Files
MobileGL/docs/Disaggregated/notes/p2/p2-results/gates-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

24 KiB
Raw Blame History

P2 package E — p2/gates — result

Tree ~/w7/p2-gates, branch p2/gates, six commits on top of the P2 contract commit 9c6a8a25. Not pushed.

The tree had to be created from the SHA, not from the tag: p2/contract is ambiguous in ~/w7/pipe (a branch of the same name exists), so git worktree add … p2/contract fails with fatal: ambiguous object name and wsl_p2_tree.sh passes the ref straight through. Both refs resolve to 9c6a8a25d88593d802e69a5d4fdee498a5b6816a.

Build directories: build-linux (pull), build-push, build-verify from wsl_p2_tree.sh, plus build-retrace (-DMOBILEGL_BUILD_TRACE_REPLAY=ON) and build-bench (-DMOBILEGL_BUILD_BENCHMARK=ON) added here for this package's own pre-flights.


1. Commits

sha subject
f02ca420 [Test] (Pipe): reproduce the handle ABA through public GL and pin the CSO content-addressing switch
1b25f2c8 [Feat] (Trace): record per-frame thread CPU time beside wall time so a paired A/B can be read as CPU cost
718818f1 [Feat] (Bench, Pipe): run the blend-toggle case in CI, give G7 a negative control, and record the two campaign devices
3766d842 [Test] (Pipe): register the CSO control in the pull build too, so pull and push name the same tests
bf62597c [CI] (Pipe): make the dirty-surface report a gate and run the push-only unit tests on the verify runtime
aef43358 [Test] (Pipe): one case per CSO lane, because two of them would race on the lane's log

16 files, +1867 / −32 (before aef43358, which is a small net deletion).

Ownership (C.5): no violation. Every path touched is E's — .github/workflows/test.yml, MobileGL/MG_Benchmark/**, MobileGL/MG_IntegrationTest/**, android-plugin/app/src/trace/cpp/**, scripts/g7_negative_control.sh, tools/device_bench/**, tools/trace_replay/run_android_retrace_local.py.


2. C.4's table, row by row

row state
CREATE Scenarios/HandleRecycleScenario.cpp (D18) 4 cases, 3 arms, 5 ctest lanes
CREATE Scenarios/CsoContentAddressingScenario.cpp (G12) 1 case, 2 arms × 2 backends
MODIFY MG_IntegrationTest/CMakeLists.txt sources, a build-time capability-detection block, 9 new lanes
MODIFY trace_benchmark.{hpp,cpp} CLOCK_THREAD_CPUTIME_ID beside the wall delta; Report::frameCpuMs
MODIFY trace_replay_core.{hpp,cpp} SummarizeSeries factored out and reused for both series; meanFrameCpuMs/medianFrameCpuMs/p95FrameCpuMs and the whole frameCpuTimesMs[]
MODIFY run_android_retrace_local.py format_benchmark prints cpu mean/p50/p95/p99, p50 and p99 reduced host-side from frameCpuTimesMs[] over the same trailing window
CREATE devices/{xiaomi-adreno830.env, oppo-mali.env} both PROFILE_VERIFIED=0 — deviation D-6
MODIFY MG_Benchmark/Driver/CMakeLists.txt DriverBenchStateToggle runs mc_state_toggle
CREATE scripts/g7_negative_control.sh (D19) done
MODIFY .github/workflows/test.yml pipe-gates becomes a gate; ctest -L unit on the verify runtime; the two controls by name; a rename-proof arming grep

