test: validate real scheduler drain on accepted PDB evidence baseline - #170
Conversation
Record exact main d01bb92 as parent before opening the independent fixture PR. Its tree equals the scheduler parent used for development, so the four fixture files and resulting tree are unchanged. No history rewrite or unvalidated production opt-in.
First-source MSVC compilation failure retained; targeted ADL correctionSubmitted head Actual MSVC diagnostic at probe line453 is C2676, attempting to append Downloaded initial default build artifact 10056947054, 2,896bytes, verified SHA256 Targeted correction: parenthesize the JSON helper call A corrected source head will receive its own complete Windows/real-matrix/native/independent-ABBA checks. No cases will be removed, no same-head retry used to hunt for green, and first-head runs/artifacts remain separate provenance. No CLI or writer-lease authorization. |
Initial independent ABBA crossed the declared no-op gate — not cleared by the JSON fixRead-only audit of first-source The external timings-off no-op paired median is +1.6946ms/+13.7988%, MAD0.91615ms, maximum+2.8174ms,3/4pairs slower. This crosses the predeclared >1ms AND >10% gate. The workflow status is green because it collects evidence, not because the acceptance threshold is encoded in that job.
Investigation so far checked exact source scope (product source/headers and benchmark unchanged), expected external clock, pair/orientation identities, original per-pair arithmetic and all instrumented semantic counters. The regression appears in both orientations; it cannot simply be attributed to always running candidate second. The JSON contains no external timing breakdown or raw stdout/stderr/benchmark executable archive that would establish a root cause. No cause has been isolated; no samples are removed as presumed noise. Instrumented modules-cold also retains median+162.796ms/+9.32%, maximum+601.879ms. Favorable results in other scenarios are not used to offset this gate. The targeted source correction First-source Native Debug#707,documentation#162,Cross-Stage#187 passed; Native Release#601 was cancelled by the genuine source correction, not counted as full validation. Default/private failed in probe compilation with zero ownership observations. The first 150 synthetic contract outcomes were also independently checked:113accepted,37rejected, all expectations match; SHA256 matches exact CRLF collector source. The old head remains rejected, all its diagnostics and raw artifacts retained. No same-head benchmark rerun was requested. |
Corrected-head independent ABBA completed; hold automatic merge pending anomaly reviewBase All76pairs/19scenarios and AB/BA/AB/BA orientations were checked; all72instrumented pairs preserve complete counters/breakdowns/compile-link-archive hit-miss. Four external timings-off samples have unavailable counters, not zeros. Paired medians are computed per pair, not differences of marginal medians. Four-sample nearest-rank P95 is the maximum, not a population-tail estimate. Only the last row uses timings-off external elapsed; every other row is instrumented.
The corrected head's external paired-median gate is not crossed. However, the initial head's +1.6946ms/+13.7988% gate crossing remains unexplained. The one-call ADL fix is not a performance fix. The new +3.946-second scale-cold tail, +1.076-second single-TU tail and +84.69% scale-single-TU instrumented paired median are retained, not dismissed as noise. No favorable sample is used to cancel an adverse one; no speedup/zero-overhead claim, no same-head rerun, and historical cold-tail debt remains open. Current recommendation: do not automatically merge #170 just because the new no-op median is below the threshold. Finish the real-tool and native review, then keep an explicit hold unless the timing anomaly investigation supports acceptance. These data do not prove that test-only source changes caused product regression, but the missing attribution also does not prove they are harmless. The artifact is paired comparison JSON, not a raw process/OS trace or archived pair of benchmark executables. The corrected probe has now actually compiled successfully on the separate Windows build VM. The150synthetic contract expectations are independently rechecked from corrected-head artifact10057257509:113accepted/37rejected, allmatch, collector hash matches exactCRLFsource. Real default70/private42 are still separate gates. No CLI cancellation or writer-lease authority is granted. |
Iviesever
left a comment
Iviesever
left a comment
There was a problem hiding this comment.
Final iteration review — real scheduler fixture validated; merge held for unresolved timing evidence
Implementation/source review with independent read-only artifact/statistics recomputation, not an independent human approval.
Exact base d01bb920e030202c33639f59139491ac7f49117c; reviewed head 283b50733601926840fdcd0474d10a06a4f75c7d; tree e62083fae3e43550a97abcd9d0410da5d4748f2c. Native test-merge 3ebce37b047aed9e756f497f9bff7e546d70ec4f has the identical tree and exactly those base/head parents. Full final diff reviewed: four fixture/collector/contract/workflow files, 157 additions/15 deletions. No product source/header, scheduler implementation, manifest, native executable inventory, benchmark, CLI/freshness or VERSION change.
Completed corrected-head execution
All this head's first-attempt workflow executions completed successfully:
- Native C++ #708 /34229725695: actual Debug build, all paired subshards, ambient/include/runtime/parameter and freshness/side-output checks, final gate.
- Native Release #602 /34229725662: Stage0, all four Release shard jobs, full self-host and validation of the exact runtime-only package. publish-release skipped.
- Documentation #163 /34229725639, Cross-Stage #188 /34229725648.
- Default Endpoint #8 /34229725685: actual probe build on a different VM,150synthetic contract outcomes,7C++ helper checks,2refusal guards,real70-case matrix.
- Private Ownership #10 /34229725690: real42-case matrix.
- Independent Performance #550 /34229725755: actual exact base/head checkouts, both Release builds, instrumentation check,19x4ABBA and upload. This is measured evidence, not a skipped job.
