You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The corresponding TSan mid-GC and both lifecycle gates pass. Those successes do not explain or waive this functional failure.
The microtask-churn profile accumulates callback-form Lock.asyncHold results through an ordinary shared global += after its hold has been released. #875 documents why that ownership boundary does not protect post-settlement reaction counters. The same profile also has ordinary asyncJoin/waitAsync totals. This is a proven unsynchronized-oracle seam, but the observed deficit still needs failure-only accounting to distinguish lost updates from missing or pending delivery. Do not infer a GC lifetime defect from the score alone.
Acceptance:
Retain the original hosted witness and prove callback/settlement ownership for each affected total.
Add failure-only exact delivery counts/unique-marker accounting and compare a clean parent control without changing the 900-program profile, seeds, GC requirements, run-loop or wait bounds.
Correct unsynchronized test totals with atomic integer sum/count pairs, preserving all bundle identity and rejection checks. Keep any genuinely lock-protected counters unchanged with an ownership explanation.
Original source/profile coverage, exact sums and completion counts remain mandatory. No sleeps, deadline changes, skipped assertions, serialized engine scheduling or retry-until-green.
Pass the canonical affected gates and post the evidence before closing the threading qualification frontier.
Related: #875, #1123, #1118. This is a separate mid-GC profile from the property-waitAsync oracle corrected in 1d5f058. The regex-cache OOM fix tracked by #1124 remains isolated; no causal edge to that source defect is asserted.
The clean production parent diagnostic confirms lost counter updates. Across 600 original microtask cases, all 40 ordinary-sum failures had exact atomic sums, exact counts, complete unique-marker masks and zero delivery errors. The ordinary total remained wrong on the later read. This rules out missing or duplicate delivery for those retained witnesses.
The canonical correction uses atomic sums, counts, unique-marker masks and persistent failure flags for post-settlement waitAsync, asyncJoin and callback-form asyncHold reactions. Existing bundle/root/seed checks, profile membership, allocation pressure, seeds and timing limits are unchanged. The no-function asyncHold total remains ordinary because it owns the hold until release().
Atomic correction patch SHA-256: bdc4d6248a1095d6eb5517fd8589204ceb7fbe4e92673755e2aac1fa6bcbb9c3. Production source is unchanged. Evidence and receipts are retained under Verification/settlers-retained-replay/native-growth-audit/managed-frames-integration/midgc-microtask-accounting/. This issue remains open pending final source integration and canonical qualification.
After integration onto d59e081, the unchanged canonical ordinary mid-GC gate passes 900/900 and the unchanged TSan gate passes 90/90. Runner and engine __tsan_init, __tsan_read8 and __tsan_write8 symbols are verified; TSAN_OPTIONS and the timeout override are unset. The final rebase adds only the separately delivered ecosystem workflow change f9cd1a1. Runtime and fuzzer bytes are identical to the qualified 68224dd2 snapshot; the correction's patch hash is still bdc4d6248a1095d6eb5517fd8589204ceb7fbe4e92673755e2aac1fa6bcbb9c3.
The earlier canonical seed-13 collection failure remains retained in #1126. Failure-only telemetry independently reproduces round-limit aborts in 2/100 script Worker and 1/100 module Worker cases, so these green accounting gates do not close the GC convergence frontier. Hosted final integration remains pending.
Receipts: midgc-microtask-accounting/integrated-canonical-midgc-20-receipt.json and integrated-canonical-tsan-midgc-2-receipt.json under the retained managed-frames-integration evidence root.
A separate unchanged profile family failed during the current final-source mid-GC qualification: seed1, runMidScriptAsyncHoldReleaseWaiterCleanupGc, reported midgc asyncHold release promises did not settle 2. The full 900-program run is being retained rather than retried for a green result.
This family is outside the two GC phase-boundary corrections being tested for #1126. Its ordinary captured joinCount++ and joinSum += v run in post-settlement reactions, after the asyncHold release. That resembles this issue's oracle ownership seam, but the new deficit does not yet distinguish a lost test update from a genuinely missing delivery. The existing 100,000-pump bound, one-millisecond property waits, original counts/sums, GC assertions and source workload remain unchanged.
