Files
MobileGL/docs/archive/notes/p3a/p3a-results/gates-v2.md
T

26 KiB
Raw Blame History

P3a package D — the gates, rework round 1 (gates-v2)

Branch p3a/gates, worktree /home/swung/w7/p3a-gates, rebased onto refs/heads/feat/disaggregated (5cb826b0) per ID-9. HEAD 2bafb316. Not pushed.

Answering gates-review-v1.md's REWORK (four majors) plus the integrator's ID-9 / ID-10 / ID-11. Everything in the review's §5 rework list is addressed; the optional minors (m5, m6) are done too.

# sha message
— 5cb826b0 (base: contract e01c0ccc + the dev@9eae9858 merge)
1 9e08c563 [Test] (Pipe): reproduce the buffer handle ABA through public GL and prove the pre-rekey guards are what stops it
2 ecbddcea [Test] (Pipe): assert a regrown adopted store costs one map-persistent round trip per storage definition, not one per draw
3 20dfb64c [CI] (Pipe): gate that the buffer pool, the deferred-release drain and the three rings did not move
4 331bb425 [Test] (Pipe): count the map-persistent round trips over a re-specified adopted arena in both arms, now that the VAO rebind fix is in the branch
5 2dfac81a [Fix] (Pipe): raise the four CSO stats lanes to the P3a subsystem mask so they stop asserting on a configuration nothing ships
6 854b1be3 [CI] (Pipe): select the P3a subsystem and map-persistent lanes in the only job that unpacks a push build
7 a9d4011a [Fix] (Pipe): restore and rebuild on every exit path of the vertex-input negative control, and keep its logs out of the repository
8 2bafb316 [CI] (Pipe): add FlushPendingRangesNow's three-tier drain to G5's byte-identical set

Commits 1–3 are the original three, rebased (the duplicate contract patch dropped out automatically). Commits 4–8 are the rework. Diffstat 20dfb64c..HEAD: 6 files, +294 / −156.


1. The rebase (ID-9)

git rebase refs/heads/feat/disaggregated — rc 0, no conflict. Git merged LargeArenaAdoptionScenario.cpp cleanly: dev's two cases were appended in one hunk and gates' case in another, far enough apart that the union is what git produced. It was verified by name rather than assumed, because ID-9 names this file as where a name can be lost:

side TEST_F(LargeArenaAdoptionScenario, …)
gates pre-rebase 80f6bf1b SubDataAfterAnInFlightDrawReachesTheNextDraw, ReadbackSeesTheLatestCpuWrite, GpuWriteIntoTheArenaIsReadBack, AnAdoptionCostsExactlyOneMapPersistentRoundtrip (4)
dev side 5cb826b0 SubDataAfterAnInFlightDrawReachesTheNextDraw, RespecifiedVertexArenaKeepsVaoBindings, RespecifiedIndexArenaKeepsVaoBinding, ReadbackSeesTheLatestCpuWrite, GpuWriteIntoTheArenaIsReadBack (5)
HEAD all 6 — the union, nothing renamed, nothing dropped

The 11 PersistentBufferOrderingProbeTest.* names from 9eae9858 are in a file this package does not touch and arrived untouched. G14 below confirms both sets at the ctest level.

feat/disaggregated has moved on since this rebase. It was 5cb826b0 when this package rebased and when every number below was taken; by the end of this round ~/w7/pipe reads 12e6bfcf (package A's wire landed; p3a/client has also moved to b6706362). Under ID-2 gates integrates last, so one more rebase is expected and is the integrator's; nothing in this package touches MG_Pipe/**, MG_Impl/** or MG_Backend/**, so the only file with any chance of conflicting again is MG_IntegrationTest/CMakeLists.txt. G1 and G14 must be re-measured against whatever pull baseline the integrator holds at that point — the ~/w7/p3a-before-* captures used here are 5cb826b0's.


