test: verify real target admission preserves cache barriers - #179
Conversation
Drive the accepted explicit target API through actual compiler/linker/Windows process execution. Six fixed Debug/Release complete/cancel/failure-cancel cases preserve cold/warm controls, raw diagnostics, typed source outcomes and target/source cache byte/mtime/file-identity snapshots. Stopped targets cannot link or publish their target cache; completed TU caches may remain valid without claiming project rollback. Add 32 record-contract expectations and isolated build/observe CI. Product files, original benchmark/ownership gates and VERSION5.5.0 unchanged.
Exact source and first Windows record gate checkedBase Windows contract job 102758109379 in Target Admission#1/34441769199 completed. Downloaded artifact 10138183315,1157bytes,SHA256 Scope review retains actual Debug linker policy, real product compile/link inspection and complete source-index mapping. Native cache/executable snapshot handles are closed before target operations and never serve as locks. Raw cache bytes are copied for independent hashing; executable bodies are compared in-process and represented by native hash/mtime/fileID records, not archived in this diagnostic ZIP. The fixture makes no all-writer,transaction rollback or CLI integration claim. Native probe build and six-case execution, original70/42 and six-batch controls, full native/selfhost/package and original independent ABBA remain separately required. No post-submission source edit or same-head rerun has been requested. |
Actual target API / MSVC / cache barrier evidence independently replayedExact base 6/6 real target cases pass on their first attempt. Each calls the actual explicit target entry for cold/warm/subject phases, backed by the real product compile executor/linker/cache coordinators and unmodified WindowsProcessRunner requests. All six cold builds execute3compiles+link+program successfully. All six warm calls perform0compile/0link and preserve all five artifacts' byte/hash/mtime/file identity records.
The two failure cases preserve the exact original stdout/stderr/static_assert marker and source1 identity through the nested target error and per-source attempt, even while admission_stop_observed=true. No third native result is invented, no successor is hidden, and no recovery run overwrites these original failures. These separately labeled static_assert cases are not historical C1041 reproduction. In all four stopped cases, existing target executable and actual link-cache remain identical in bytes,mtime,volume and file ID. Pending source2 cache remains unchanged; failed source1 cache also remains unchanged. Completed sibling caches change and are valid according to actual read-only product inspection. The old target cache being preserved does NOT mean it is fresh: actual link inspection requires relinking in all stopped cases. Completed subjects publish new target-cache snapshots, actually link/run, and return fully reusable source/target state. Inspection itself changes none of the snapshots. This is product cache behavior, not a fixture commit sentinel or project rollback. Independently checked 48original compile/link/run results,96stdout/stderr files,96raw cache snapshot copies and120artifact snapshot records. Executable bodies are compared in native memory and represented by recorded SHA/identity; they are not separately archived in this diagnostic ZIP. All six subject runs retain two distinct real cl.exe handles/image/creation identities before the stop decision and observe their completion on joined return. No stoppable ProcessSpec was accepted; no compiler result is labelled terminated. This still does not prove all shared-service writers have stopped or authorize a lease. 24input source comparisons and20executed-argv comparisons agree across modes after removing only the declared failure marker and normalizing fixture directories. Two12000-function sources and pending source remain fixed; effective /Zi,/FS,/bigobj and /MTd or /MT retained. The recorded argv is the exact forwarded ProcessSpec, not separate OS command-line telemetry. Snapshot handles close between invocations and are not locks. Actual input artifact 10138340324,537669bytes,SHA256 This closes the new native target gate, not the full merge decision. Original native Debug/Release,self-host/package,default70/private42,the six underlying batch cases and independent ABBA remain separately required. No post-publication source edit or same-head native/CI/benchmark rerun requested;VERSION5.5.0 and historical/#172/CLI/lease limits remain unchanged. |
Independent ABBA and original tool regressions retained in fullBase All76pairs,AB/BA/AB/BA order,stored arithmetic,paired medians/MAD/nearest-rankP95 and marginal medians agree. All72instrumented complete counters/breakdowns/compile-link-archive hit-miss values match;four external counter sets are unavailable,not zero. Only the final row uses timings-off external elapsed;other rows use mqb.timings. Four-pairP95 is the maximum,not a population-tail estimate.