New result: the actual scheduler API is exercised
14/14 scheduler-drain controls pass across seven profiles and both service origins. Read each actual A/scheduler.json, A/drain.json, original A/work0.result.json, parent observation and B's original results:
- Exact API
BoundedWorkScheduler::run_with_admission_stop, worker_count1/started_count1, stop_requested/admission_stop_observed/stopped_before_all_items alltrue. - Event actually observed and forwarded while the callback was active; bridge nativeerror0, no callback or scheduler exception.
- Original A compile succeeded; pending A TU was not dispatched. Retained A/B compiler overlap observed, retained A process signaled on return, original server survived, original B compile/link/run all succeeded.
- All14 prior direct-drain controls remain and pass. Independently compared 56 source files and70 argv records between the fourteen reference/API pairs: sources byte-identical; argv identical after normalizing only case-directory identity. This is input equivalence, not a timing comparison of policies.
The functional fixture gap from #168 is closed for this single-A-worker, one-active/one-pending scope. Neither compiler receives a terminating token. callback_active includes compile/capture/diagnostics and does not establish an in-flight RPC. This is not real multiworker A cancellation coverage, CLI Ctrl+C, orchestration-wide cancellation propagation, owner-crash recovery, external-writer enumeration, durable quiescence or a write-lease certificate. Both safety fields remain false.
Raw matrices, failures and provenance
Default 70/70, private 42/42, zero collection/control/infrastructure failures. Both modes'14unmanaged controls pass. All28A-started managed endings nevertheless still kill their original service. Default preserves 3 original B failures: /Zi Debug cancel(C1090), /ZI Debug cancel(C1001), Modules Release normal(C1090). Private preserves 1 original B failure, /Zi Release A-started normal(C1051/C2471). Failed B link/run were unattempted, not passed. All112later recovery compiles succeeded and did not replace these four original failures; favorable B successes despite service death remain too.
Actual default toolchain: MSVC14.51.36231 /Hostx64-x64,Windows10.0.26100,runner image20260824.214.3,PowerShell7.6.5,fourCPUs. All70default preflight/post-discovery lists empty and typed outer envelopes clean. Exact prebuilt executable members were independently hashed and matched to observer/source provenance; these are the observer inputs, not the performance workflow's binary pair.
Independent audit checked 717 default/415 private original tool results, 1434/830 stdout/stderr files. Default B owner identity appears only 9/70, private 5/42. In scheduler-drain, A-at-request identity is8/14 and B identity1/14; direct-drain has6/14 and2/14. A-after sets are empty in all28drain cases; this does NOT prove no remaining writers. Scheduler request-to-A-result observations span 4319.2973–10005.2188ms including capture/diagnostic overhead; the bridge10mswait is not a cancellation deadline. No latency guarantee or policy speedup is claimed.
Verified raw ZIP digests:
- default 10057916488,19802533bytes,4096entries:
5081cd6141a1792e7c9efeebb3241b21e562a4f69a185ffdc078ce73ef19d839 - private 10057682494,4918804bytes:
cfa647a2a9472a11477320e1a7fa2e76b39284cbe99668ea6444274342f79862 - contract 10057257509,1393219bytes,6051entries:
a4ad5fa4432f2baeeed8dc32cc057627154fb023b4fd748d3c15f19dde871829—150/150expectations,113accepted/37rejected; exactCRLFcollectorhashverified. - exact input 10057579709,1175508bytes:
f524b3fbc743d58fb77eb1e636f44f86d56c374f6c5cfc95f1d217e0d22732dc - probe build 10057580695,4587bytes:
d7fb5fe44b59739772f04d8d1a81d2d3efdab1f834cf2a3d6a8e35d83785eaaf - corrected ABBA 10057522616,35144bytes:
f2479c1e3d740d096b32f3380144c76bfb780297af118703644733cd5358dcd9
Diagnostic ZIPs omit generated binary bytes; their916default/536private hashes are post-case inventories, not cancellation-instant snapshots. The separate input ZIP contains its two exact executables. Actions retention30days; raw ZIPs and reproducible audit scripts also retained in this delivery.
Merge decision: HOLD, despite completed CI and a below-threshold new no-op median
The initial ABBA review actually crossed the predeclared >1ms AND >10% external paired-median gate, +1.6946ms/+13.7988%. The corrected head's full independent19-rowABBA has external median+0.14575ms/+1.6312% and therefore does not cross that gate. All72instrumented semantic identities match on both runs. Nevertheless, the source correction only fixes JSON-helper ADL, not a product runtime path. A later lower median is not evidence explaining the earlier gate crossing.
New unfavorable evidence remains: scale-cold maximum +3946.398ms, single-TU maximum +1076.118ms, scale-single-TU median +87.006ms/+84.69%, cold median +141.396ms/+11.01%. No source-causal regression is proven, but missing external-span/binary/OS attribution also does not establish harmlessness. Both ABBA runs and all favorable/adverse pairs remain, with no same-head rerun or noise-based deletion. I do not authorize merge this iteration. Keep the PR open as Draft; no automatic or protection-bypassing merge. The timing-investigation requirement and historical cold-tail debt stay open.
