Files
MobileGL/docs/Disaggregated/notes/p3a/p3a-results/gates-review-v1.md
T
swung0x48 aba329cc53 [Docs] (Disaggregated): P6's plan and contract draft; the working notes move from a local directory into the repository
Two things land together: the plan and contract draft P6 needs before it starts, and the
working notes that until now existed on exactly one machine.

1. P6's plan and contract draft (neither started; not one line of code)

     docs/Disaggregated/P6-SPAWN-PLAN.md      the package plan
     docs/Disaggregated/P6-CONTRACT-DRAFT.md  the contract draft (moves to
                                              MG_Remote/CONTRACT-P6.md when c6 lands)

   The three findings that mattered, all written into the documents:

   a. P6 is SMALLER than its roadmap cell suggests. The transport primitives are already in
      the tree and already tested: SocketDoorbell including hangup detection and the two
      lost-wakeup fences (Doorbell.cpp:211-360, four cases in FdPassingTest), SCM_RIGHTS fd
      passing, ShmSegment::Adopt, and the SurfaceOp / WindowKind schema already defined in
      protocol.fbs. inproc uses a real second shm mapping rather than an alias, and
      SessionRings.h:22-28 says that was done so today is not the first time it runs. P6 is
      assembly plus a control plane plus a process, not "write a socket transport".

   b. The evidence for "P6 is only a transport swap" has EXPIRED. It was bought by P5c, whose
      audit found 59 direct memory accesses in a tree that already looked clean. P5e changed
      the tree again. So the first package, a6, is a read-only audit that writes no code, and
      every contract row that depends on it is marked PENDING a6 rather than decided in
      advance.

   c. P6 does not get a game running on the phone. ANativeWindow* is meaningless in the child,
      and protocol.fbs:186-196's own comment says Android transfers the window out of band -
      that band is P12. P6 lands pbuffer / surfaceless / offscreen, which is what HeadlessGL,
      the split-equivalent lane and trace replay need, and which is exactly what P6's own exit
      gate measures. A real window reaching the server is a named refusal. This gets its own
      section because "P6 is done" would otherwise read as "the game runs now".

   The contract adds rule G: a role may not name a process-local handle. It extends P5c's
   rule E (a role may not name the other role's memory) from memory to identity - EGLDisplay /
   EGLSurface / EGLContext / ANativeWindow* / fd number / pid / mapped address either cross as
   a token this contract defines or they do not cross, and a descriptor crosses only through
   SCM_RIGHTS, where the kernel renumbers it.

   What P5e left for P6 gets its own section. The latch semantics are already designed
   (ARCHITECTURE.md:504) and are copied unchanged. What is new is that after run-ahead the
   client no longer waits while the peer process can DIE: the client holds records that were
   published and will never be applied, and Android's freezer suspends the child without
   closing anything. So slow and dead must be distinguishable WITHOUT a timeout threshold.
   That is P6's only new design; everything else is carriage. It shares a predicate with E1's
   outstanding re-specification (ID-122), and the two should be designed together.

   Two gates carried over from P5e's lessons: G2 requires integration-spawn's test-name set to
   match integration-split's name for name, because a differing set means the lane is quietly
   running less; and the arm-proof gate requires every entry to record the child pid and
   transport=spawn in its OWN private log - ID-124 established that this kind of false green
   really happens, and a spawn lane that actually ran monolith would pass every other gate.
   S2's shape must be settled before it is written, and the document says why: E1 is the
   cautionary case.

2. The working notes move into the repository (docs/Disaggregated/notes/, 415 .md files)

   ROADMAP / CURRENT_STAGE_PROGRESS / MEASUREMENTS cite ruling numbers, audit line numbers and
   the provenance of gate figures from ~/w7/notes/, a directory that exists on one machine and
   is under no version control. A record cited hundreds of times that lives only locally is not
   a record, so it moves in whole, with its directory structure intact.