3. Verification — commands run and what they actually said

3.1 Section-A gates reachable on this tree

gate result
G1 symbol_report.py --before ~/w7/p2-before-libMobileGL.so --after build-linux/libMobileGL.so --threshold 0 27799 -> 27799 defined symbols: 0 added, 0 removed, 4 resized, 0 renamed; .text +160 (+0.001%). All four resized are inherited from the contract commit — RenderState::RenderState() +148, RenderState::SetCapability +77, RenderState::IsCapabilityEnabled +29, and _GLOBAL__sub_I_DirectGLES.cpp −9 (see C-2). This package changes no file the pull library is built from.
G2 ctest -N names, pull vs push diff empty, 1402 entries each
G5 awk '/namespace RenderStateImpl {/,/} \/\/ namespace RenderStateImpl/' … | sha256sum d8fd1c48716056c536752fde09db3a9e5aa8494b70404798bfc15d11220efe27 — equal to ~/w7/p2-before-syncrenderstate.sha
G7 bash scripts/g7_negative_control.sh build-push rc 2, refusing: no test matches RenderStateSpans\.SetterConsistency. Designed behaviour before package A's c2 — a missing test is not a pass
G7 mechanism … --verify-patch-only rc 0: the ColorMasks demotion applies, the patched table still compiles (the partition stays sorted/non-overlapping/complete, which is what makes the break invisible to the compiler and visible only to the test), sources restored byte-for-byte, build-push rebuilt from them, git status --porcelain MobileGL/MG_Pipe/ empty
G8 ctest --test-dir build-verify -R HandleRecycle --no-tests=error --output-on-failure green; .Legacy runs and passes on both backends, .Handles and .AbaControl skip with the reasons in §3.3
G9 python3 scripts/gen_pipe_dirty_surface.py --check rc 2 — the flag is package B's and does not exist yet (deviation D-7). --summary still works
G12 ctest -R CsoContentAddressing --no-tests=error green; every entry skips with the tracker-not-landed reason
G13 grep -rc 'pGLContext' MobileGL/MG_Backend | grep -v ':0$' empty
G13 python3 scripts/check_include_closure.py rc 0 — wire-header OK 2 headers in closure, 0 forbidden; 4 probes, 0 skipped, 0 problem(s)
G13 gen_pipe.py --check / --self-test / git diff --exit-code -- MobileGL/MG_Pipe/generated / symbol_report.py --self-test rc 0 / 0 / 0 / 0
G13 the pipe-gates stdio alternation over MG_Backend + MG_State no match
G14 ctest -N names, contract tree vs this tree 1368 → 1402: 0 removed, 34 added. The p2-before-ctest-names.txt baseline is unusable — see C-1

3.2 Suites

command result
cmake --build {build-linux,build-push,build-verify} -j 12 rc 0, 0, 0
ctest --test-dir build-linux -L unit --no-tests=error -j 8 100% tests passed, 0 failed out of 1489
ctest --test-dir build-push -L unit … 100% passed, 0 failed out of 1489
ctest --test-dir build-verify -L unit … 100% passed, 0 failed out of 1489, 13.65 s — the measured cost of the new CI step
ctest --test-dir build-linux -L integration-gpu --no-tests=error (serial, as CI runs it) 100% tests passed, 0 tests failed out of 918
ctest --test-dir build-{linux,push} -L integration-gpu -j 4 914/918 — four failures, pre-existing and local-only, see §3.6
ctest --test-dir build-verify -R 'HandleRecycle|CsoContentAddressing' -j 4 100% passed, 0 failed out of 44
ctest --test-dir build-bench -R DriverBench --output-on-failure DriverBench Passed 2.23 s; DriverBenchStateToggle Passed 0.69 s

DriverBench mc_state_toggle on this box, for the record: mc_state_toggle,120,46,3.232,70259.1,309.4 (case, frames, ops/frame, median_frame_ms, ns_per_op, fps).

3.3 The new entries, and the exact skip reasons

34 new names. Lanes: {DirectGLES,DirectVulkan}.HandleRecycle.{Handles,Legacy}., DirectVulkan.HandleRecycle.AbaControl. (DirectVulkan only — the knob reverts two DirectVulkan guards and steers nothing on DirectGLES), and {DirectGLES,DirectVulkan}.CsoContentAddressing.{On,Off}..

What actually executed rather than skipping, which is the load-bearing part of the evidence:

DirectGLES.HandleRecycle.Legacy.…AVertexArrayAtARecycledAddress…    Passed (15 ms)
DirectGLES.HandleRecycle.Legacy.…ATextureAtARecycledAddress…        Passed
DirectGLES.HandleRecycle.Legacy.…AFramebufferAtARecycledAddress…    Passed ( 4 ms)
DirectVulkan.HandleRecycle.Legacy.…AVertexArrayAtARecycledAddress…  Passed (11 ms)
DirectVulkan.HandleRecycle.Legacy.…ATextureAtARecycledAddress…      Passed
DirectVulkan.HandleRecycle.Legacy.…AFramebufferAtARecycledAddress…  Passed ( 1 ms)
HandleRecycleScenario.TheReproducerRecyclesEveryName                Passed (every lane)

They passed rather than skipping, which means the ABA really was constructed: glGenVertexArrays / glGenTextures / glGenFramebuffers handed every deleted name straight back, and in the vertex-array case glGenBuffers did too, so the replacement VAO and its buffer carry the dead objects' names with a byte-identical attribute configuration. Had any name not been recycled the case would have reported inconclusive, not proven (the ObjectLifetimeIdTest shape), not a pass. This is also the evidence that today's lifetimeId + weak_ptr guards do stop this ABA — which is what the re-key has to be at least as strong as.