Before acceptance, the next focused iteration should improve attribution with a predeclared fixed sample budget, preserved exact binary/environment identity and external invocation records, including an A/A calibration alongside A/B rather than retrying until green. Historical missing traces cannot be invented retroactively; any eventual risk acceptance must explicitly preserve that limitation. Do not propagate production cancellation or transfer project leases on the strength of this fixture while that decision is pending.
First failure and local validation corrections retained
First head f52a8adc... failed actual probe compilation with C2676 (std::quoted ADL); both real matrices had zero observations. Raw default/private build ZIPs10056947054/10056939288 and initial contract/ABBA10056801357/10057161311 are preserved. Initial Debug/documentation/Cross-Stage passed; Release was cancelled by the genuine source correction, not counted as full validation. The one-call fix was reproduced with <iomanip> in a local strict GCC syntax regression; corrected JSON escaping also passed ASan/UBSan. These helper tests are not local MSVC execution.
Two read-only local audit-development/display mistakes (a missing wrapper parameter, then a nonexistent display key) were corrected and logged without changing source, raw artifacts or requesting a CI/benchmark rerun. The earlier initial-ABBA comment's MAD transcription should read 0.91645ms, not0.91615ms; the saved recomputation is correct and gate status unchanged. No failures are relabeled successful.
VERSION remains5.5.0. #168 is already independently merged; this review neither reverts it nor claims #170 is merged. M1b remains incomplete; final VERSION/release work is separate.
Iviesever
changed the title
Attribution iteration: fixed experiment declared before new measurementsReconfirmed main The new investigation will build A=exact main and B=exact held head using the existing MQB-native Release build scripts and one pinned seed. Preserve the actual two executables, hashes, source-input trees, build transcript, allowlisted environment (not credentials), runner/PowerShell/clock identity and unchanged fixture definitions. Fixed budget: eight AA pairs plus eight AB pairs = 32 complete fixture reports, each with 19 measured scenarios and two priming invocations (672 invocation records). AA uses the SAME A executable path and bytes in both roles; AB uses A and B. Pair orientation alternates AB/BA; contrast order alternates AA/AB versus AB/AA by block. Persist the entire plan before launching. No adaptive sample count, no same-slot retries, no outlier deletion and no subtraction of AA variation from AB effects. An explicitly opt-in observer keeps the PowerShell native invocation and original fixture inputs, but records monotonic before/after directory-switch/native-call-and-capture boundaries and the complete captured merged TEXT lines, exit code and failure BEFORE parsing timing JSON. It adds four clock reads and out-of-interval record writes, so these new observed samples are labeled separately from the historical external stopwatch. The unchanged/default benchmark path and independent 19x4 PR ABBA remain separate validation. The observer does NOT magically obtain native-child PID/CPU, per-child ETW or retroactive OS traces; those unavailable dimensions will remain explicit. External minus internal totals are residuals, not proven OS queue time. Record writes may perturb later samples; AA helps expose this but cannot certify no observer effect. Before collection, actual-helper contracts must pass for schedule balance, strict recorded clock/outcome checks, complete output including nonzero exits, cwd restoration, malformed timing retention, no overwrite and unavailable counters. Full relevant native/release and independent exact-base/head gates remain required for the instrumentation PR. Its external no-op median gate remains worse by BOTH1ms and10%; no thresholds are relaxed. For #170: analyze AA and AB separately, preserve ALL first-source/new-source results, and keep HOLD if evidence remains insufficient. A lower new median alone will not be called an explanation for the first +1.6946ms/+13.7988% crossing. No automatic merge, source-causality claim or CLI/lease authorization is encoded in the collector. Historical missing performance binaries/traces and cold-tail debt cannot be reconstructed by running again. |
…alibration Accept opt-in attribution tooling after exact-tree Native Debug #710, full Release/self-host/package #604, Documentation #165, 32 Windows observer contracts, independent observer-off ABBA #553 and fixed 32-report/672-invocation AA/AB study #2. Preserve original fixture and inactive-observer clock, actual binaries, raw merged diagnostics, exact identities and all favorable/adverse samples. Same-binary AA tails narrow new-run latency scopes but do not reconstruct #170's historical no-op gate crossing. #170 remains Draft/HOLD; no speedup, zero-overhead, OS-cause, CLI cancellation or writer-lease authorization. Keep first-source contract failure and corrected-source provenance. VERSION stays 5.5.0; no historical tag/asset or formal release change.