   Only .md moves. Scripts, run logs, census and gate result JSON, patches, APKs and nm dumps
   stay where they are: they are evidence DATA rather than documents - large, regenerated per
   run, and mostly carrying absolute paths and device serials that mean something only on the
   machine that produced them. The mapping and this split are written in
   docs/Disaggregated/notes/README.md.

   Two exceptions are recorded there too: P6's BRIEF has been promoted to a first-class
   document and keeps no copy under notes/, so the two cannot drift apart; and recovered/ holds
   material salvaged from failed workflow runs, which may duplicate the like-named phase
   directories and is kept because a few of those were the only complete version at the time.

   CI's citation lint takes docs/Disaggregated/*.md only (non-recursive), and notes/ is
   deliberately outside it: those notes' file:line citations were written against their own
   code heads, and resolving them against today's HEAD would fail wholesale - which is not an
   error, it is their correct state as a historical record. To check one, use the revision it
   declares (--rev).

3. Inconsistencies fixed in passing

   - README.md's status line still said P5e was in progress while the stage table already said
     closed.
   - CURRENT_STAGE_PROGRESS's next-step item 1 still listed the exit gate as pending, though
     (a) and (b) were run in the previous commit; rewritten to closed plus the one item that
     actually remains (E1).
   - ROADMAP's P6 row still had the pre-phase plan in three of its cells; it now carries the
     package order, the scope boundary and the two new gates.

Checked with check_doc_citations.py: 125 citations across docs/Disaggregated/*.md, 0 problems.
2026-09-18 23:31:39 -04:00

27 KiB
Raw Blame History

P3a package D (gates) — adversarial review v1

Reviewed 1ef91dec / 588d25e9 / 80f6bf1b on p3a/gates (worktree /home/swung/w7/p3a-gates, parent 39722687), against BRIEF-P3A.md A/B/C.3/D.2/D.3/E, INTEGRATOR-DECISIONS.md ID-1/ID-2/ID-8 and the newly added ID-9, contract-v1.md + contract-review-v1.md (item 11), wire-v1.md, espryt-v1.md, and the package's own gates-v1.md + handlerecycle-before.log.

Nothing in p3a-gates was modified. A private worktree /home/swung/w7/p3a-review-gates was created at 80f6bf1b, used to run both scripts, and removed (git worktree list clean; p3a-gates git status empty).

Verdict: REWORK — four majors, all mechanical (three edits and one integrator decision)

The package is well built and the two scripts are genuinely honest instruments: I ran them and they answer correctly on every path I could reach, including against the real P3a-rewritten Managers.cpp. Nothing here passes on a broken tree and no test name was removed or renamed. What blocks acceptance is narrower: an explicitly assigned open review item is two-thirds undone while the report says it is done (M1), eight new CI entries whose whole purpose is the P3a A/B are never selected by any push-build CI run (M2), a script that breaks its own written exit contract on the one path that matters (M3), and a G5 set that omits the function the brief's risk table names as G5-protected (M4).


1. Majors

M1 — contract-review item 11 is one-third addressed; the four CSO stats lanes still pin 0x7f, and the report claims the item closed

Evidence. contract-review-v1.md:344-347 reads: "The .Handles and stats itest lanes still pin MOBILEGL_PIPE_PUSH=0x7f (MG_IntegrationTest/CMakeLists.txt:863, 965, 975) … Package D owns that file: G12's A/B needs an explicit 0x1ff arm (or the pins re-read as the control they now are)."

At the contract commit those three line numbers are:

  • :863 → MGL_ITEST_HANDLES_ARM_KNOBS — fixed, now 0x1ff at MG_IntegrationTest/CMakeLists.txt:930.
  • :965 → MGL_ITEST_GLES_CSO_ON_ENVIRONMENT — unchanged, still 0x7f at HEAD :1032.
  • :975 → MGL_ITEST_VULKAN_CSO_ON_ENVIRONMENT — unchanged, still 0x7f at HEAD :1042.
  • and their Off twins at :1037, :1047 still 0x800000000000007f.

gates-v1.md §1 and D4 describe the change as the MGL_ITEST_HANDLES_ARM_KNOBS pin only and never mention the CSO lanes; the report presents item 11 as handled.