Unrounded external paired median +0.22385ms/+1.9108219% does not cross the unchanged BOTH>1ms AND>10% gate. No speedup/zero-overhead claim. Keep timings-enabled-no-op4/4slower,public-header/cold3/4slower and the common-header+12.482ms tail. Smaller tails than older runs do not resolve historical C1041/performance debts. Product source and original benchmark are unchanged;favorable values are not an attributed optimization. No noise deletion,AA subtraction or same-head remeasurement. The comparison JSON is not an archive of the actual benchmark binary pair or OS trace. Separately replayed original default 70/70 and private 42/42,with all required positive unmanaged/direct/scheduler controls passing. Actual input executables match their provenance;all14scheduler/direct pairs match56source files+70argv records after directory normalization. 150/150 ownership record expectations match113accepted/37rejected and the exact Windows CRLF collector hash. New target study does not replace these gates. Preserve original nine default +five private adverse managed-ending B failures(C1090/C1051/C1083/C2471 as recorded),failed/unattempted link/run and separate recovery results. All28A-started managed endings still kill the original shared service;no policy authorization follows. Audited705default/407private original results and1410/814stdout-stderr files;owner identity14/70and4/42 is incomplete,not writer-quiescence evidence.895/521binary inventory entries are post-case hashes,not incident-time snapshots. The original six true-MSVC batch regressions also pass independently and preserve their two injected C2338/exit2 originals as failed despite cancellation;their sentinel remains fixture-only,distinct from the new target's actual cache evidence. All original batch result/diagnostic files and exact input bytes were checked. Core default artifact10138433170 SHA256 |
Iviesever
left a comment
Iviesever
left a comment
There was a problem hiding this comment.
Final review — accept real target API/cache end-to-end coverage
Implementation self-review with separate read-only source reconstruction, native-artifact inspection and statistics replay; not an independent human approval.
Exact base: d863b1a5dbab2cc1d8b1c09ae9ced5953fa52d64.
Reviewed head: 2b0b22b918a4ffc83a711288c6172c36cef85a0b.
Candidate tree: 24346d15b7f245396c972dfd2478b2ef0ecaa3a4.
Native test-merge: 342a4c75df929255f13bb1b51ab427fa2a95d66a, identical tree and exactly those base/head parents.
Four new test/CI files, 933 additions/0 deletions; all414existing files remain byte-identical,418candidate files. Canonical base/candidate archives and exact patch reproduce both trees. Product implementation, scheduler, CLI, source manifest/native79 inventory, original benchmark/ABBA and accepted observer tools are unchanged. VERSION5.5.0 remains unchanged.
Verified blobs: workflow 6fa38adc1e6865e512d0faf273edc8e3eeb316ce; native probe 8c5454ba090e4d75b8dc0667be62703258986cb6; collector 8b965002adb820799f17260bbcc1aa3c4c0349bf; verifier eabf529d99dac5cce351b339c6867c2929e040b1.
Actual target path and source review
The fixture constructs the real compile executor, compile coordinator, linker, link coordinator and target coordinator. Cold, warm and revised-source subject phases all call run_with_compile_admission_stop. The recording runner forwards the unchanged product ProcessSpec to WindowsProcessRunner, saves original argv/exit/stdout/stderr and rejects a stoppable token. It does not fabricate compiler output or substitute target/cache implementations.
The fixed three-source executable target has two12000-function active inputs and one pending source,2workers and a shared header. The first two subject invocations synchronize before starting; actual retained cl.exe image/creation identities and simultaneous liveness are then required before the stop decision. Missing overlap fails the fixture instead of changing the input or repeating it. Effective /Zi,/FS,/bigobj and /MTd or /MT are preserved; original /Z7 precedes /Zi. Actual Debug link argv keeps /DEBUG and /INCREMENTAL, while Release keeps /INCREMENTAL:NO. No linker-policy workaround or compiler Job is added.
Actual target link-cache, target executable and three source caches are read for bytes/SHA256/mtime/volume/fileID between joined operations. Handles close before target execution; they are not locks. Raw cache bytes are archived. Executable bytes are compared in the native process and represented by its digest/identity records, not archived as independent executable-body snapshots. These finite records do not prove all external writers have stopped.
All nine current-head workflows completed successfully
- Native Debug #735/34441769183: actual current Debug MQB/library build, all four paired subshards, ambient/include/runtime/parameter and freshness/side-output checks, final native gate. Step execution was inspected; not a classifier-induced skip.
- Native Release #621/34441769137: Stage0, all four Release shard jobs, full self-host, runtime-only staging, validation of the exact packaged artifact and upload. publish-release skipped. Self-host/package job102760001487 completed, not merely queued.
- Real Target Admission #1/34441769199: actual PowerShell parsing and native probe compilation,32record expectations and all six real target cases passed on separate build/observe VMs.