Next is failure-only atomic delivery/unique-marker accounting on this exact family and a clean-parent control. No causal claim is made about the array-result rooting fix or the new collection predicate. Current source is production363e plus the separate34-line fuzzer-only phase-boundary candidate; the candidate does not change this helper's body. The new log is retained at native-growth-audit/managed-frames-integration/midgc-finalization-progress/completion-final-debug-midgc900.log with its source patch receipt.
Activity
chrisbbreuer commentedon Nov 10, 2026
The clean production parent diagnostic confirms lost counter updates. Across 600 original microtask cases, all 40 ordinary-sum failures had exact atomic sums, exact counts, complete unique-marker masks and zero delivery errors. The ordinary total remained wrong on the later read. This rules out missing or duplicate delivery for those retained witnesses.
The canonical correction uses atomic sums, counts, unique-marker masks and persistent failure flags for post-settlement waitAsync, asyncJoin and callback-form asyncHold reactions. Existing bundle/root/seed checks, profile membership, allocation pressure, seeds and timing limits are unchanged. The no-function asyncHold total remains ordinary because it owns the hold until release().
Qualification on production parent 1d5f058:
Atomic correction patch SHA-256: bdc4d6248a1095d6eb5517fd8589204ceb7fbe4e92673755e2aac1fa6bcbb9c3. Production source is unchanged. Evidence and receipts are retained under
Verification/settlers-retained-replay/native-growth-audit/managed-frames-integration/midgc-microtask-accounting/. This issue remains open pending final source integration and canonical qualification.test(threadfuzz): make mid-GC settlement accounting atomic
chrisbbreuer commentedon Nov 10, 2026
Delivered in 2e47eb06.
After integration onto d59e081, the unchanged canonical ordinary mid-GC gate passes 900/900 and the unchanged TSan gate passes 90/90. Runner and engine
__tsan_init,__tsan_read8and__tsan_write8symbols are verified; TSAN_OPTIONS and the timeout override are unset. The final rebase adds only the separately delivered ecosystem workflow change f9cd1a1. Runtime and fuzzer bytes are identical to the qualified 68224dd2 snapshot; the correction's patch hash is still bdc4d6248a1095d6eb5517fd8589204ceb7fbe4e92673755e2aac1fa6bcbb9c3.The earlier canonical seed-13 collection failure remains retained in #1126. Failure-only telemetry independently reproduces round-limit aborts in 2/100 script Worker and 1/100 module Worker cases, so these green accounting gates do not close the GC convergence frontier. Hosted final integration remains pending.
Receipts:
midgc-microtask-accounting/integrated-canonical-midgc-20-receipt.jsonandintegrated-canonical-tsan-midgc-2-receipt.jsonunder the retained managed-frames-integration evidence root.fix(regexp): allocate cached matcher headers on first use
chrisbbreuer commentedon Nov 12, 2026
A separate unchanged profile family failed during the current final-source mid-GC qualification: seed1,
runMidScriptAsyncHoldReleaseWaiterCleanupGc, reportedmidgc asyncHold release promises did not settle 2. The full 900-program run is being retained rather than retried for a green result.This family is outside the two GC phase-boundary corrections being tested for #1126. Its ordinary captured
joinCount++andjoinSum += vrun in post-settlement reactions, after the asyncHold release. That resembles this issue's oracle ownership seam, but the new deficit does not yet distinguish a lost test update from a genuinely missing delivery. The existing 100,000-pump bound, one-millisecond property waits, original counts/sums, GC assertions and source workload remain unchanged.Next is failure-only atomic delivery/unique-marker accounting on this exact family and a clean-parent control. No causal claim is made about the array-result rooting fix or the new collection predicate. Current source is production363e plus the separate34-line fuzzer-only phase-boundary candidate; the candidate does not change this helper's body. The new log is retained at
native-growth-audit/managed-frames-integration/midgc-finalization-progress/completion-final-debug-midgc900.logwith its source patch receipt.fix(threadfuzz): synchronize post-release join totals