Failure scenario. After P3a lands, DirectGLES/DirectVulkan.CsoContentAddressing.{On,Off} — the P2 control that proves CSO content addressing is real — runs with bits 7 and 8 cleared, i.e. on a configuration nothing ships. It keeps passing. This is verbatim the objection the package itself writes at CMakeLists.txt:934-941: "Pinning it at 0x7f after P3a would leave the Handles arm asserting the P2 shape while the … handles it is supposed to be about stayed switched off — a lane that still passes and no longer measures the key that ships." The rule was applied to one lane family and not to the two the review named.

Refutation attempted. Could the CSO lanes need 0x7f deliberately? The counters they read (csom, csob) are render-state, untouched by bits 7|8, so raising the mask is behaviour-preserving for what they assert — and D's own stated principle ("the mask is the PHASE default, not a hand-picked bit") points the other way. I found no comment anywhere in the file or the report defending 0x7f for these lanes. If there is a reason, it is unwritten, which is itself the finding.

Fix. 0x7f→0x1ff at :1032, 1042 and 0x800000000000007f→0x80000000000001ff at :1037, 1047, re-run the four CSO lanes; or write the decision down in the file and in gates-v1.md §D4 and say item 11 is closed by argument rather than by change.


M2 — the ResourceSubsystemOn. / ResourceSubsystemOff. lanes (8 entries) never execute in a push build in CI, so the G12/D-A3 A/B cannot go red where it was installed

Evidence. .github/workflows/test.yml:600-602 (as edited by this package):

ctest --output-on-failure -L integration-gpu \
  -R 'HandleRecycle|CsoContentAddressing|ResourceSubsystemControl|MapPersistentRoundtrips' \
  --no-tests=error -j 4

None of the four alternatives matches the new lanes: DirectGLES.ResourceSubsystemOn.LargeArenaAdoptionScenario.SubDataAfterAnInFlightDrawReachesTheNextDraw does not contain ResourceSubsystemControl, and …ResourceSubsystemOn.LargeArenaAdoptionScenario.AnAdoptionCostsExactlyOneMapPersistentRoundtrip does not contain MapPersistentRoundtrip**s** (the case name is singular). The only place the full label runs is the pull build-linux job (test.yml:293-298, ctest -L integration-gpu), where MOBILEGL_PIPE_PUSH steers nothing at all (Config.h declares the field inside the push guard) — so both arms are the same legacy path there. integration-verify is the only job that unpacks a push build (test.yml:573), and it runs only the filtered step above.

Net: the eight entries D added — the four ResourceSubsystemOn. and four ResourceSubsystemOff. LargeArenaAdoptionScenario entries, whose stated purpose (CMakeLists.txt:1176-1182) is "if the handle path and the legacy BufferBackendOps path disagree about any of it, one of these two lanes goes red and names which" — are, in CI, either a duplicate legacy run or not run.

Failure scenario. Espryt's handle arm and its legacy BufferBackendOps arm disagree about an adopted store's in-flight SubData / readback / GPU-write; the lane that exists to name that disagreement never runs with the bits set; CI is green; the divergence ships and is found on device.

Refutation attempted. gates-v1.md §8 item 5 does say the label-wide A/B is the integrator's. That is true of MOBILEGL_PIPE_PUSH=0x7f ctest -L integration-gpu, but D also chose to add these eight entries to CI's push build and then did not select them there; ROADMAP.md:7 asks a gate to be able to go red where it is installed. The pull job does run the entries — but with the mask inert, which is the "green that asserts nothing" the package's own comments reject elsewhere.

Fix. One string: -R 'HandleRecycle|CsoContentAddressing|ResourceSubsystem|MapPersistentRoundtrip' (drop Control, drop the plural). Confirm afterwards that all 16 new push-build entries appear in the step's ctest listing.