- Typed MSVC Batch #3/34441769275: all six existing real two-worker batch regressions passed independently.
- Default Endpoint #25/34441769205 and **Private Ownership #27/34441769115:**70/70 and42/42 complete, required positive controls passed.
- Documentation #182/34441769153, Cross-Stage #212/34441769122 passed.
- Independent Performance #596/34441769135: exact head/base checkouts, both actual Release builds, instrumentation check, original19x4ABBA measurement and upload completed.
The unchanged File Event/PDB Open research workflows were not triggered and are not described as current rerun passes. No post-publication source correction, same-head native/CI/benchmark retry or extra sample was requested. Native build diagnostics retain existing C4100,C4996 and D9025 warnings; this is not a warning-free/WX claim.
Six actual target cases — original errors and real cache publication
Complete native target/cache audit
Every case first performs a real cold3compile+link+program run, then a warm target call with0compile/0link and all five artifact snapshots unchanged. Subject results, in both Debug and Release:
| Subject | Original compiler exits | Target outcome | Subject link/run | Source reusable | Target link reusable |
|---|---|---|---|---|---|
| complete | 0,0,0 | success | 1/1 | true,true,true | true |
| cancel | 0,0; third not entered | cancelled(code9) | 0/0 | true,true,false | false |
| failure-cancel | 0,2(C2338); third not entered | compile_failed(code7) | 0/0 | true,false,false | false |
The two intentionally injected static_assert failures preserve their exact raw stdout/stderr, marker, source1 attribution and nested native exit through the target error and per-source attempt while admission_stop_observed=true. They are failures, not cancelled-success. No third result or successor is invented and no later recovery replaces the original failure. The static_assert is not a historical C1041 reproduction.
All four stopped cases preserve the old actual target executable and link-cache bytes,mtime,volume and fileID. Pending source2 cache remains unchanged; failed source1 cache also remains unchanged. Completed siblings may retain changed, valid source-cache entries. Actual product read-only inspections classify those completed sources as reusable but the failed/pending sources and unlinked target as dirty. Thus preserving an old cache does not incorrectly satisfy freshness. Inspection itself changes none of the snapshots. Successful subjects actually relink/run and publish their target cache, rather than merely passing a fixture sentinel.
Independently replayed48original compile/link/run result files,96stdout/stderr records,96raw cache copies and120artifact snapshot records. Full recomputed reports equal the saved native audits/summary.24input-source comparisons and20executed-argv comparisons agree after only case-directory and declared static_assert normalization. Six subjects retain two unique real cl.exe identities and observe completion on target return. These are exact forwarded ProcessSpec arguments, not separate OS command-line telemetry.
Actual native input probe SHA256 dbe7cf88b191f16c8a59783221e2d5527a8e11e67ff460636715451521de1a54 matches source/build/observer provenance. Toolchain MSVC14.51.36231/Hostx64-x64,Windows10.0.26100,PowerShell7.6.5,image20260824.214.3. The probe input is not the independent ABBA executable pair.33generated binary inventory entries are post-case hashes, not failure-instant snapshots.
Positive and negative record checks both exercised
Windows contract artifact10138183315 was independently checked: 32/32 expected outcomes,7accepted including exact raw CRLF and25damaged-record refusals. This is not an auditor that passes by rejecting everything. Missing/extra native outputs, hidden link, wrong source/outcome/flags, pending-as-success, stale-cache-as-fresh and changed snapshot bytes/mtime are rejected. Synthetic data tests record validation, not native execution. The actual published verifier was replayed over all six native cases, not only its self-test.
Original regressions and adverse shared-service policy remain visible
Independent ABBA and original-matrix audit
Default70/private42 retain every required unmanaged/direct/scheduler positive control.150ownership-record expectations match113accepted/37rejected and the actual CRLF collector hash. All14scheduler/direct pairs match56sources+70argv after directory normalization. Prebuilt executable provenance, original seven-helper and refusal records were checked. The unchanged six true-MSVC batch cases separately retain their two original C2338 failures and fixture-only successor gate; they are not a substitute for the new actual target cache evidence.
Original unsafe managed-ending experiments preserve nine default and five private B first-build failures, exact C1090/C1051/C1083/C2471 diagnoses, unattempted link/run and separate recoveries. All28A-started managed endings still kill the original service. Audited705/407tool results and1410/814diagnostic files. B-resource identity14/70 and4/42 is incomplete, not quiescence evidence.895/521binary inventories are post-case observations. None of these negative-policy records authorizes per-cl termination or safe writer-lease transfer.