Attribution iteration delivered; this PR remains Draft/HOLDIndependent tooling PR #171 is now merged to main Complete fixed AA/AB study, raw artifact identity, all19 contrasts and precise scope localization The predeclared runtime pair A=d01/B=283b completed32reports/672invocations without repeated slots, deleted samples or AA subtraction. The two actual executables, build transcript and original merged output/timing records are now preserved. All288instrumented scenario-pairs have identical semantic counters, and16timings-off pairs keep counters unavailable. New observed no-op medians AA+0.1216ms/+1.0806%, AB+0.1088ms/+0.9631% are below the investigation threshold, but do not explain this PR's earlier +1.6946ms/+13.7988% crossing. Concrete NEW-run evidence: the same A binary/path in AA produced scale-cold external+1634.2199ms while internal total grew+1633.615ms and cwd only+0.0780ms. Another same-A single-TU pair grew+1012.9075ms, with link phase40.711→846.015ms and cache-save work3.107→213.241ms. The cache-save scope includes serialization plus directory/open/write/flush/close/replacement operations; these numbers do not isolate disk waits, AV activity or a specific syscall. Parallel/nested work is not summed as elapsed. Actual A/B whole images differ in19bytes (timestamp/decorated-name fields); .text is identical, but this is NOT whole-image equality or a runtime-cost proof. New host-boundary records are not historical ETW/per-child CPU traces. Observer overhead and AB path/metadata effects remain possible. Historical missing artifacts cannot be retroactively reconstructed. Decision remains HOLD, not a source-causal regression verdict. The next focused experiment should distinguish cache persistence subspans and link/process launch/wait/capture with credible per-tool/OS evidence, a fixed predeclared sample budget and retained raw output. Do not blindly repeat whole-suite medians, infer historical causes from new AA tails, or modify freshness/durable-write rules as a performance workaround. If this PR is later updated onto the new main, validate its new exact base/head/tree separately; the old pair is not new-head evidence. CLI cancellation, multiworker propagation, external-writer quiescence and lease transfer remain unauthorized. |
Accept controlled investigation tooling after exact-tree Native Debug #713, full Release/self-host/package #606, Documentation #167, Cross-Stage #190, ten synthetic and seven actual-file checks, eight fixed query-on/off cases, unchanged default56/private42 and independent ABBA #558. Preserve exact physical file identities, unavailable queries, original diagnostics and all unfavorable timing/managed-service results. Four query-on and four query-off original builds pass, but finite non-recurrence does not explain #172's historical C1041 or release #170's performance hold. Keep both old PRs held; no CLI cancellation, M1b completion, external writer quiescence or lease-transfer authorization. VERSION stays 5.5.0; no historical tag/asset or formal release change.
…ness Accept the event-backed readiness iteration after exact-tree Native Debug #718, full Release/self-host/package #610, Documentation #171, Cross-Stage #195, 36 synthetic contracts, actual withheld-provider refusal and four complete calibrated MSVC traces, unchanged default56/private42, and independent ABBA #566. Require actual same-session event delivery and single opening calibration before compiler admission; rematch witnesses against authoritative ETL records. Preserve original C2398/old incomplete traces, C1041 controls, managed-service failures and all adverse performance samples. No same-head retry or weakened gates. Keep #170/#172 held; no M1b completion, CLI cancellation, external writer quiescence or writer-lease authorization. VERSION remains5.5.0; no historical tag/asset or formal release change.
…ed PDB events Accept bounded investigation tooling after exact-tree Native Debug #721, full Release/self-host/package #612, Documentation #173, Cross-Stage #198, original PDB investigation, 40 invocation plus 36 trace contracts, eight fixed invocation traces and four reference traces with native refusal/calibration controls, original default56/private42 and independent ABBA #570. Preserve initial cross-volume correlation failure, ambiguous B child candidates, original outcomes and all adverse timing samples. Six sharing-violation events are coordinator-probe events, not proven compiler failures; historical C1041 and #170/#172 holds remain unresolved. No CLI cancellation, M1b completion, external writer quiescence or lease-transfer authority. VERSION5.5.0 and historical tags/assets/releases unchanged.
…daries Record same-QPC native API intervals, PID+creation and TID only in explicit invocation investigation. Preserve native returns before timestamp/serialization, RM cleanup and metadata-open outcomes. Add strict offline matching without treating temporal overlap or public API success as compiler/RPC causality. Keep fixed eight/four studies, original tracing/ownership/ABBA, product code and VERSION5.5.0 unchanged. Add 38 record contracts; preserve historical missing records and #170/#172 holds.
…vidence Accept test-only diagnostic capability after exact-tree Native Debug #727, full Release/self-host/package #617, all original default56/private42 positive controls and retained studies, 33 fault plus 154 prior record contracts, four fixed cold-B calibrations, and independent ABBA #583. Both injected C1041/exit2 cases retain native MSPDBSRV sharing failures and separate recovery; original controls and historical failures remain distinct. Explicit scoped risk acceptance revises only #176's PR-level hold, not historical C1041/root-cause or production reliability status. Preserve the failed first path-matching audit, original historical C1041, all adverse timing including +4.628-second tail, and #170/#172 holds. No product behavior, CLI cancellation, writer-lease, M1b completion or formal release authority. VERSION5.5.0 and historical tags/assets unchanged.
Resume after #176 — exact updated-base scheduler fixture integrationReconfirmed main Scope: preserve every accepted #171–#176 capability and restore #170's existing actual Predeclared gates for the NEW combined head: full Debug/all Release shards/self-host/exact package, documentation and relevant cross-stage;150/150existing scheduler observation-record expectations; all70default/42private cases with all required positive controls;14actual scheduler summaries, original A result and pending-suppression/real A+B overlap/server/B compile-link-run checked separately from14direct references. Preserve all accepted native trace/refusal, invocation/query/fault studies and their contracts. Independent original observer-OFF ABBA19x4,72complete instrumented semantic identities, four unavailable off-counter sets; unchanged external no-op paired-median gate worse by BOTH>1ms AND>10%. No same-head retry, sample deletion, workload reduction or weaker evidence check. Historical disposition is explicit BEFORE new measurement: #170's original +1.6946ms/+13.7988% crossing remains unexplained, not cleared by the ADL fix or later smaller medians. #171's completed fixed AA/AB investigation archived the actual runtime images and showed same-binary tails; it did NOT explain the original crossing. #176 is now accepted as diagnostic tooling with unresolved historical C1041 retained. Subject to all current gates and source review, this review may make a similarly limited, explicit risk acceptance for test-only integration, rather than requiring missing historical traces to be reconstructed. This is not automatic acceptance or a performance fix; a newly failed mandatory control/current performance gate still blocks merge. Production cancellation/lease/release reliability and #172 remain unapproved. Next scope after this fixture, not bundled here: typed original-failure/cancellation propagation and real multiworker execution. A failed compiler must not become cancellation-success or permit downstream link/cache publication. |
Resolve #170 against exact main 4fac681, preserving all accepted observer/query/fault capabilities and the original scheduler_drain helper. Retain all direct-drain investigation entry points and explicitly reject their mislabeling as scheduler runs. Default70 keeps original56 plus14 actual API cases; private42 unchanged. Preserve historical failures and performance hold evidence; require new exact-combined-tree native/full release/real studies and independent ABBA before a scoped test-only acceptance decision. No production scheduler/caller, CLI, ProcessSpec, freshness, benchmark or VERSION5.5.0 change. No history rewrite.