Skip reasons, verbatim from ctest -V:

  • HandleRecycle.Handles — "the Handles arm needs the backend's {slot, gen} re-key, and this build does not have it: no source under MobileGL/MG_Backend/DirectGLES mentions the Track H subsystem constant (P2 package C for DirectGLES, package D for DirectVulkan). The arm is registered and visible, and arms itself when that package lands."
  • HandleRecycle.AbaControl — "the AbaControl arm needs MOBILEGL_PIPE_HANDLE_ABA_CONTROL to have a consumer, and this build has none: MG_Config parses the knob (ConfigLoader.cpp) but no source under MobileGL/MG_Backend/ reads Features.PipeHandleAbaControl, so the two guards the knob is supposed to defeat are still in force and the ABA cannot be reproduced. P2 package D owns that consumer."
  • CsoContentAddressing, pull build — "this library was built without MOBILEGL_PIPE_PUSH, so there is no render-state CSO to mint, no cso[] bracket in the summary line and nothing for the content-addressing bit to steer. The entry is registered here anyway so that ctest -L integration-gpu names the same tests in the pull build and the push build (gate G2)."
  • CsoContentAddressing, push build — "the CSO counters have no emitter in this build: MG_Impl/Pipe/Tracker.cpp does not exist, so nothing mints or binds a render-state CSO and csom / csob are structurally zero. P2 package B owns the tracker; this entry arms itself when it lands."

Every one of those skips is decided by the build, not by a hand-written guard. MG_IntegrationTest/CMakeLists.txt looks for MG_Backend/DirectGLES/SlotTables.h, MG_Impl/Pipe/Tracker.cpp and the strings kMGPipeSubsystemMagmaVertexInput / PipeHandleAbaControl inside DirectVulkan/Renderer/VertexInputStateFactory.cpp; the two existence answers use file(GLOB … CONFIGURE_DEPENDS) (re-evaluated before every build, costing nothing until the file appears), the one content grep adds that single small file to CMAKE_CONFIGURE_DEPENDS (so an edit to it reconfigures, and an edit anywhere else in the backend does not). Each verdict is printed at configure time:

Integration tests: DirectGLES has no SlotTables.h - HandleRecycle.Handles will SKIP on it
Integration tests: no MG_Impl/Pipe/Tracker.cpp - CsoContentAddressing will SKIP
Integration tests: DirectVulkan's vertex input is not re-keyed yet - HandleRecycle.Handles will SKIP on it
Integration tests: MOBILEGL_PIPE_HANDLE_ABA_CONTROL has no consumer - HandleRecycle.AbaControl will SKIP

3.4 The CPU series (D.4.1), pre-flighted on the desktop CLI

build-retrace configured and built (rc=0, 165/165). One real replay of minecraft-1.17-main-menu-854:

mobilegl_trace_replay --trace … --backend DirectGLES --mobilegl-library build-linux/libMobileGL.so \
  --benchmark --benchmark-finish=0 --benchmark-tail-frames=50
→ benchmark completed; frames=2, tailFrames=2, meanMs=655.506, medianMs=655.506, p95Ms=1290.787,
  meanCpuMs=271.766, medianCpuMs=271.766, p95CpuMs=535.271, fps=1.526

benchmark.json checked programmatically: frameCpuTimesMs[] present, same length as frameTimesMs[], every entry ≥ 0, sum(cpu) ≤ sum(wall) (wall [1290.787, 20.225], cpu [535.271, 8.261]).

format_benchmark on that JSON:

frames=2 total=1.3s tail=2 mean=655.506ms median=655.506ms p95=1290.787ms fps=1.5 \
  | cpu mean=271.766ms p50=8.261ms p95=535.271ms p99=535.271ms

and with frameCpuTimesMs removed: … | cpu unavailable (no per-thread CPU clock in this run).

3.5 The device-profile guard

bash -n clean on bench.sh and session.sh. Both exit 2 against an unverified profile with the refusal text; --allow-unverified-profile warns (three [warn] lines saying the run is not comparable with a pinned one) and proceeds. odinlite.env sets no PROFILE_VERIFIED and the guard defaults to "verified", so its behaviour is unchanged.