Independent original ABBA — no altered thresholds or erased samples
The full19-row table and all sample details are in the linked comparison review. All76pairs,AB/BA/AB/BA order,stored deltas,paired medians/MAD/nearest-rankP95 and marginal medians were recomputed. 72/72 instrumented complete counters/breakdowns/compile-link-archive hit-miss identities match;four external timings-off counter sets remain unavailable,not zeros.
External no-op paired median +0.22385ms/+1.9108219% does not cross the original BOTH>1ms AND>10% gate. No speedup or zero-overhead claim. Keep timings-enabled-no-op4/4slower,public-header/cold3/4slower and common-header+12.482ms tail. Favorable samples are not attributed to optimization; all product source and the original benchmark are unchanged. Smaller current tails do not explain historic C1041/performance debts. Only the external row uses timings-off elapsed; four-pairP95 is the maximum,not a population-tail estimate. No noise deletion,AA subtraction,same-head remeasurement or historical-debt closure. Comparison JSON is not an archive of the actual ABBA executable pair or OS trace.
Decision and next handoff
Accept ordinary merge with expected head 2b0b22b918a4ffc83a711288c6172c36cef85a0b, subject to unchanged main and branch protection. The originally declared current gates pass without a waiver. This closes the true-MSVC/link/cache end-to-end coverage gap for the tested three-source executable target API. It does not change CLI defaults, validate every target mode, complete M1b or settle historic reliability/release debts. VERSION5.5.0,#172HOLD and all writer/transaction limitations remain.
Current scope excludes DLL/static/modules/PCH-parent,small/forced paths,late-link cancellation,partially started worker failure and owner-crash recovery. Existing deterministic tests for some boundaries are not relabeled native coverage. Target return and finite file snapshots do not authorize a project lease.
Next return to #164's foreground build/run separation in a separate small implementation: isolate successful build completion from foreground artifact execution, preserve argv/cwd/environment/stdin/output/exit-code behavior and identify the future transaction boundary before starting the program. Do not add a new observer or enable CLI cancellation/leases merely from this test result. Ownership of discovery/toolchain/scan/compile/link/archive writes and abnormal recovery remain later explicit transaction requirements.
Reproducibility and retention
Eleven immutable core artifact ZIPs were SHA256/CRC-checked; complete canonical base/head archives,exact patch,metadata and portable read-only replay also pass. Target10138373075 SHA256 89fa8306566d45335b6af46091040033871e2f16f982679a44f5b8f2bd13fca9;ABBA10138389083 81091ed2652f7534e9bd428b5ebbde00f21bca57926b92d92231a28eaae7c805;default10138433170 40a1d5cfda29cf4a1d7df28d2c1f264a6d9f4373621836c4fccdd8ac89185a5d;private10138431607 ffe5fef0ef13da623fee48e0fe62eb833bfb0098463bb2d4e29077b26c7e47ba. This is the core evidence,not every Actions artifact; retention30days. Generated executable bodies are excluded from diagnostic archives,not silently described as retained bytes.
Local Python32checks and strict GCC syntax with actual product headers/declaration-only Win32/BCrypt shims are not local MSVC/PowerShell/ETW execution. Prepublication generic-path/raw-byte fixture corrections and local shim/formatting diagnostics remain documented; only final verified blobs entered the commit. No published failure or measurement was overwritten. Later title/ready-triggered skipped perf jobs cannot replace actual#596; post-merge push CI is not claimed already completed here.
Iviesever
changed the title
1 participant
Final decision: accept actual MSVC/link/cache coverage of the target API
Final exact-source review, all current gates and acceptance decision
Accept ordinary expected-head merge. This test iteration closes #178's real-tool target-endpoint coverage gap without modifying product behavior or enabling CLI cancellation. #172 remains independently held; M1b and historic C1041/performance requirements remain open.
Exact source
d863b1a5dbab2cc1d8b1c09ae9ced5953fa52d64.2b0b22b918a4ffc83a711288c6172c36cef85a0b.24346d15b7f245396c972dfd2478b2ef0ecaa3a4.342a4c75df929255f13bb1b51ab427fa2a95d66a, exactly those base/head parents.What actually runs
The fixture uses the real product target,compile,link and cache coordinators. Its runner records and forwards the unchanged ProcessSpec to WindowsProcessRunner and rejects any stoppable process token. There are no simulated compiler outcomes or substitute cache implementations.
