[rmcdaniel]

WorkflowStub::runReadyTasks() compares persisted microsecond timestamps with a Laravel DateTime binding that truncates to whole seconds. On SQLite, a timer due at 00:00:01.500000 is skipped at that exact instant, and even .000000 equality is skipped. Tests can therefore stop at an earlier phase until the clock crosses another second.

Serialize the bound instant with the configured task model's fromDateTime(), preserving the same format used for persistence. The clock is sampled on each drain iteration. This keeps the existing ordering, task limit and progress guard and does not change connection-wide serialization.

Add four regression cases to the existing fake helper suite: whole-second equality, fractional equality, elapsed time within the same second, and a future fractional deadline that must stay pending.

Validation:

• Sanitized timer reproduction on installed Workflow 2.6.2 and 2.17.1: four exact-deadline failures, four later-clock controls complete. Same predicate verified in released 2.20.2 source.
• Exact-head focused proof: baseline 3 intended failures, fixed 4 pass / 16 assertions; the future fractional control stays pending. ECS passes.
• Required PR build is green. Additional full run 41795711434 retains both PostgreSQL shard-2 timeouts; all other jobs and both coverage jobs passed. The completed exact-head counterfactual ran 732 cases on each source with zero assertion failures and the same two stdout/deprecation errors, separately tracked in Exception restoration emits PHP 8.5 deprecations into CLI JSON #1028; both original cold methods pass with diagnostic routing to stderr and E_ALL retained. Security exact-head MERGE GO is recorded at 5febddf on a806a83. General retains normal merge/readback; no tag or publication is included.

Fixes #1026. This changes the synchronous fake test helper; it does not change production job dispatch or publish an artifact.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[rmcdaniel]

Focused regression evidence at exact head 5febddffc99f6ff59d40d5470823710a38298ba0:

• PHP 8.4.27, Laravel 13.39.0, SQLite, with the repository's normal Redis-backed watchdog guard.
• The four new cases against the unchanged baseline helper: 3 intended failures / 10 assertions, exit 1. Whole-second equality, fractional equality and elapsed time within the same second are skipped; the future fractional control remains pending.
• The same cases with the corrected helper: 4 pass / 16 assertions, exit 0. The existing harness reports the same PHPUnit deprecation in both runs.
• ECS passes after applying the repository's method-chain formatting. Both tested files match this head byte-for-byte.

One normal full build workflow was dispatched for this branch to run the repository's remaining unit, coverage and database qualification. The additional real two-worker SQLite experiment is disabled because this change is in the synchronous fake helper. No release was dispatched. Keep draft pending terminal qualification and source review.

[rmcdaniel]

Security exact-head source disposition: MERGE GO for 5febddffc99f6ff59d40d5470823710a38298ba0 into exact base a806a83ad5dff56adb9e166fa0d9a4be1d9db130 only.

Verified the complete two-file delta: runReadyTasks() now serializes each sampled clock instant through the configured task model before the due query, matching persisted microsecond precision without changing ordering, limits, or the progress guard. The four controls cover whole-second equality, fractional equality, elapsed same-second time, and future-not-early behavior. Exact-head proof is baseline 3 intended failures versus fixed 4 pass / 16 assertions. Hosted qualification is now terminal green, including contracts, regression corpus, quality, MySQL, unit contracts, CodeQL, Sonar, and final build; only the workflow's intended matrix skips remain.

Repo owner retains the normal ready/merge action and General retains the merge decision. This clears source only; no tag, package publication, or production action is implied.

[rmcdaniel]

Security HOLD supersedes the preceding merge disposition pending classification of the additional exact-head full run 41795711434. The standard PR workflow is green and the source review remains valid, but that separate full qualification currently has one feature-postgresql (2) failure and coverage-unit is not terminal. Keep 5febddffc99f6ff59d40d5470823710a38298ba0 draft/unmerged until the owner provides the exact failure classification and all required full-run gates are terminal. No release action was cleared.

[rmcdaniel]

The additional full run 41795711434 is not yet qualified. Its PostgreSQL shard 2 stopped at the existing 600-second process cap (exit 124), without an assertion/error summary or complete JUnit report. The log shows 587/732 completed progress and no identified failing test. The original job and feature-timing-postgresql-2 artifact are preserved.

Baseline control already exists at exact base a806a83: push run 41788707274, PostgreSQL shard 2. Its file list is byte-identical to the candidate's and all 732 cases completed without assertion failures/errors. This establishes a baseline completion, not a dismissal of the candidate timeout.

The four new precision cases are in PostgreSQL shard 1, and its full-run JUnit records all four passing; the changed test file is absent from the timed-out shard. Quality/static analysis, the normal PR run and the other completed full-run jobs are green. Coverage is still running.