2. Per-finding disposition

M1 — the four CSO stats lanes still pinned 0x7f — FIXED (2dfac81a)

MG_IntegrationTest/CMakeLists.txt, at the review's line numbers, now at HEAD's:

lane was now line at HEAD
MGL_ITEST_GLES_CSO_ON_ENVIRONMENT (:1044) 0x7f 0x1ff :1046
MGL_ITEST_GLES_CSO_OFF_ENVIRONMENT 0x800000000000007f 0x80000000000001ff :1051
MGL_ITEST_VULKAN_CSO_ON_ENVIRONMENT 0x7f 0x1ff :1056
MGL_ITEST_VULKAN_CSO_OFF_ENVIRONMENT 0x800000000000007f 0x80000000000001ff :1061

The control semantics are unchanged: the only bit that separates the two arms is 63 (kMGPipeBehaviourNoCsoContentAddressing, MG_Pipe/MGPipe.h:90); every other bit is the shipping mask and now moves with the phase (kMGPipeSubsystemsMigratedAtP3a = 0x1ff, MGPipe.h:95). A comment block at :1031-1043 states that rule at the site, so the next phase's author does not have to rediscover it, and closes contract-review item 11 explicitly for all three lane families.

Re-run (M1's own requirement). ctest --test-dir build-push -R CsoContentAddressing → rc 0, 6 entries (the 4 lanes pass; the 2 ambient entries skip by design). The arms are still distinguishable at the new mask — the lane logs read csom=3 csob=17 (content-addressed) against csom=17 csob=17 (no content addressing), i.e. 3 mints for 17 binds versus one mint per bind. The generated ctest environment was checked directly: the CSO On lanes now carry MOBILEGL_PIPE_PUSH=0x1ff and no 0x800000000000007f string survives anywhere in the build tree.

Every remaining pin in the file re-read (grep -n 'MOBILEGL_PIPE_PUSH=0' MG_IntegrationTest/CMakeLists.txt), which is the second half of the instruction. Thirteen value-carrying pins remain:

line value verdict
:930 MGL_ITEST_HANDLES_ARM_KNOBS 0x1ff phase default
:944, :947 HandleRecycle.Legacy (both backends) 0 named control — the pre-handle arm, documented at :895-900
:950 HandleRecycle.AbaControl 0 named control — D18's lane verbatim
:1046, :1056 CSO On 0x1ff fixed here
:1051, :1061 CSO Off 0x80000000000001ff fixed here; bit 63 is the control
:1130 ResourceSubsystemControl.On 0x1ff phase default
:1135 ResourceSubsystemControl.Off 0x7f named control — G12's armed/cleared A/B
:1164, :1169 MapPersistentRoundtrips. ×2 0x1ff phase default
:1200 ResourceSubsystemOn. 0x1ff phase default
:1203 ResourceSubsystemOff. 0x7f named control

HandleRecycle.AbaControlHandles (:953-956) pins nothing of its own — it composes MGL_ITEST_HANDLES_ARM_KNOBS, so it followed 0x1ff automatically. No lane runs with P3a's subsystems off unless it is a named control.

M2 — the eight ResourceSubsystemOn./Off. entries never ran in a push build in CI — FIXED (854b1be3)

.github/workflows/test.yml:615, the integration-verify job's controls step:

-R 'HandleRecycle|CsoContentAddressing|ResourceSubsystem|MapPersistentRoundtrip'

ResourceSubsystemControl → ResourceSubsystem and the plural MapPersistentRoundtrips → singular, for exactly the reasons the review gives. A 14-line comment above the step records that these alternatives are test-name PREFIXES, that each is the shortest string that still selects only what it means to, and that this is the only CI job unpacking a push build — so an entry the filter misses either never runs under the P3a bits or runs only in the pull integration job where the mask steers nothing.

Measured, in the job's own build directory (build-verify, -L integration-gpu -R …):

filter entries selected
old (as reviewed) 80
new 96
of which P3a lane entries (ResourceSubsystem{On,Off,Control}. / MapPersistentRoundtrips.) 16

The step run for real: rc 0, 96/96 pass. Of the 12 ResourceSubsystemOn./Off. entries, 10 assert and 2 skip — the two skips are AnAdoptionCostsExactlyOneMapPersistentRoundtrip in each lane, which needs package B's counter and says so; the other five cases per lane (including dev's two Respecified*Arena*) run on both arms today. That is review item 9 ("confirm the lanes actually assert") discharged, at least for the pre-B tree.

Every other new lane checked the same way. All 38 names D adds are present in ctest -N on build-linux, build-push and build-verify; the 22 that carry ResourceSubsystem or MapPersistentRoundtrip in their names (16 lane entries + 6 ambient) are all inside the new filter. The HandleRecycle families were already reachable and remain so.