3.6 A pre-existing -j 4 flake this run surfaced — not caused by this package

ctest -L integration-gpu -j 4 fails four entries, identically in the pull and the push build:

DirectGLES.UnlocatedIoBlocks.…TheEmulationIsActuallyArmedWhenTheEnvironmentPinsItOn
DirectVulkan.PrimGenReroute.…TheRerouteIsActuallyArmedWhenTheEnvironmentPinsItOn
DirectGLES.PointSizeDemotion.…TheDemotionIsActuallyArmedWhenTheEnvironmentPinsItOn
DirectVulkan.PointSizeDemotion.…TheDemotionIsActuallyArmedWhenTheEnvironmentPinsItOn

Diagnosis, from LastTest.log: the failure message ends Log appended by this test: followed by nothing. Each of those pinned lanes registers a whole scenario (TEST_FILTER "<Scenario>.*") against ONE MOBILEGL_LOG_FILE_PATH, and the library opens its log fopen(path, "w") — so under -j a sibling case in the same lane truncates the file between this case's draw and its read. This is exactly the hazard MG_IntegrationTest/CMakeLists.txt documents for the arming lane; those three lanes do not follow the rule.

Evidence that it is a race and not a regression:

  • ctest -L integration-gpu serially (which is how CI runs it — test.yml lines 293-298 use no -j) is 100% tests passed, 0 tests failed out of 918;
  • the same four run alone: all pass;
  • the three pinned lanes together at -j 4, with none of this package's entries in the selection: 21/21 pass — the race is timing-dependent, not selection-dependent;
  • the failing lanes contain no entry this package added, and this package changes no library source.

Not fixed here, deliberately: the fix is to give each arming case its own filtered lane and log (six more entries across three unrelated scenarios), which is scope this package has no mandate for and which churns the ctest name set that G2/G14 are read against. Recorded for the integrator: it is latent, CI is safe from it today only because CI runs integration-gpu serially, and the rule it violates is "a log-reading case owns its lane", not "…owns its log path".

The same defect existed in this package's own first draft — two CSO cases in one lane — and is fixed in aef43358 by folding them into one case.


4. Deviations from the brief, with reasons

D-1 (blocking) — D18's AbaControl arm cannot assert the corruption on this tree, and the fix is not mine to make. D.2 says that on the contract tree .AbaControl PASSES by asserting the corrupted pixels, and C.4 says that red is the "重键前红" artefact E records first. It cannot be produced from package E alone: Features.PipeHandleAbaControl is parsed at ConfigLoader.cpp:267 and read by nothing. Its consumers — VertexInputStateFactory::ComputeHash hashing attr.Buffer.get() and LookupVaoDrawMemo skipping the vaoLifetimeId compare — live in MobileGL/MG_Backend/DirectVulkan/**, which C.5 gives to package D. Per instruction I did not edit it and record it here instead.

What was done so the gate is not lost: the arm is registered, always-on and visible; it skips with a message naming the missing consumer and its owning package; and the skip lifts automatically, with no further edit to this package, the moment PipeHandleAbaControl appears in VertexInputStateFactory.cpp. Integrator action: package D must implement the knob (D18 specifies exactly what it does), and D.2's "red before" log has to be taken after D lands the knob and before D enables its re-key, not on the contract tree. The alternative is to move the knob's consumer into E's scope, which changes C.5.

D-2 — .Handles and the CSO control skip for the same structural reason. .Handles needs C/D, CsoContentAddressing needs B's tracker. D.2 already expects .Handles to skip; the CSO skip is the same shape and is stated because C.4 does not mention it.

D-3 — the CSO lanes register in the pull build too. C.4 implies a push-only registration (ctest --test-dir build-push -R CsoContentAddressing), which is what I wrote first. It breaks G2: the push build would then name four tests the pull build does not. Commit 3766d842 registers them unconditionally and skips in the pull build with the reason; the brief's command still works unchanged.

D-4 — g7_negative_control.sh patches MGPipeRenderStateSpans.h as well as .cpp. D19 says it patches the .cpp. In the landed contract the chunk table is a constexpr boundary array in the header; the .cpp only holds the two half-arrays whose lengths follow from it. Tree wins for facts. The script also relaxes the four measurement pins (== 7, == 8, == 396, == 772) because those pin the shipped table rather than the invariant — leaving them would turn the control into a build break, proving the static_asserts work rather than that the test still checks.

D-5 — the demotion is spelled as two extra boundaries. The table alternates dynamic/pipeline by index parity, so inserting exactly two boundaries (at ColorMasks and at FramebufferSrgbEnabled) demotes ColorMasks alone and leaves every later chunk's parity, and therefore its half, unchanged. Verified: the patched table compiles.