Iviesever
changed the title
Combined source and actual Windows record contracts verifiedPublished merge-main head ffaa1a5, base 4fac681; native test-merge 21f2a4a has identical tree e518e8b694df53cd45c72761f386f3b998ab3d2d and exact base/head parents.409files total,405byte-identical to main; only the four declared test/CI files differ. The original scheduler helper and three nonconflicting blobs match corrected old #170. Production sources, parameter/freshness rules and all accepted observation/fault modules remain unchanged. Actual Windows ownership contract artifact 10130123194 (1393096bytes), SHA256 Old initial/corrected #170 ABBA ZIPs have been independently reread: initial external median +1.6946ms/+13.7987894% still crosses the original gate; corrected old median +0.14575ms/+1.6311554% does not. Both data sets remain separate. Their missing historical execution traces are not fabricated; no claim that the old anomaly disappeared. The temporary perf title was set only after the new head push: synchronize#586 is skipped, actual new-head#587 is the independent measurement. No native/CI/benchmark retry was requested. Full real studies/native/ABBA remain pending their own raw review; no merge authorized by these contract results. |
Iviesever
left a comment
Iviesever
left a comment
There was a problem hiding this comment.
Final updated-base review — accept real scheduler fixture integration; retain historical runtime debts
Source/evidence self-review with separate read-only source reconstruction and artifact/statistics replay. This is not an independent human approval.
Exact base: 4fac6817748aabb77654e8ffad9c4f6ce206dcdd.
Reviewed head: ffaa1a530a5a7766c5aeb443ca86fa034f2b8e93.
Tree: e518e8b694df53cd45c72761f386f3b998ab3d2d.
Native test-merge: 21f2a4aa647009025d1e7ea7d025716aa0407f11, identical tree and exactly those base/head parents.
The head is an ordinary merge of old #170 head283b507... and current main4fac681..., with explicit conflict resolution, not rewritten history. Complete canonical base/candidate archives and the four-file patch reproduce the recorded trees. Four test/CI files,160 additions/14 deletions;405of409base files byte-identical. Three nonconflicting files and the scheduler_drain helper are byte-identical to corrected old #170. No production src/include, scheduler implementation/caller, ProcessSpec, benchmark, manifest/native inventory, CLI/freshness or VERSION change. All accepted #171–#176 tooling remains; #172 is not imported.
Verified blobs: workflow fb8d8f745d1348ad5fe9c26e66c5259d5fbefe43; probe 1ad20d942a7f2131591b573904675284bb9a0770; collector c5b3952922dd5c10bdbf9d951f97f0813bc2111c; contract 2a1090d53029be894fc3995758169bc9b19c24b8.
Actual current-source execution
All nine applicable workflows completed successfully on this combined source:
- Native Debug #729/34418829397: actual build, all four paired subshard jobs, ambient/include/runtime/parameter and freshness/side-output checks, final gate.
- Native Release #618/34418829374: Stage0, all four Release shard jobs, full self-host and validation of the exact runtime-only package. publish-release skipped.
- Documentation #179/34418829415, Cross-Stage #206/34418829364, existing PDB Open Investigation #9/34418829368 passed.
- Default Endpoint #22/34418829373, private ownership #24/34418829380, File Event #12/34418829379 passed their actual native jobs and required controls.
- Independent Performance #587/34418867045 ran both exact checkouts, both Release builds, instrumentation check, ABBA and upload. Synchronize-triggered #586 was skipped, not a prior measured batch.
No same-head native/CI/benchmark retry was requested. Old results are not substituted for this combined-source validation.
Real scheduler API boundary —14/14 actual cases
Default 70/70 and private 42/42 complete. The original56default cases remain, including all14direct-drain references; all required unmanaged controls pass. Each of the 14scheduler-drain cases actually calls BoundedWorkScheduler::run_with_admission_stop(2,1,...).
Independently checked each actual A/scheduler.json, A/drain.json, original A/work0 result and parent/B results: worker_count1/started_count1, three stop flags true, event observed/forwarded during the callback, bridge_error0, no callback/scheduler exception, first A compile0, pending TU unattempted. Retained A/B compiler overlap, A process completion, original service survival and original B compile/link/run0 all hold. The actual helper's error checks are retained; no compiler receives a terminating token.
Compared all14scheduler/direct reference pairs: 56source byte comparisons and70argv comparisons agree after normalizing only case-directory identity. Actual default input executable members match recorded binary/source provenance. Seven C++ helper checks and both refusal outputs were inspected. 150/150 ownership-record expectations match (113accepted/37rejected), including original failure-vs-recovery preservation; the collector SHA matches the actual CRLF checkout.