M3 — the G7 script's exit contract — FIXED (a9d4011a)

scripts/p3a_vertex_input_negative_control.sh restructured:

  • A single repair() (:103-123): restore the header from the pre-patch byte copy, cmake --build, then re-run the suite to prove the tree went back. Idempotent, a no-op until the patch is applied.
  • PATCH_APPLIED=1 is set before the patcher runs (:178), not after, so a python that died mid-write is repaired too. trap 'repair' EXIT (:174) replaces trap 'restore' EXIT, so mid-way failures repair: the "patched header did not compile" branch (:203) and any interrupt now go through it.
  • Step 4 (from :212) only records a verdict (tripped / did-not-trip / wrong-reason); step 5 (from :227) calls repair and clears the trap before anything is printed or exited, and a failed repair downgrades the verdict to 2 ("a build directory that could not be put back is 'could not run'"). All three outcomes share one block.
  • No file is written into the repository. LOG_DIR is <build-dir>/p3a-g7-logs (absolute, created with mkdir -p, falling back to mktemp -d if the build dir is unwritable); the header backup lives there too, and both former cp … ./p3a-vertex-input-control-*.log lines are gone. Every verdict prints the log path.
  • The header's exit-code block (:39-63) now states the rebuild as part of the contract and explains why the "did not trip" path is the one that needs it most.

Proved by forcing each path once. A rehearsal ran a byte-identical copy of the shipped script (cmp clean) inside a throwaway repo at /tmp/g7rehearsal, whose MobileGL/MG_Impl/Pipe/VertexInputEmit.h and VertexInputEmit.Roundtrip ctest entry are tiny stand-ins — so the control flow exercised is the real one, and only the two constants' contents differ:

forced condition rc header after binary after repo root
the walk names IsBgra when it is zeroed 0 original built from the restored header clean
the walk stopped comparing (stays green) 1 original built from the restored header clean
the walk fails without naming the field 1 original built from the restored header clean
the patched header does not compile (Bgra enum class) 2 original built from the restored header clean

In every case the transcript shows restored …; rebuilding … / the tree is restored, rebuilt and green again before the verdict line (or, for the exit-2 case, from the trap), and ls of the fake repo root shows nothing new. The rehearsal tree was deleted.

On the real tree the script still answers rc 2 — MG_Impl/Pipe/VertexInputEmit.h is package B's and does not exist yet — and leaves git status clean.

M4 / ID-11 — FlushPendingRangesNow in G5 — DONE (2bafb316)

scripts/p3a_untouched_regions.sh: FUNCTIONS is now ten names, appended last so the first nine lines of the D.0 baseline keep their positions. EXPECTED_FUNCTION_COUNT=10 replaces the hard-coded 9. The header table gains the row and a paragraph recording why the tenth is there (ID-11 resolving BRIEF-P3A.md:420 / :1708 against D-F's nine-function table) and what byte-identical means for it: the handle arm must CALL it, not carry a copy — a push arm that re-spells the ladder beside an untouched #else copy satisfies G1 and defeats G5, because two ladders drift.

The self-test grew a control for it. SELF_TEST_FUNCTIONS="ClearBufferPool FlushPendingRangesNow", one negative control each, each run from the pristine copy and each required to be named in the message. ClearBufferPool is the easy shape (small, no forward declaration); FlushPendingRangesNow is the opposite on purpose — 70 lines, three nested tiers, its own early returns — so an entry that silently extracted the wrong extent cannot pass. Both trip:

[p3a-untouched] positive control: 10 bodies extracted from the working tree
[p3a-untouched] positive control: an untouched copy compares equal
[p3a-untouched] positive control: an edit outside the ten bodies is invisible
[p3a-untouched] negative control: a perturbed ClearBufferPool body is reported, and named
[p3a-untouched] negative control: a perturbed FlushPendingRangesNow body is reported, and named
[p3a-untouched] self-test passed

The new D.0 baseline (p3a_untouched_regions.sh HEAD, rc 0):