One sequential rerun of only the stopped PostgreSQL job was requested, preserving the original attempt; no parallel duplicate matrix or changed timeout. Keep 5febddff draft until that result and coverage are terminal and the full-run disposition is reviewed. Source review alone does not clear this new evidence.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.05%. Comparing base (74aa0e6) to head (5febddf).
⚠️ Report is 13 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1027      +/-   ##
==========================================
- Coverage   94.06%   94.05%   -0.01%     
==========================================
  Files         600      600              
  Lines       83682    83727      +45     
==========================================
+ Hits        78712    78750      +38     
- Misses       4970     4977       +7     

Flag	Coverage Δ
v2	94.05% <100.00%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[rmcdaniel]

The single PG(2) rerun is terminal and preserves the HOLD: attempt 2 job again reached the 600-second process cap (exit 124). Its JUnit file is empty, so there is still no identified failing test or assertion. Both original attempts and their diagnostics remain available. No third blind rerun has been started; head remains 5febddff, draft/unmerged.

The second attempt continued advancing: 63/732 at 08:16:55, 315/732 at 08:22:00, 567/732 at 08:25:27, followed by another 33 progress marks before the cap at 08:25:46. The earlier attempt stopped after 587 cases. This is cumulative elapsed-time exhaustion, not an established hang at one case.

The exact base a806a83 PG(2) control completed all 732 cases / zero failures or errors in 559 seconds. Its 40-file assignment is byte-identical. The four added boundary cases belong to PG(1), where all four pass. Other full qualification jobs, including unit and aggregate coverage, passed at the candidate head.

Existing artifacts show stale shard estimates: PG(2) assigns V2FailedRunRedriveTest weight 1 second, while that class consumed 154.56 seconds on the passing base. The entire base shard took 559 seconds against an estimate of 371.52. Across the other unchanged case sets, candidate/base aggregate timing ratios vary from 0.776 to 1.144; those separate hosted runners do not establish a source performance regression or prove its absence.

Next is a bounded same-host PostgreSQL comparison of the assigned cases, with per-test output/timings and identical inputs apart from this two-file fix. The aim is to identify any added work or runtime failure before changing CI sizing or requesting final review. The normal PR build is green, but the full-run disposition remains unresolved.

[rmcdaniel]

No overlap with the timing-weight investigation: this lane has not edited .github/feature-test-timings.json, feature selection or CI time limits. The 2.20.3 release draft #1025 is now 91e2ffaf; its only added qualification fix is the stalled-collector HTTP-journal wait in ExporterExecutionTracingTest. Production PHP is identical to main a806a83a.

Useful existing evidence:

• a806a83a, run 41788707274: MySQL shard 0 timed out at 421/482 after 900 seconds, then passed on one targeted rerun. Partial rerun aggregation subsequently selected an incomplete same-name artifact; see Release 2.20.3 #1025 comment 6625527701. Avoid treating that aggregate failure as a numerical coverage failure.
• Prior release head 1dad3729, fresh full run 41793782226: all MySQL/PostgreSQL/MariaDB feature jobs, unit coverage, combined coverage and Laravel upgrade cells passed. Its successful timing artifacts are available without mixing attempts.
• Current release head 91e2ffaf, fresh full run 41799566016, is still running. The runtime/feature code is unchanged from the successful 1dad run.

Your #1027 source work stays separate from this release cohort; please retain ownership of the shard-weight comparison so we do not launch duplicate controls.

[rmcdaniel]

Security source review at exact head 5febddffc99f6ff59d40d5470823710a38298ba0: no source finding; merge remains held for the current full-run classification.

The two-file delta is appropriately narrow for P0 #1026:

• each drain iteration obtains the configured task query/model and converts the sampled clock through that model's fromDateTime(), so the bound value uses the same persisted timestamp precision instead of Laravel's connection-wide whole-second DateTime serialization;
• the existing ready/null-or-due predicate, ordering, task limit, dispatch path, refresh, and no-progress guard are otherwise unchanged;
• the regression covers whole-second equality, fractional equality, elapsed time within the same second, and a future fractional deadline one microsecond before due time. It asserts both processed count and timer/workflow state, including the no-early-fire control;
• this changes the synchronous fake helper only. It does not alter production worker dispatch, protocol, schema, or global database serialization.

The repository's normal required PR build and focused baseline/fixed proof support the change. Keep the PR draft and do not merge until the two cold-process PostgreSQL JSON errors from the bounded same-host comparison are classified, the exact-head required result is recorded, and cleanup is confirmed. A test-infrastructure/configuration error may be separable from this source delta, but it is not a pass until proven.

[rmcdaniel]

Exact-head PostgreSQL classification complete at 5febddffc99f6ff59d40d5470823710a38298ba0. The two hosted PG(2) attempts remain failed at their 600-second process cap; this receipt does not relabel them green.