M3 — p3a_vertex_input_negative_control.sh breaks its documented exit-1 contract: the "did not trip" path leaves the build directory holding the library built from the patched header

Evidence. Header contract, scripts/p3a_vertex_input_negative_control.sh:40-44: "1 the control did not answer: either the suite stayed green with the field dropped, or it went red without ever naming IsBgra … and both leave the tree restored AND rebuilt". The second of those two branches honours it (:186-197 restore → rebuild → re-ctest → exit 1 at :204). The first does not:

161  if ctest --test-dir "$BUILD_DIR" -R "$TEST_NAME" … ; then
163    say "NEGATIVE CONTROL DID NOT TRIP: …"
166    cp -f "$LOG_DIR/ctest-after.log" ./p3a-vertex-input-control-failure.log
168    exit 1
169  fi

exit 1 runs only the trap 'restore' EXIT installed at :126, which copies the header back and does nothing else. The script's own comment at :171-174 states the rule it is violating: "leaving a build directory holding the broken header is worse than any exit status."

Separately, both exit-1 branches cp a log into $REPO_ROOT (:166, :200). .gitignore carries no rule matching p3a-*.log, so a run leaves untracked files in the working tree.

Failure scenario. G6 has stopped checking (exactly the condition this control exists to detect). An engineer sees "NEGATIVE CONTROL DID NOT TRIP", opens the suite, and re-runs ctest --test-dir build-push — against a libMobileGL.so in which IsBgra is hard-zeroed. Every downstream reading from that build directory until someone happens to rebuild is taken from a deliberately corrupted library.

Refutation attempted. restore() uses cp -f, which refreshes the header's mtime, so the next cmake --build does repair it. But nothing in the script, its message, or its exit code tells the caller to build, and ctest alone will not. The sibling branch proves the author considered the requirement; this branch simply misses it.

Fix. Hoist the restore → rebuild → re-ctest block (:186-197) above the verdict so all three outcomes share it, and keep the failure logs in $LOG_DIR (or a gitignored path) rather than the repo root.


M4 — G5's set omits FlushPendingRangesNow, which the brief twice calls G5-protected and names as a named risk's only mitigation

Evidence. scripts/p3a_untouched_regions.sh:58 implements D-F's nine-function table verbatim. But BRIEF-P3A.md:420 says "FlushPendingRangesNow's three-tier drain (:1004-1072) is a G5-protected function", and E's risk row at BRIEF-P3A.md:1708 makes it the whole mitigation: "FlushPendingRangesNow's three-tier drain silently changes tier … → The function is in G5's byte-identical set and its kill switch … gets its own integration pass in D.3." It is not in the set. Managers.cpp:1004 has exactly one definition (calls at :1401, :1710), so the extractor would take it as a tenth entry unchanged.

Failure scenario. During the P3a rewrite the tier-1 threshold or the whole-buffer test in FlushPendingRangesNow drifts and drops the hot path into tier 3. No gate sees it — G5 does not cover the function, and the brief's risk table records the risk as mitigated by a gate that does not exist. The symptom is the Mali WAR stall and MC 26.3's p99, i.e. exactly ROADMAP.md:19's headline number, and under ID-6 performance is recorded rather than gated, so nothing else stops it either.

Refutation attempted. D-F's authoritative "Decision: nine functions" list omits it, and package D implemented that list exactly — so this is first a contradiction inside the brief. But D owns G5, the contradiction is in the same document D was implementing from, and gates-v1.md does not note it. Note also that P3a may legitimately need to edit the function (D-C reshapes resource_flush_range), in which case adding it would force a red — which is precisely why this needs a written decision rather than silence.

Fix (integrator decision). Either add FlushPendingRangesNow to FUNCTIONS (ten bodies; re-capture the D.0 baseline) and update the script's header table, or record in gates-v1.md/D.5 that E's risk row's mitigation is withdrawn and name what replaces it.