ba79b92dc6ebf2cb2af6aa79f23ba1941818d131812965e9f96a89caffc8e90a  IsPoolable
6563286e007650a5abb0a1bbdd7acddd901776f3e55d3c0087c8fee60aa6cc4c  EnrollIntoPool
dc9d3066b2429208e18adc27c639d16260e74b7f17f7c3f290d54f58534f089b  AcquireFromPool
12656cfd716dec50a4a4893459a7a6aaa0a635f79dda8fa51b88a819da27e5e0  TrimBufferPool
c44a12746bcd82e55d8a635adf766e1fe2cf961ded5d9d745896fd65535bde16  ClearBufferPool
15259e547fb514dcdcf087bbde1d048fc2d9e7901ea6c6c94a1109fb7987073a  ProcessDeferredBufferReleases
2daaa5254c38db63742ef8d6dacd25f8868c7fcd3d9745956862134cf76911d5  CreateRingStorage
6221df2e4c0b185fbac061e956253dacd8f8d886bb3f3341cd76a26b5e1f84b1  RingAvailable
709b6f4190c735517034b9c5d62967a9ae61f2198d2eb441dfe4e7af437cb784  RingAllocate
c65570022a5856c1afccfd0cb42fefac707689f5bb3154c3a8e17d2495ccdf8e  FlushPendingRangesNow

The integrator should replace ~/w7/p3a-before-untouched.sha with this list (it is 44c2b5cf's too — the gate is rc 0 between them).

D6 — the second-arena workaround undone (331bb425)

ID-9 makes d7655247 an ancestor of this branch, so both counting workloads now re-specify an adopted store under a live VAO instead of routing around it:

  • LargeArenaAdoptionScenario.cpp:457-474 — glBindBuffer(m_arena); glBufferData(kArenaBytes, nullptr, …) inside the counted window, and the attribute pointers are deliberately not re-declared afterwards, so the draws below can only land if the backend VAO followed the new store on its own. m_secondArena and its TearDown lines are gone.
  • ResourceSubsystemControlScenario.cpp — DefineAnArenaAndDrawFromIt → DefineTheArenaAndDrawFromIt (:225): index 0 creates the store and declares the attributes, every later index re-specifies the same buffer and declares nothing. std::vector<GLuint> m_arenas → a single GLuint m_arena (:250). Two storage definitions either way, so the counter's expectations (mpr == 2 on, 0 off) are unchanged.

Both comments now say why the shape is the hard one and cite ID-9 rather than the old detour. Because the first of these runs under ResourceSubsystemOn./Off., the respecify path is now exercised on both arms of the P3a A/B, which is what ID-9 asks for; a handle arm that re-implemented the retire without the rebind is a dead draw or a fault there rather than a divergence found on device.

The minors

# disposition
m1 (the -V artefact) Done. Both logs regenerated post-rebase and kept: ~/w7/p3a-handlerecycle-before.log (the D.2 command) and ~/w7/p3a-handlerecycle-before-verbose.log (ctest -V, 158 KB, 22 [ HandleRecycle ] verdict lines), both copied to ~/w7/notes/p3a/p3a-results/. The AbaControl green is now readable as a verdict line, not inferred from an exit code.
m2 (D6 wording) Moot — the workaround is gone. For the record the review was right: StorageBufferRegrowScenario never used a second arena (it re-specifies an SSBO with no VAO attributes).
m3 (the stale workaround) Fixed — see D6 above.
m4 (G14 numbers) Re-measured against 5cb826b0's baseline; see §3. D's contribution is +38 (34 + the 4 new ResourceSubsystem{On,Off}.…Respecified*Arena* lane entries the union produced).
m5 (digit separator) Done. mask() skips a ' preceded by a digit or a hex digit that itself follows a hex digit — a C++14 separator (16'777'216, 0xff'ff) no longer opens a char literal that could blank a brace a function away.
m6 (perturb() brace) Done. The opening brace is located in the masked text and used as an offset into the original (mask() preserves length), so a brace in a comment or a default argument on the signature line cannot misdirect the perturbation.
m7 (return type on its own line) Not changed. Still true that none of the ten is written that way, FlushPendingRangesNow included (void FlushPendingRangesNow( is one line).
m8 (build-bench) Still not run — there is no build-bench in this worktree and no benchmark source changed. Integrator item, unchanged.
m9 (588d25e9's message) Unchanged: the messages are C.3's verbatim and the rebase preserved them (ecbddcea). The D5 RESOURCE_LOCK fix is still carried by that commit; this report is the place it is written down.
m10 (no DirectGLES ABA positive control) Unchanged; carried to package C's scope as the review suggests.

3. The G5 verdict against espryt's tree

git -C /home/swung/w7/p3a-espryt rev-parse HEAD has moved: it is now cdbd5535 ([Test] (Espryt): extend the slot-table suite to the buffer kind and pin that its death crosses as resource_destroy), not the 203120fa the brief names. The gate was run against both, and answers identically.