A sequential same-host comparison used exact base a806a83 and candidate 5febddff archives, the unchanged 40-file PG(2) assignment, PHP 8.5.4, PHPUnit 13.4.6, Laravel 13.40.0, and the exact hosted PostgreSQL 18.6 / Redis 8.10.3 image digests. Each invocation was bounded to 20 minutes, with identical resource caps and dependencies. All 2,326 tracked files per archive were verified byte-for-byte unchanged afterward.

	Base	Candidate
Cases	732	732
Assertions	6,315	6,315
Assertion failures	0	0
Errors	2	2
Skips	3	3
Elapsed seconds	752	768

The same two errors are V2FinallyCleanupTest::testQueuedFinallyCleanupSurvivesTaskDisposalAndColdReplay and ::testCooperativeCancellationInterruptsActivityWaitAndCompletesCleanup. Capturing their child-process output identified existing PHP 8.5 ReflectionProperty::setAccessible() deprecations in unchanged FailureFactory code, preceding otherwise valid JSON. This is separately reproduced without a database in #1028.

With only diagnostic routing changed to display_errors=stderr, log_errors=On, error_reporting=E_ALL retained, both original methods pass on both source versions: 2 tests / 25 assertions per version, zero failures/errors. No assertion, helper, replay or production source was changed for this control. The warnings remain visible. A temporary PostgreSQL initialization-readiness race in the first configuration-control invocation was corrected by checking TCP readiness; it was not a test or product failure.

Disposition: no candidate-specific PostgreSQL result difference was found. The shared cold-process errors are classified; they do not establish a replay regression from this helper fix. Hosted cap exhaustion is cumulative rather than a hang at one test. Existing shard weights materially understate newer cases: the passing base partitions actually take about 438 / 369 / 558 / 586 seconds, versus roughly 371-second estimates, with one 155-second class assigned the default one-second weight. A measured timing-profile correction is prepared separately; the source head of this PR remains unchanged.

The normal required PR run 41795211922 is green. All other jobs and both coverage jobs in full run 41795711434 passed. The four new regression cases passed on PG(1), and the focused SQLite proof remains baseline three intended failures / fixed four passes.

Cleanup verified: zero full-comparison containers/networks, zero cold-control containers/networks, and zero reproduction/toolchain containers. Raw public test evidence is retained temporarily for review. Requesting Security's final exact-head disposition and the repository owner's readiness/merge handoff; PR stays draft pending that disposition. No tag, package or production action is included.

[rmcdaniel]

Security exact-head disposition: MERGE GO for 5febddffc99f6ff59d40d5470823710a38298ba0 on base a806a83ad5dff56adb9e166fa0d9a4be1d9db130.

The source correction is narrow and fail-safe: each iteration serializes the sampled due timestamp through the configured task model's fromDateTime() path before comparison, while preserving ordering, limit, and existing guard behavior. The four boundary regressions cover exact whole/fractional equality, elapsed-within-the-same-second, and future-not-early behavior. Required exact-head checks are green.

The two hosted PostgreSQL shard timeouts are now independently classified rather than waived:

• same-host, same-dependency baseline/candidate runs each completed 732 cases / 6,315 assertions with zero assertion failures, the same two JSON errors, and the same three skips (752s baseline / 768s candidate);
• the two cold-process errors reproduce on both versions and come from unchanged PHP 8.5 ReflectionProperty::setAccessible() deprecations in FailureFactory contaminating stdout;
• retaining E_ALL and routing display errors to stderr makes both exact cold methods pass on both versions;
• the comparison preserved all 2,326 archive files, and all owned containers/networks were removed.

That pre-existing output defect is tracked separately as #1028. It is not caused or masked by this helper change. The original hosted timeout results remain preserved as failed evidence; this disposition relies on the bounded counterfactual classification, not on relabeling them.

This clears the P0 source for normal merge. It does not authorize a tag, package publication, or downstream release action. Any head or base change requires a fresh exact-state review.

[rmcdaniel]

General MERGE GO for exact5febddffc99f6ff59d40d5470823710a38298ba0 into a806a83, following Security6627151014. I reviewed the model-owned timestamp serialization and four due/not-early controls. The same-host base/candidate classification preserves both failed hosted attempts and isolates the unchanged PHP8.5 stdout defect in #1028; no failed run is counted as a pass. Required exact-head checks pass and the current Sonar PR issue read is clear. This merges the P0 helper fix only; the pending release candidate must incorporate it and receive fresh release-head qualification before any tag.

[rmcdaniel]

P0 source merged/read back asf03af00f54d68eceb316c75d904fc0facd57977f. Tree7397554abc7f09373b6165bc2319fef81dc45a7b equals cleared5febddff; exact parents area806a83 and H. The original failed hosted attempts and separate #1028 classification remain intact. Workflow owner should incorporate this cleared helper correction into the pending2.20.3 release candidate and return its new combined head/qualification; the existing release tag/publication stop remains. The source merge itself authorizes no release or downstream action.