2. Minors

  • m1 — the kept D.2 artefact does not contain the verdict lines the report quotes. handlerecycle-before.log is 153 lines of --output-on-failure (the brief's D.2 command) and contains zero arm= lines: gtest's std::cout verdict from HandleRecycleScenario.cpp:288-291 surfaces only for failing tests. gates-v1.md §5 prints a block of [ HandleRecycle ] arm=AbaControl expected=STALE observed=STALE … lines under "That is the artefact ROADMAP.md's 重键前红 asks for", annotated "(ctest -V)", and that -V log was not kept. The artefact that goes into MEASUREMENTS.md therefore proves the AbaControl arm only by its "Passed" — the inference-from-exit-code the design says it wants to avoid. Fix: keep the ctest -V log beside it and cite it in §5.

  • m2 — gates-v1.md D6 says "Both counting workloads were rewritten to define a SECOND arena". StorageBufferRegrowScenario.cpp:266-272 re-specifies SetUp's m_buffer three times (it is an SSBO with no VAO attributes, so d7655247's crash does not apply). Only LargeArenaAdoptionScenario.cpp:150-158 and ResourceSubsystemControlScenario.cpp:222-225 use a second arena. Report text only.

  • m3 — the D6 workaround is now stale, and it costs the coverage ID-9 asks for. ID-9 (added since gates-v1.md was written) merged dev@9eae9858 — d7655247 included — into feat/disaggregated as 5cb826b0, and requires the fix to be present in both the legacy and the handle arm of the respecify/retire path. After gates rebases, the two workloads that route around the crash stop exercising respecify-of-an-adopted-store-under-a-live-VAO, which is now the path that most needs arm-parity coverage. I verified the merge does not disturb G5: p3a_untouched_regions.sh 44c2b5cf 5cb826b0 → rc 0 (d7655247's edit is at Managers.cpp:1081-1084, inside Ops_Respecify, outside the nine). Fix after the rebase: restore a re-specify window in at least one of the two counting cases; d7655247's own four Respecified*Arena* cases arrive by union anyway.

  • m4 — the report's G14 numbers are against the superseded baseline. Measured now: build-linux/build-push = 2521; ~/w7/p3a-before-ctest-names.txt = 2502 (re-captured at 5cb826b0). D's own contribution is exactly the claimed +34 and every one is a new name — no rename, no removal. The 15 names that comm -23 now reports as "removed" are the un-rebased dev additions ({DirectGLES,DirectVulkan}.LargeArenaAdoptionScenario.Respecified{Vertex,Index}Arena* and 11 PersistentBufferOrderingProbeTest.*). G14 must be re-measured after the rebase, and the union must keep all 15.

  • m5 — p3a_untouched_regions.sh's masker mishandles an odd number of ' in code. A C++ digit separator (16'777'216) opens a char literal at :115 and blanks to the next ', which can hide a brace and shift a body's extent silently. There is none in Managers.cpp today (verified: no [0-9]'[0-9]; every one of its 272 apostrophes is inside a comment or a balanced literal, and comments/strings are masked first), and the extractor survives the real P3a rewrite (below). Cheap fix: skip ' when the preceding character is a digit.

  • m6 — perturb() locates the body brace in the ORIGINAL text (:192, text.index('{', begin)) rather than the masked text, so a brace in a comment or a default argument on the signature line would perturb the wrong place. SELF_TEST_FUNCTION=ClearBufferPool is chosen to avoid this and it works; the failure mode would be a loud self-test failure, not a silent pass.

  • m7 — the hash starts at the line containing name( (:164), so a return type on its own preceding line falls outside the hashed body. None of the nine is written that way today.

  • m8 — C.3's cmake --build build-bench && ctest -L benchmark was not run (D11: no build-bench in the worktree). No benchmark source changed, but it is a brief-mandated verification step; the integrator should run it once on the finished tree.

  • m9 — commit 588d25e9's message says nothing about the RESOURCE_LOCK fix (D5) it carries for four pre-existing lanes. A future bisect for that flake will not find the fix by message. The brief prescribed the three messages verbatim, so this is a note rather than a rule break.

  • m10 — the buffer ABA has no positive control on the backend P3a actually re-keys. The .AbaControl arm is DirectVulkan-only (the knob has no DirectGLES consumer), so on Espryt — the buffer path P3a lands — the corruption is never demonstrated reproducible; only the .Handles arm asserting FRESH stands between a broken re-key and green. The gate can still go red for its reason (a broken re-key reads STALE), and HandleRecycleScenario.cpp:319-321 names exactly where a Features.PipeHandleAbaControl consumer over the resource slot table would go. Worth carrying into package C's scope rather than losing.