$ bash scripts/p3a_untouched_regions.sh 44c2b5cf 203120fa      -> rc 2
$ bash scripts/p3a_untouched_regions.sh 44c2b5cf cdbd5535      -> rc 2
[p3a-untouched] FlushPendingRangesNow: expected exactly one definition, found 2
[p3a-untouched]   ...definition at line 1127, 3 lines
[p3a-untouched]   ...definition at line 1307, 70 lines
[p3a-untouched]   the handle arm must CALL the untouched FlushPendingRangesNow, not carry a copy of
[p3a-untouched]   it: two ladders drift (BRIEF-P3A.md:1708, ID-11)

The gate was not weakened. Instead the len(hits) > 1 diagnostic was made to print the two extents and the rule, because this is now the expected shape of the failure rather than a limitation of the extractor.

The exact hunk, for espryt's rework (MG_Backend/DirectGLES/Managers.cpp at both refs; the whole region is #if MOBILEGL_PIPE_PUSH at :902 … #else at :1188 … #endif at :1424):

what where lines sha256 (16)
push arm — FlushPendingRangesFrom(GLESBufferResource&, const Uint8* hostBase, SizeT frontendSize), a re-spelled copy of the whole three-tier ladder :1051-1125 75 37fc94ffc5991923
push arm — FlushPendingRangesNow(GLESBufferResource&, BufferObject&), a thin forwarder to it :1127-1129 3 7ac74243aec7b6f8
#else (pull) arm — FlushPendingRangesNow(GLESBufferResource&, BufferObject&) :1307-1376 70 c65570022a5856c1 = byte-identical to $BASE:1004

New callers of the re-spelled core are at :2000 and :2541 (the handle arm); :1708 and :2677 still call FlushPendingRangesNow.

Read this precisely, because it is not "a body moved":

  1. All nine original G5 functions are byte-identical at espryt's HEAD — verified by extracting them from espryt's file directly; every sha matches 44c2b5cf. The review's earlier rc 0 still holds for the nine.
  2. The pull arm's FlushPendingRangesNow is byte-identical too. Nothing was edited in place.
  3. What the gate objects to is that inside a push build the ladder exists twice — once as FlushPendingRangesFrom (75 lines, re-parameterised) and once as the #else copy that G1 freezes — and the gate cannot say which is "the" body, so it refuses (rc 2) rather than picking one.

This is a real conflict between G1 and G5-with-the-tenth, and it is the integrator's to settle. Espryt's deviation D1 exists because refactoring these helpers in place resized five pull symbols (FlushPendingRangesNow +14 among them), which G1's empty admitted-resize set forbids; the #if/#else split is what bought G1 back. The two obvious resolutions both cost something:

  • Espryt restores it byte-identical and the handle arm calls it. This is what ID-11 asks for and it is what makes the risk row at BRIEF-P3A.md:1708 true. It needs the handle arm to reach the same body without a BufferObject&, which is the constraint D1 was written around — so it is not a one-line change, and whether it is achievable without resizing a pull symbol is espryt's call, not this package's.
  • The tenth entry stays and the phase records a written, argued exception for it — naming the two copies, stating that the #else copy is frozen by G1 and the push copy is covered by CrossFrameBufferScenario's thirteen cases plus StreamedArenaScenario's two, and that the exception retires with the pull path at P13. That is a decision, not a silence, and it is what D.5 would record.

What this package will not do is drop the tenth entry or teach the extractor to prefer one of two definitions: either would turn ROADMAP.md:19's headline p99 number back into a risk with a mitigation that does not exist. The finding is recorded here so espryt's rework and the integrator can choose.


4. Verification transcript

All in /home/swung/w7/p3a-gates, CCACHE_BASEDIR=/home/swung/w7, at 2bafb316.