Existing observer/query/fault entry points remain direct-drain only, with an explicit guard against relabeling those studies as scheduler evidence. Callback-active covers compile/capture/diagnostics, not an in-flight RPC. This is still one A worker/one active/one pending; it is not real multiworker coverage, production CLI cancellation, external-service writer enumeration or a write-lease certificate.
Accepted diagnostic capabilities survive the merge
Replayed the exact published trace/invocation/observer/fault auditors against unmodified current raw archives and compared their results with the recorded reports. The original8PCH/Modules,4calibration,4traced+4untraced preexisting and4RM-on+4RM-off studies remain complete. Their 292original tool results/584stdout-stderr records, all four native missing-readiness controls and original success/failure calibrations are retained. 23 ambiguous tool intervals remain ambiguous, not assigned by order. Each normal observation explicitly has scheduler_api_used=false.
The cold-B fault calibration remains separately labeled:2baseline originals exit0,2injected originals produce C1041/exit2, each with2retained-service sharing-failure events;4recoveries separately exit0. Its16results/32diagnostics and negative readiness test passed the original fault oracle. It is not a warm/concurrent historical reproduction or a scheduler failure-propagation test. Across five studies there are28calibrated traces and4untraced workloads, not32successful original compiler outcomes.
Original 187Python expectations (36trace+54invocation+64observer+33fault) pass both Windows and local replay. Actual three observer input binaries hash-match all five studies' provenance. They are not the independent performance binary pair. Recorded runner images differ between some workflows (20260824.214.3 versus20260907.229.1); no identical historical host state is claimed.
Original adverse service policy still unsafe
Current default retains two original adverse managed-policy B failures: /ZI A-started cancel(C1083/C2471) and PCH Release A-started normal(C1090); both B compilers exit2 and link/run remain unattempted. Recovery0 does not replace those results. Private B builds happen to succeed this time. Across both modes all28A-started managed endings still kill the original service. These are preserved negative-policy cases, not failures of the new scheduler/direct positive controls.
Independently checked 719default/417private tool results and1438/834diagnostic files. B-resource identity appears only6/70default and1/42private snapshots; absence is not writer-quiescence evidence. Binary hash inventories (920default/540private) are post-case records, not retained binary bodies or failure-instant snapshots.
Independent unchanged ABBA — all19scenarios
Artifact 10130363823,35690bytes,SHA256 91b4689eab4d2d8cd348a639ad5b9cbd1f05bd6ef15b90abbc18569757ac343b independently downloaded/CRC/hash checked. All76pairs, per-pair arithmetic, AB/BA/AB/BA orientation, paired medians/MAD/nearest-rankP95 and marginal medians were recomputed. All72instrumented complete counters/breakdowns/compile-link-archive hit-miss agree. Four timings-off external counter sets are unavailable, not zero. Only the final row uses external elapsed; four-pairP95 is merely the maximum, not a population-tail estimate.
| Scenario | Paired median Δms | Paired median Δ% | MAD ms | Paired P95 Δms | Slower |
|---|---|---|---|---|---|
| cold | −18.175 | −1.07 | 114.804 | +189.549 | 2/4 |
| no-op | −0.002 | −0.03 | 0.063 | +0.099 | 2/4 |
| single-tu | −36.291 | −26.08 | 24.422 | −3.581 | 0/4 |
| public-header | +4.392 | +4.40 | 12.643 | +25.876 | 3/4 |
| build-run | −1.310 | −5.87 | 1.693 | +0.504 | 2/4 |
| link-only | −25.770 | −31.75 | 14.509 | −1.557 | 0/4 |
| target-scale-cold | +8.788 | +0.30 | 688.823 | +1166.507 | 2/4 |
| target-scale-no-op | −6.529 | −13.64 | 7.185 | +1.858 | 1/4 |
| target-scale-no-op-auto | +0.297 | +0.92 | 1.527 | +2.009 | 2/4 |
| target-scale-no-op-j1 | +1.671 | +3.81 | 1.128 | +14.445 | 4/4 |
| target-scale-common-header-no-op | +0.201 | +0.55 | 0.660 | +28.413 | 2/4 |
| target-scale-single-tu | +13.539 | +11.01 | 3.294 | +262.485 | 4/4 |
| discovery-cold | −0.018 | −0.00 | 39.441 | +3104.250 | 2/4 |
| discovery-no-op | +2.657 | +32.41 | 2.681 | +30.829 | 3/4 |
| discovery-header | +17.762 | +16.92 | 17.305 | +413.239 | 3/4 |
| modules-cold | −33.952 | −1.75 | 109.745 | +764.338 | 2/4 |
| modules-no-op | +0.149 | +3.01 | 1.071 | +4.162 | 2/4 |
| timings-enabled-no-op | +0.035 | +0.92 | 0.070 | +0.263 | 2/4 |
| timings-disabled-no-op (external) | +0.100 | +0.89 | 1.236 | +2.512 | 2/4 |
Unrounded external paired median +0.0995ms/+0.8877037% does not cross the unchanged BOTH>1ms AND>10% gate. No speedup or zero-overhead claim. Retain the+3.104second discovery-cold tail,4/4slower scale-j1/scale-single-TU and all other favorable/adverse samples; no presumed-noise deletion, AA subtraction or offsetting one scenario against another. The comparison JSON does not archive the actual ABBA executable pair or OS trace.