3. What I checked and found correct

Recorded because several of these were the likeliest places for this package to be wrong.

  • The G8 "red before" is genuine, and the report's D.2 claims match the log exactly. ExpectPixelsFor (HandleRecycleScenario.cpp:287-299) sets expectsStale only on Arm::AbaControl and then asserts the whole viewport is the stale colour, so a tree on which the corruption stopped reproducing turns the control red, not green. In handlerecycle-before.log: .Legacy Passed on both backends (#2479, #2485), .AbaControl and .AbaControlHandles Passed (#2491, #2497), .Handles Skipped (#2467 DirectGLES, #2473 DirectVulkan) — exactly as gates-v1.md §2 states.
  • The two-buffer window is a real D-H probe. Shared POSITION buffer, differing COLOUR buffer, elements blob byte-identical by construction, so an inherited set_vertex_buffers shows as own-geometry-in-the- dead-VAO's-colour. It cannot distinguish which binding was inherited, and does not claim to. Correctly gated on the arm marker only (P2's vertex-input key), not on the new resource marker. D1's split into its own case is justified by a measurement recorded at the site and adds a name without removing one.
  • mgl_itest_join_environment usage is right on all six new lanes — it is a function with PARENT_SCOPE joining on \\; (CMakeLists.txt:312-322), and every new lane appends ${MGL_ITEST_CAPABILITY_ENV} ${MGL_ITEST_COMMON_ENV} (or ${MGL_ITEST_VULKAN_ENV}), so the EGL/ICD pinning survives the property's REPLACE semantics.
  • mgl_itest_probe_for_symbol clears its output variable (:396, set(mglItestProbeHit "")), so the new foreach(DirectGLES DirectVulkan) at :461 cannot leak the first iteration's hit into the second. Both new probes answer correctly here (no MGPipeResourceOps under MG_Backend/, no MapPersistentRoundtrips under MG_Impl/Pipe/) and will fire on the real trees: p3a-espryt/MobileGL/MG_Backend/DirectGLES/Managers.cpp:2177 declares const MG_Pipe::MGPipeResourceOps g_glesResourceOps, and p3a-client/MobileGL/MG_Impl/Pipe/PipeFill.cpp names MapPersistentRoundtrips.
  • No PipeStatsWindow timing flake. Log.cpp:84 fflushes every line, and MOBILEGL_PIPE_STATS_PERIOD is frames-per-line with a floor of 1 (ConfigLoader.cpp:279), so the summary line is on disk before the assertion reads it. Last() takes the last MGPipe stats: marker, so trailing log traffic is harmless.
  • mpr= cannot be "absent" on a push build, so the Off arm's mpr == 0 is a reading and not a missing-field failure: PipeStats.cpp:431 appends " mpr=" + to_string(...) unconditionally inside the #if MOBILEGL_PIPE_PUSH cso[ bracket, and CounterOrAbsent's " " prefix matches it. This was my main suspected false-red on G12 and it is refuted.
  • Private-log discipline is complete. Every new log-reading entry owns a MOBILEGL_LOG_FILE_PATH nothing else writes and selects exactly one case (ResourceSubsystemControlScenario.cpp has exactly one TEST_F; the two MapPersistentRoundtrips. registrations each name one case). The ResourceSubsystemOn/ Off. lanes configure no log and the counting case skips saying so. MGL_ITEST_COMMON_ENV (:279-287) carries no log path. I enumerated all 20 log-owning registrations: the four D5 locked (UnlocatedIoBlocks, PrimGenReroute, both PointSizeDemotion) are exactly the ones that give a whole scenario one log and read it back; the CSO, verify-arming, verify-corrupt and poison lanes each filter a single case, and the Verify. lanes have no reader. D5 is correct and complete.
  • G12's direction is sound in both dead-switch modes: bits stuck ON → the Off arm reads 2 and goes red; bits stuck OFF → the On arm reads 0 and goes red; and both arms additionally assert the same green viewport, so a switch that changed what is drawn also fails.
  • G10's numbers are far enough apart to be unambiguous: 3 definitions vs 12 dispatches, 1 adoption vs 5 draws, with 0 separately named in both messages.
  • The G7 script's hard-coded assumptions hold against package B's real header. p3a-client/MobileGL/MG_Impl/Pipe/VertexInputEmit.h:88 is wire.IsBgra = attrib.IsBgra ? 1 : 0;; the script's regex matches it exactly once (dry-run confirmed) and rewrites it to wire.IsBgra = 0 /* … */;; VertexInputEmitTest is registered at MG_Test/Pipe/CMakeLists.txt:147,171. Missing header → 2, header without the field → 2, bad build dir → 2 (I ran the last one).
  • Exit codes verified by running (in the private worktree): --self-test → 0 with all three controls printing; 44c2b5cf HEAD → 0; single-ref → 0 with 9 lines; nonexistent ref → 2; three arguments → 2. Worktree git status clean afterwards — neither script leaves the tree modified.
  • G5 survives the real P3a rewrite. Run against p3a-espryt@203120fa — which rewrites Managers.cpp for the whole handle conversion and adds the resource op table at :2177 — the extractor finds all nine and reports byte-identical vs 44c2b5cf. This is the strongest available evidence that the brace-matching extractor neither false-reds nor fails to extract after the file is rewritten. Extraction is also correct in the tricky cases: RingAllocate's forward declaration (:958) and its call sites are rejected on the )-followed-by-{ test, and RingAllocateSlow / UboRingAllocate / UnpackRingAllocate are excluded by the \b anchor.
  • G5 will not false-red after the ID-9 merge: 44c2b5cf → 5cb826b0 → rc 0.
  • test.yml. fetch-depth: 0 and job-level BASELINE: "44c2b5cf" are both required and correctly scoped (and correctly not the workflow's baseline_sha symbol input). The workflow has no pull_request trigger, so the two G5 steps' if: runs them on every feat/disaggregated push and skips them on dev — as intended. include-graph-check untouched, the TEMPORARY trigger lines untouched, no other job changed (the whole 61-line test.yml diff is confined to the two regions).
  • Commit hygiene. Three commits, three single-line messages verbatim from C.3, no bodies, no trailers; files strictly inside C.5's package-D set; wsl_p3a_gate.sh / p3a_ab.sh correctly not written (ID-3 supersedes C.3's table). Both scripts are mode 644, matching scripts/g7_negative_control.sh.