check result
cmake --build {build-push,build-linux,build-verify} -j 28 rc 0, rc 0, rc 0
G1 symbol_report.py --before ~/w7/p3a-before-libMobileGL.so --after build-linux/libMobileGL.so --threshold 0 --fail-on-symbol-set-change --fail-on-added-bytes 0 rc 0 — Removed 0, Added 0, Resized 0, Renamed 0. 0/0/0/0 against ID-9's re-captured 5cb826b0 baseline
G2 ctest -N (ID-4 extractor), pull vs push identical, 2540 each (build-verify 3382)
G14 comm -23 ~/w7/p3a-before-ctest-names.txt <build-linux names> empty — 0 removed. comm -13 = +38 added, all new names; the union kept dev's LargeArenaAdoptionScenario.Respecified{Vertex,Index}Arena* on both backends and in both new subsystem lanes
ctest -L unit -j 12 on all three builds rc 0 — 1582/1582 each
ctest -L integration-gpu -j 8 on build-push rc 0 — 958/958
ctest -L integration-gpu -j 8 on build-linux rc 0 — 958/958
ctest -L integration-verify -j 4 on build-verify rc 0 — 842/842
hard rule 4 ctest --test-dir build-verify -R HandleRecycle --no-tests=error rc 0; only visible skips (the .Handles buffer arm and the Verify. prefixed copies, each naming what is missing)
hard rule 4 ctest --test-dir build-push -R 'StorageBufferRegrow|LargeArenaAdoption|ResourceSubsystem' rc 0; the only skips are the four counting entries + the two lane copies, all saying no source under MobileGL/MG_Impl/Pipe/ names MapPersistentRoundtrips
M1 re-run ctest --test-dir build-push -R CsoContentAddressing rc 0 — 6 entries, 4 lanes assert, arms distinguishable (csom=3 vs csom=17 at 17 binds)
M2 re-run the CI step as CI runs it, on build-verify rc 0 — 96/96; 80 → 96 entries, the 16 P3a lane entries now included; 10 of the 12 On/Off entries assert, 2 skip naming package B
G5 p3a_untouched_regions.sh 44c2b5cf HEAD rc 0 — ten bodies byte-identical
G5 p3a_untouched_regions.sh --self-test rc 0 — 3 positive controls + 2 negative controls (both named)
G5 vs p3a/espryt (203120fa and cdbd5535) rc 2, naming the two definitions — §3
G7 p3a_vertex_input_negative_control.sh build-push on this tree rc 2 — package B's header absent, message says so
G7 all four exit paths, forced on a byte-identical copy 0 / 1 / 1 / 2, tree restored and rebuilt on every one, nothing written to the repo root
git status clean

Two D.2 artefacts kept (~/w7/ and copied to ~/w7/notes/p3a/p3a-results/): handlerecycle-before.log and handlerecycle-before-verbose.log. Every other log this round created — ~/w7/p3a-gates-build.log, the /tmp build and label logs, the /tmp/g7rehearsal tree — was deleted. ls ~/w7/p3a-gates-* is empty.


5. What remains for the integrator

gates-v1.md §8's list stands, with items 8 (carry D6) and the G14 re-measurement discharged. What is left, in the order it bites:

  1. Settle the FlushPendingRangesNow duplication with package C (§3). This is the one item that is not mechanical: either espryt's handle arm calls the untouched body, or the phase writes down an argued exception in D.5. Until then G5 is rc 2 on any tree carrying espryt, and rc 2 must not be read as "the script is broken" — the message names the two definitions and their line extents.
  2. Re-run p3a_untouched_regions.sh 44c2b5cf HEAD after every merge, plus --self-test. Re-capture ~/w7/p3a-before-untouched.sha with the ten-line listing above (§2, M4).
  3. p3a_vertex_input_negative_control.sh build-push on the tree carrying package B — it must report negative control tripped, naming IsBgra (rc 0). Anything else is a finding about G6, and rc 2 now means something different from what it means here. Its logs land in <build-dir>/p3a-g7-logs/.
  4. The .Handles buffer arm must stop skipping once package C registers MGPipeResourceOps; the four counting entries must stop skipping once package B emits MapPersistentRoundtrips, and then read mpr == 3, 1, 2 (On) and 0 (Off). A green Off arm beside a skipping On arm is not the A/B (review item 11).
  5. MOBILEGL_PIPE_PUSH=0x7f ctest -L integration-gpu -j 8 beside the default 0x1ff — the label-wide third arm. The two ResourceSubsystemOn/Off. lanes now cover the adopted store's whole life including the respecify path in both arms, but the label-wide A/B is still the integrator's.
  6. cmake --build build-bench && ctest --test-dir build-bench -L benchmark once (C.3's block, m8).
  7. gh workflow run test.yml --ref feat/disaggregated — pipe-gates runs the two G5 rows (branch scoped, BASELINE: "44c2b5cf", fetch-depth: 0); the integration-verify controls step now selects 96 entries instead of 80.
  8. The DirectGLES buffer ABA has no positive control (review m10) — carry into package C's scope rather than losing it: HandleRecycleScenario.cpp:319-321 names where a Features.PipeHandleAbaControl consumer over the resource slot table would go.