Historical disposition and final decision
Pre-measurement scope and limited-risk policy explicitly allowed a test-only decision if all current gates passed; they now do.
The original f52a8ad external +1.6946ms/+13.7988% crossing remains unexplained. Recomputed both old raw ABBA ZIPs again: the first gate is still crossed and corrected old283b507 is below threshold; a later lower median is not a performance fix. Their raw hashes remain f751e428... and f2479c1e..., with initial C2676 probe-build failure ZIPs retained too. #171's accepted fixedAA/AB investigation and archived runtime images narrow the later observations but do not explain the original missing traces. #176's actual cold-PDB failure calibration likewise does not fix the historical warm/concurrent C1041.
Accept ordinary expected-head merge of ffaa1a530a5a7766c5aeb443ca86fa034f2b8e93, subject to unchanged main and protection. This intentionally changes the PR-level HOLD into limited acceptance of test-only real-scheduler integration. It does not declare historical performance/root-cause requirements satisfied, waive a current failed check, approve #172, complete M1b or authorize production cancellation/lease/release. Original positive controls stay enabled and future failures still block their applicable gates. VERSION5.5.0 and historical tags/assets remain unchanged.
Next return to a separate small implementation of typed original-failure/cancellation propagation and real multiworker scheduling. A failed compile must remain a failure and block downstream link/cache publication. No additional observer expansion is required by this acceptance. External writer authority, project transactions, build/run separation and sessions/residency/release retain their original dependency gates.
Reproduction and limits
Twelve current coreZIPs were CRC/SHA verified, plus four unmodified historicalABBA/build-failureZIPs retained; this is not every Actions artifact. Default10030535114 is NOT the artifact id: the correct current default artifact is 10130535114,SHA256 44f576389e637b7a48179cc3b9d379afdba5ab3fda86c9df4f6f1b66073dfb2f; private10130368557 SHA256 d0b35faa941d3e7026bac67ddc1fead1e5b72b5b3c6a65bc6e4f65d8193ceca6. Full current manifest, source archives, patches and read-only audit outputs are retained in delivery. Actions retention30days.
Local Win32 declaration-shim syntax/Python replay is not local MSVC/ETW execution. A temporary-path equality check initially rejected two expected ENOENT diagnostics because Linux/Windows randomized temp paths differ; all non-path fields and exact missing relative filenames/ENOENT remain checked, and both raw variants are retained. An ad hoc inspection syntax typo was corrected before execution. No repository source, native records, assertions in the product tests or CI measurement changed. An initial unreferenced transcribed blob was rejected before any commit; only the verified1ad20d94... blob entered the published tree. No evidence is relabeled successful by these local audit corrections.
Title/ready-triggered skipped performance jobs cannot replace actual#587. Post-merge push CI is not claimed already completed by this review.
Iviesever
changed the title
… mapping Accept the bounded read-only inventory after exact-tree Native Debug #749, full Release/self-host/package #630, 28 pure inventory and 33 mapping plus 39 unchanged domain checks, original default70/private42 and real batch/target-cache controls, and independent ABBA #619. Preserve unresolved native effects, per-entry errors, missing-parent/leaf and case-alias distinctions; no complete write-set, reservation/recovery or CLI lease authority. Retain all adverse managed-service and timing results and historical #170/#599/C1041 debts. VERSION5.5.0, #172 HOLD, historical tags/assets and release status unchanged.
…ection Accept exact base eaf3437 / head a012b2b, tree 4f5594d, after substantive review 5190239753. Native Debug #751 and Release/self-host/package #631; actual new44/original28/target84 and domain39/mapping33 checks in both configurations; original default70/private42 positives and batch6/target6; unchanged independent observer-OFF ABBA #622 (76 pairs, 72 equal complete instrumented vectors). Preserve default2/private3 unsafe cleanup failures, all diagnostics and adverse timing samples including +3.9745 ms/+34.5689858%. Only collection APIs, no CLI cancellation/write lease/recovery authority. Begin 5.6 feature freeze and cumulative correctness/compatibility closure; #164 retains the long-term 6.0 roadmap. Historical C1041/#170/#599/common129 risks remain release-blocking; #172 HOLD. VERSION remains 5.5.0. No release authorization, historical tag or asset change.
…regression Accept test-only base b7e322c / head a876eba, tree 280326b, after substantive review 5190397138. Debug/Release 79 native programs each; four new sequences/32 actual calls; original failure run34748925471 and all not-run records retained. Pinned-seed Stage0/1/2 self-host, seven-file package and installer lifecycle verified. Independent original ABBA625: 76 pairs, 72 matching complete instrumented vectors; external median +0.79255ms/+6.8856773% does not trigger the predeclared conjunctive median gate. Preserve adverse external +63.0149ms/+392.6039687% and +1.4072ms/+12.2515432%, plus all other tails; no zero-overhead claim or history erasure. Only test collector/workflow added; product freshness and VERSION5.5.0 unchanged. Continue cumulative correctness/compatibility closure, especially actual PCH target configurations and failure-repair coverage. #172 HOLD; historical C1041/#170/#599/common129 remain release-blocking. #164 retains long-term6.0 roadmap; no release authorization, CLI cancel/write-lock default activation, historical tag or asset changes.