4. Judgement on the deviations and the two carried findings

# judgement
D1 (new case, not a second window) Accept. The reason is measured, recorded at the site, and adds a name rather than removing one (G14). The alternative genuinely does not work: one ABA memo entry per process carries the resolved layout as well as the bindings.
D2 (a new resource marker, not P2's) Accept, and it is the right call. Reading MGITEST_HANDLE_REKEY_<backend> would have armed the buffer arm on a tree where no buffer is keyed on a handle. Verified the new probe is absent here and present in p3a-espryt.
D3 (AbaControl expectation follows the measurement) Accept. The reasoning is at the assertion, and the log backs it (#2491, #2497 Passed).
D4 (0x7f→0x1ff) Accept as far as it goes — see M1. The Handles pin is right; the two CSO pins the same review item named were not touched and the report does not say so.
D5 (RESOURCE_LOCK on four flaky lanes) Accept — this is a good fix. The diagnosis (whole-scenario log + fopen(path,"w") + ctest -j) matches the file's own documented rule, RESOURCE_LOCK avoids the rename G14 forbids, the four locked lanes are exactly the ones with that shape (I enumerated all 20 log-owning registrations), and three consecutive 950/950 runs is adequate evidence. Only m9 (message) against it.
D6 (dev@d7655247 routed around) Superseded, and now a small liability. ID-9 merged the fix as 5cb826b0 before gates rebases, so item 8 of §8 is already discharged. The workaround is not wrong (same number of storage definitions), but it now removes the only coverage of respecify-under-a-live-VAO at exactly the moment ID-9 requires that path to work in both arms. See m3.
D7 (branch-scoped G5 rows, BASELINE, fetch-depth: 0) Accept. Correct, necessary, and correctly distinguished from baseline_sha.
D8 (single-ref listing mode) Accept. Makes D.0's capture a plain redirect; verified rc 0 / 9 lines.
D9 (new PipeStatsWindow.h) Accept. Shared reader, CsoContentAddressingScenario's copy left alone — right call for a package with no other reason to touch that file.
D10 (mode 644) Accept. Matches the nearest precedent.
D11 (build-bench not built) Accept with m8. No benchmark source changed; the integrator should run it once.