Six fixed ordinary executable targets: Debug/Release each complete,cancel,failure-cancel. Every target calls the explicit API for a cold build and actual program run, a zero-tool warm no-op, then a revised-source subject build. Two12000-function sources and one pending source,2workers,shared header,/FS/effective /Zi,/bigobj and correct MTd/MT are preserved. Actual Debug link keeps /INCREMENTAL;Release /INCREMENTAL:NO. Cancellation follows observation of two retained live cl.exe identities. Missing overlap fails the fixture instead of growing or repeating it.
Completed current-head gates
All nine current-head workflows completed on their first attempt. Untriggered old File Event/PDB Open studies are not described as rerun. No post-publication source correction or same-head native/CI/benchmark retry. Existing C4100,C4996,D9025 build warnings remain in diagnostics; no warning-free claim.
Actual failures, cancellation and real target cache barriers
Full native target/cache and provenance inspection
Both complete subjects compile0,0,0,actually link/run and publish their target cache. Both cancelled subjects preserve original0,0 and unentered third item,report cancelled and invoke no link/run. Both failure-cancel subjects preserve original0,2/C2338,exact stdout/stderr/source1 and unentered third item,report compile_failed,not cancellation-success. These static_assert failures are explicitly injected,not historical C1041 reproduction.
All four stopped targets preserve actual old EXE/link-cache bytes,mtime,volume and fileID. Pending sourcecache remains unchanged;failed sourcecache also remains unchanged. Completed siblings may save valid source caches. Real product read-only inspection marks failed/pending sources and the stopped target dirty while completed sources remain reusable. Unchanged old cache is not falsely accepted as fresh. Inspection changes none of the snapshots. This is not project rollback or a writer-lease guarantee.
Independently replayed48native result files,96diagnostic files,96raw cache copies,120artifact snapshot records,24source comparisons and20executed-argv comparisons. Native executable bytes are compared in-process with digest/identity records but not archived as independent body snapshots. Exact input probe hashes match source/provenance;input binaries are not the ABBA executable pair. All six targets use real selected cl.exe handles and joined completion,no terminating token.
Original negative evidence and performance retained
Full19-rowABBA and original matrix audit
Default/private matrices retain nine/five original B failures in the intentionally adverse managed-ending policy,exact diagnostics,unattempted link/run and separate recoveries. All28A-started managed endings still kill original services.705/407original results and1410/814diagnostics checked. Sparse owner identity14/70 and4/42 is not complete-writer evidence. All14scheduler/direct input pairs and150typed record expectations match;these independent gates were not replaced by the new target study.
External timings-off no-op paired median +0.22385ms/+1.9108219%,2/4slower,does not cross the unchanged BOTH>1ms AND>10% gate. All72instrumented counters/breakdowns/compile-link-archive hit-miss match;four external counter sets unavailable,not zero. Only the final row is external elapsed;four-pairP95 is a maximum. Keep timings-enabled-no-op4/4slower,public-header/cold3/4slower and common-header+12.482ms tail. No speedup/zero-overhead claim,noise deletion,AA subtraction or historical debt closure. Comparison JSON is not an OS trace or actual ABBA binary-pair archive.
Original acceptance criteria and handoff
The predeclared exact-source fullnative/selfhost/package,32record expectations,6real cold/warm/subject targets with original result/cache/freshness checks,original70/42positive controls,six true-MSVC batch cases and independent19x4ABBA/72semantic identities/unchanged threshold all passed. No current gate was waived. Successful new controls do not explain historic C1041 or performance incidents.
Next separate small implementation: foreground build/run separation from #164. Isolate build completion from launching the built program while preserving argv,cwd,environment,stdin/output and exit codes. Establish the future transaction boundary before foreground program execution without claiming that compiler exit proves external-writer quiescence. Do not add another observer,enable CLI cancellation or introduce leases through this test change. Current native coverage is three-source executables,not DLL/static/modules/PCH-parent,small/forced,late-link or partial-worker-start/owner-crash cases.
Eleven immutable coreZIPs,canonical base/head source archives,exact patch,metadata and executed portable replay are retained;not every Actions artifact. Retention30days. Generated target binary bodies are excluded from diagnostic ZIPs;cache bytes retained and finite post-case hashes do not represent incident-time state. Local32record tests and declaration-only Win32/BCrypt GCC syntax checks are not localMSVC/PowerShell/ETW execution. Prepublication path/newline/shim corrections are documented;only verified blobs entered the commit. Temporaryperf title changed only after actual#596completed;later skipped jobs or new post-merge push CI are not substituted for these completed results.