; invalid titles and formatting failures still fail required checks.

Upstream's SQLite transaction now serializes owner inspection and cleanup, superseding this PR's earlier same-version lock workaround. Upstream's compatibility limitations remain unchanged. The earlier isolated macOS CLI-output timeout remains unexplained and unchanged. Recurring Windows history Python startup/child failures motivated the disclosed fixture setup and deadline change; their underlying host cause remains unproven. The prior 2125c0a Windows login-readiness failure also remains unproven and retained in the benchmark. The removed eight-way ambient-import stress scenario is not claimed fixed in production.

Hosted runner queues and setup outliers remain included in the measurements. In the earlier 91625fe cohort, one job waited 305 seconds for a runner; another cached-Node worker spent about 24 seconds collecting tool versions and 21 seconds installing pnpm. Cache hits do not eliminate this observed variance, and the logs do not establish its underlying cause.

Do not merge until required checks, an approving review, and the ten-run timing comparison pass. Verify the full main-branch workflow for the exact merge commit afterward; the earlier main runs are not that verification.

Public disclosure review

• No customer, partner, prospect, or user identities, data, or identifying details are included.
• No credentials, personal data, private source, scan findings, or nonpublic links or tickets are included.
• I reviewed the branch name, title, description, commits, changes, comments, logs, screenshots, attachments, and links for public disclosure.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-26T13:49:34.254401Z	ec9c3e3	Manual request
🔒 Security Review	✅ Completed 2026-08-26T03:58:01.465773Z	410ab2b	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[kmbroai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep it up!

Reviewed commit: 27c9128ef7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: 6fa83bcd6c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. More of your lovely PRs please.

Reviewed commit: 7be8379372

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. 🎉

Reviewed commit: 617db52c8f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep them coming!

Reviewed commit: 384d242907

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Swish!

Reviewed commit: 91625fe19f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review the updated head 7952cf6.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. What shall we delve into next?

Reviewed commit: 7952cf67fb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review the updated head ef4aa31.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Chef's kiss.

Reviewed commit: ef4aa3103d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review the updated head 2125c0a.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Can't wait for the next one!

Reviewed commit: 2125c0adef

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

CI benchmark

Candidate: ec9c3e3a1d4f1bdeaf70617481f69d86607807ec, based on a3b157ce1014594ec3f0bc43894a818f2435526a. This comment is the current checkpoint; the PR description retains the earlier 5f004f6 checkpoint and its preceding trial table. The description was not edited before this push, to avoid starting another superseded full workflow.

Prospective declaration: deep-tool-a covers the first ten full PR workflow attempts on this unchanged candidate. This declaration was published before the candidate push. All ten must pass, with nearest-rank p90 ≤214 seconds against the frozen original 428-second p90. Separately, the combined p90 of every completed full attempt on this head, including failures and slower attempts, must also be ≤214 seconds. Stop and diagnose failures; stop for timing futility after two samples over target. No sliding window or dropped trial.

Verified: 10/10 full attempts passed, in 213, 205, 190, 192, 191, 220, 165, 199, 199, 213 seconds. All 33 jobs passed in each. Fixed-cohort and all-current-head nearest-rank p90 are both 213 seconds, versus 428 seconds originally: 2.009× faster (50.23% less elapsed time). The 220-second sample remains included. The collector stopped after attempt 10; no eleventh attempt was requested. A fresh workflow enumeration confirms these are all completed full attempts on this head.

Elapsed time is workflow start through the last job completion, including queueing, setup, dependencies, tests, packaging, report uploads, and aggregates. The frozen original ten runs had 5 passes / 5 failures; the final candidate had 10 passes / 0 failures. This is the observed ten-run result, not a guarantee of the long-run failure rate.

Current attempt	Outcome	Seconds
32976359721/1	success	213
32976359721/2	success	205
32976359721/3	success	190
32976359721/4	success	192
32976359721/5	success	191
32976359721/6	success	220
32976359721/7	success	165
32976359721/8	success	199
32976359721/9	success	199
32976359721/10	success	213

The 220-second attempt's slowest Windows job spent 143 seconds in tests versus 68 seconds in the first attempt. Both workers slowed across several unrelated files; caches hit and runtime/image versions were unchanged. No isolated source bottleneck or underlying host cause was established, and no code change was made. The full duration is retained.