The two "findings the integrator must carry":

  • the d7655247 SIGSEGV — real, correctly diagnosed, correctly not fixed in flight per ROADMAP.md:88, and already resolved by ID-9's merge. What must now be carried is the inverse: undo the detour so the path is covered again (m3).
  • the RESOURCE_LOCK fix for four flaky pinned lanes — real, correctly fixed here rather than carried, and verified as the only lanes with that shape. Nothing left for the integrator except to notice it is in 588d25e9 (m9).

5. Rework list (exact)

  1. M1 — MG_IntegrationTest/CMakeLists.txt:1032, 1042: 0x7f → 0x1ff; :1037, 1047: 0x800000000000007f → 0x80000000000001ff. Re-run the four CsoContentAddressing.{On,Off} lanes. Or write the reason they stay at the P2 mask into the file and into gates-v1.md D4, and say item 11 is closed by argument. Either way, gates-v1.md must stop implying item 11 was a one-line change.
  2. M2 — .github/workflows/test.yml:601: -R 'HandleRecycle|CsoContentAddressing|ResourceSubsystem|MapPersistentRoundtrip'. Verify the step now lists all 16 new push-build entries.
  3. M3 — scripts/p3a_vertex_input_negative_control.sh: move the restore → rebuild → re-ctest block (:186-197) above the verdict so the :161-169 branch shares it; write the two failure logs into $LOG_DIR instead of the repo root, or add a .gitignore rule.
  4. M4 — integrator decision on FlushPendingRangesNow: add it to FUNCTIONS (:58) and to the header table, re-capture the D.0 baseline; or record the withdrawal of E's risk-row mitigation in D.5.
  5. m1 — keep the ctest -V HandleRecycle log beside handlerecycle-before.log and cite it in gates-v1.md §5; m2, m4 — correct the D6 and G14 wording.
  6. Optional (m5, m6): digit-separator guard in mask(); brace lookup in perturb() over the masked text.

6. What the integrator must re-run on the finished tree

gates-v1.md §8's list is correct and I have nothing to remove from it. Add:

  1. Re-measure G14 against the current baseline. ~/w7/p3a-before-ctest-names.txt is now 5cb826b0's (2502). After the rebase, comm -23 must be empty and the union must contain all 15 dev-side names (LargeArenaAdoptionScenario.Respecified{Vertex,Index}Arena* × 2 backends, 11 PersistentBufferOrderingProbeTest.*) plus D's 34. ID-9's union resolution in LargeArenaAdoptionScenario.cpp is where a name can be lost.
  2. Re-run p3a_untouched_regions.sh 44c2b5cf HEAD after the espryt merge specifically. I already verified it is rc 0 against p3a-espryt@203120fa and against 5cb826b0; the merge resolution is the place a pool or ring body can still move.
  3. After M2's fix, confirm the ResourceSubsystemOn/Off lanes actually assert — i.e. that they do not all skip — in the integration-verify job's controls step.
  4. cmake --build build-bench && ctest --test-dir build-bench -L benchmark once (C.3's block, m8).
  5. Re-check that the ResourceSubsystemControl.Off arm reads mpr == 0 and that the On arm reads a non-zero mpr on a tree carrying B and C. A green Off arm with a skipping On arm is not the A/B.