1 participant
Final decision: accept test-only scheduler integration; retain historical runtime debts
Final exact-source, complete execution, raw-artifact,19-rowABBA and limited-risk acceptance review
Accept ordinary expected-head merge. This explicitly revises #170's PR-level HOLD for the test integration only. The original +1.6946ms/+13.7988% performance crossing remains unexplained, and the historical C1041 is not fixed. All current mandatory gates pass; no current failed check is waived. #172 remains independently held, M1b incomplete, and production cancellation/lease/release acceptance remains separate.
Exact source
4fac6817748aabb77654e8ffad9c4f6ce206dcdd(test: calibrate native PDB failure and observer evidence #176).ffaa1a530a5a7766c5aeb443ca86fa034f2b8e93.e518e8b694df53cd45c72761f386f3b998ab3d2d.21f2a4aa647009025d1e7ea7d025716aa0407f11, exact base/head parents.Delivered integration and boundary
The actual
BoundedWorkScheduler::run_with_admission_stop(2,1,...)fixture now coexists with all accepted #171–#176 evidence capabilities. Its bridge requests admission stop while the first callback is active; the already-started compiler naturally completes and the pending TU is not dispatched. B compiles/links/runs in its own output/PDB space against the same service. Neither compiler gets a terminating token. Callback/bridge/scheduler errors remain failures.The default matrix keeps all original56cases and adds14scheduler-drain controls; private42 is unchanged. Existing observer/query/fault entry points stay direct-drain controls, with an explicit guard against mislabeling them as scheduler runs. No new observer or #172 code is introduced. Callback-active includes capture/diagnostics, not proof of RPC-in-flight; this remains one A worker/one active/one pending, not real multiworker or production CLI cancellation.
Completed current combined-source validation
Each of14scheduler records was checked against the actual API summary, original A/work0 result, no pending dispatch, real A/B overlap, A retained-handle completion, service survival and original B compile/link/run0. The14reference pairs match56source files and70argv records after only directory normalization. Seven native helper checks and both refusal outputs are retained. Green summary alone was not used as proof.
The five accepted studies were replayed using the exact published auditors. They retain28calibrated traces+4untraced workloads, five native missing-readiness controls,308tool results/616diagnostics and23ambiguous invocation intervals. Two separately injected cold-B originals still produce C1041/exit2 with retained-service sharing failures; their recovery0 is not substituted for the first failure. These injections are not historical reproduction or scheduler failure-propagation evidence. Actual observer input binaries match all five provenance records, but are not the ABBA executable pair.
Original default/private matrices preserve two/zero adverse managed-policy B failures in this batch. Default /ZI A-started cancel(C1083/C2471) and PCH Release A-started normal(C1090) retain original exit2 and unattempted link/run; recovery does not replace them. All28A-started managed endings still kill the original service.719/417tool results and1438/834diagnostic files were checked. Sparse owner snapshots are not a no-writer certificate.
Performance and historical-risk disposition
Scope, original gates and risk policy declared before this measurement
Current external timings-off no-op paired median +0.0995ms/+0.8877037%, maximum adverse+2.512ms,2/4slower: unchanged gate worse by BOTH>1ms AND>10% not crossed. Preserve discovery-cold+3104.250ms,4/4slower scale-j1 and scale-single-TU, and every other favorable/adverse sample in the full review. No speedup/zero-overhead claim, noise deletion, AA subtraction or historical debt closure. Four external counter sets are unavailable, not zeros; only that row uses external elapsed; four-pairP95 is a maximum.
Both original #170 raw ABBA ZIPs were recomputed again: f52 still crosses the original gate,283b remains below it. The ADL compile correction and later smaller medians do not explain that crossing. #171's accepted fixedAA/AB study and archived runtime images are additional later evidence, not reconstruction of missing historical traces. #176's cold-file injection does not explain historical warm/concurrent C1041. All those debts remain open for production/release acceptance.
This is limited acceptance of test-only integration, not fulfillment of historical root-cause requirements. Current correctness/performance requirements were not lowered. Original positive controls remain enabled; future failures still block their relevant gates. No same-head native/CI/benchmark retry was requested. #586 was a skipped synchronize trigger, not another measured batch; actual#587 supplies this comparison.
Handoff and retained evidence
Next separate iteration: typed original-failure/cancellation propagation and real multiworker scheduling. A failed compiler must not become cancellation success or permit downstream link/cache publication. This is not implemented or authorized by the present fixture. External-writer authority, single-writer project transactions, build/run separation, sessions/residency and final VERSION/release remain dependent gates. #172 is not approved here.
Delivery retains12current coreZIPs and4unmodified historicalABBA/build-failureZIPs, exact source archives, patch, manifests and replayable read-only audits. Not every Actions artifact is included. Retention30days; generated compiler binary bodies are excluded from diagnosticZIPs and finite hash inventories are post-case, not failure-time snapshots. The Python auditor checks the accepted decoded ETL stream, not a second binary ETL decoder. No local MSVC/PowerShell/ETW execution is claimed.
Local cross-platform audit development initially compared randomized temporary-path error strings too strictly; the two exact ENOENT cases retain both raw errors and require the same relative filename and all other fields. No source/native evidence/CI outcome changed. A bad unreferenced transcription blob was rejected before committing; only verified probe blob1ad20d94... entered the source tree. Title changed only after actual#587 completed; later skipped perf jobs and post-merge push CI are not substituted for completed exact-tree validation.