All 15 artifacts and 32 XML reports from both the first and tenth attempts were saved and their ZIP digests and extracted bytes checked. Raw identities, outcomes, and assertions match the passing reference and each other exactly, with no normalization, missing cases, or duplicate identities: Ubuntu Node 22 1,932 cases / 23,486 assertions; each Windows version 1,928 cases / 23,288 assertions. These are the lanes with uploaded JUnit reports; other Unix lanes are verified through their actual test steps. The corrected tool-name case passed all 15 assertions on Ubuntu and both Windows versions in both checkpoints. Both prior Windows failure deltas are restored; nothing else changed. Every report is bound to the actual tested merge, with the stated base/head parents and matching tree. Both machine-policy checks actually passed in all ten attempts. The fresh GitHub review also found no major issues.

All 71 earlier completed trials remain recorded: 51 passed, 20 failed. Including the accepted candidate, the ledger has 81 completed trials: 61 passed, 20 failed. The first 70 remain in the PR description; the latest prior failed attempt is below.

Revision	Run / attempt	Outcome	Seconds
5f004f6	32974424066/1	failure	496

That attempt includes the full wait for the old-head description run to cancel. All 33 canceled-run job records are retained; it is excluded as canceled, not selected by timing. None of the resulting queue delay is subtracted from the 496-second sample. Both Windows versions failed the same advertised Deep-tool-name test: a timed synchronous subprocess returned null status in about nine milliseconds. The logs omit its error and signal. The pinned Bun Windows cached-timer mechanism fits the observed sequence, but remains an inference rather than a native reproduction. No second attempt was requested on that revision.

All 15 artifacts and 32 XML reports were preserved. Every test identity remains present. Only the named case changed from passed/15 assertions to failed/5 assertions on each Windows version; all other outcomes and assertions match the passing reference exactly. The report run/merge properties, parents, and source tree were independently checked.

The only new change replaces that fixture's synchronous subprocess call with the existing asynchronous process helper. It still executes the real selected Node and bundled server, sends the same JSON-RPC stdin, checks both worker and reducer layouts, parses the same output, and retains all 15 target assertions and cleanup. The child deadline remains 30 seconds. There is no production change, new helper, retry, test skip, or deadline increase.

Local validation passed: four focused tests / 37 assertions; the exact eight-file worker passed 32 tests / 1,360 assertions with both recorded seeds, in 4.539 and 4.586 seconds. Raw identities, outcomes, and assertions exactly match both the unchanged Linux replays and the prior passing hosted Ubuntu subset. Linux did not reproduce the Windows failure; these runs do not establish a Windows fix. Types, formatting, and diff checks passed. No new full local suite or installed-consumer run was performed. The unchanged 277-file archive was rehashed, not repacked; this test is excluded from the npm payload.

Three fresh native reviews and a separate independent verifier cleared the exact a3b157c / ec9c3e3 publication gate with no actionable findings. Additional credential probes in the reviews were blocked by the local temporary-directory ownership restriction and are not counted as passing tests. Those local reviews did not execute Windows tests.

The full workload remains 33 jobs, 20 test lanes, and eight package-verification lanes, including both Windows machine-policy checks. The first- and last-run artifact checkpoints are complete; the failed prior-head contrast is retained separately. As disclosed in the description, two controlled overlapping imports replace the eight-way unsynchronized import stress scenario; this removes that stress coverage and does not claim to fix a production race. Existing ACL churn and fail-closed checks remain. The redundant export chunk-count assertion was removed while preserving the real large-export and forwarding checks. Production source is unchanged. Progress updates edit this comment, not the PR description.

PR checks are green, the branch is mergeable, and there are no unresolved review threads. Exact post-merge main verification remains pending; the PR is still open. Merge authorization has not been granted.

[kmbroai]

@codex review the latest revision 25ff9db, including the parallel package-verification gates, Windows pnpm tool cache, and integration with the upstream CI classifier.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Another round soon, please!

Reviewed commit: 25ff9db877

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review the latest revision d3e1514, especially the direct Python lookup and explicitly changed 10-to-30-second fixture subprocess deadlines. Functional assertions, real subprocesses, cleanup, and production interpreter discovery should remain unchanged.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Hooray!

Reviewed commit: d3e1514b1a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review the latest revision d201738, especially publication-fixture descendant readiness: the PID must be fully written before its final filename becomes visible. Real signal escalation, descendant exit checks, handoff assertions, deadlines, and cleanup should remain unchanged.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep them coming!

Reviewed commit: d201738ae5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Swish!

Reviewed commit: 5f004f6433

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[kmbroai]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Hooray!

Reviewed commit: ec9c3e3a1d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".