D-6 — the two device profiles ship PROFILE_VERIFIED=0, and bench.sh/session.sh gained a refusal. The device-specific facts those profiles need — cpufreq policy names, the real OPPs, the GPU pin node, the thermal zone type — are not knowable without the devices, and 35d0befa is a Qualcomm part while the harness pins through MediaTek nodes (/proc/ppm/policy/hard_userlimit_*, /proc/gpufreq/gpufreq_opp_freq). Guessing them is the worst available outcome: su -c 'echo … > /proc/…' fails with a zero exit, so a run against a wrong profile reports numbers it believes were taken under a frequency pin. So every device-specific field is TODO_VERIFY_ON_DEVICE, both profiles declare PROFILE_VERIFIED=0, and the two scripts that pin refuse such a profile unless --allow-unverified-profile is passed. The guard itself is a deviation (C.4 asked only for the .env files); it is inside tools/device_bench/**, which C.5 gives to E, and the README records what earns PROFILE_VERIFIED=1.

D-7 — pipe-gates now runs a flag that does not exist yet. gen_pipe_dirty_surface.py --check / --self-test are package B's (C.1); on this tree they exit 2. D.1 lands tracker before gates, so by the time E merges the flags exist — but do not run gh workflow run test.yml on a tree that has gates without tracker.

D-8 — build-linux-verify's arming grep now accepts either entry-point name. Not in the brief. Package B renames MGPipeFillForVerb → MGPipeValidateForVerb (D1), and the CI step greps nm for the old name, so it would go red on the rename for a reason unrelated to what it tests. It now requires MGPipeValidateForVerb|MGPipeFillForVerb and still fails when there is neither; MGPipeVerifyInputs is unchanged and still checked on its own.

D-9 — six commits, not C.4's three. One commit covers both scenarios (they share MG_IntegrationTest/CMakeLists.txt and the capability block, so splitting them would mean two commits editing one hunk); the extra two are D-3's G2 fix and §3.6's lane fix, both found by verification after the fact.


5. Where the tree contradicted the brief

C-1 — ~/w7/p2-before-ctest-names.txt is not comparable with any pull build. It holds 2363 names. ctest -N today reports 1364 in ~/w7/pipe/build-linux, 1368 in ~/w7/p2-contract/build-linux, 1402 here, and 2238 in ~/w7/p2-gates/build-verify. So the baseline looks like it was captured from a verify-configured directory, and G14's comm -23 ~/w7/p2-before-ctest-names.txt - reports 999 "removed" tests against any pull build, including the untouched contract tree. The integrator should re-capture it from a pull build at 48268068, or state which configuration it belongs to. G14 was evaluated against the contract tree instead: 1368 → 1402, 0 removed, 34 added.

C-2 — the pull build's resized set is four symbols, not D15's three. _GLOBAL__sub_I_DirectGLES.cpp is −9 bytes at the contract commit. This package touches no file the pull library is built from, so the delta is package A's to attribute; G1 as written ("resized is empty or exactly the one mangled RenderState::RenderState() symbol") is already violated by the contract commit alone.

C-3 — the tag p2/contract is ambiguous with a branch of the same name; git worktree add … p2/contract fails outright. Use refs/tags/p2/contract or the SHA.

C-4 — D19's "patches MGPipeRenderStateSpans.cpp" — the table is in the header (D-4). D6's offset table also differs from the landed header in one row (the header puts StencilStates[0].Func in pipeline chunk P2, D6's table puts it in P3). That is package A's business and is noted only because g7_negative_control.sh had to read the real table to patch it.

C-5 — ARCHITECTURE.md:504 / tools/bench.sh — confirmed as the brief's correction 5; the script is tools/device_bench/bench.sh. This package edits that file; the doc citation is the integrator's to fix.

C-6 — test.yml runs integration-gpu serially, not -j 4. Relevant because §3.6's flake is only reachable in a parallel local run.


6. Notes the integrator will want

  • Measured CI cost of the new steps: ctest -L unit on the verify runtime 13.65 s / 1489 entries; the two negative-control lanes ~1.8 s / 44 entries; DriverBenchStateToggle 0.69 s.
  • The CSO control's numeric contract, so package B can meet it deliberately: 8 toggle pairs = 16 draws in one frame, each changing the pipeline subset. Content-addressed arm csom <= 4 and csom < csob; bit-63 arm csom == csob; both arms csob >= 16; and the readback is all-green and byte-identical across two consecutive toggle frames. If B emits the counters anywhere other than PipeStats::AddCalls(CallClass::RenderStateCsoMints/Binds, …) the control reads zeroes, and the csob >= 16 assertion turns that into a failure rather than a pass.
  • MOBILEGL_PIPE_STATS_PERIOD=1 is what makes the control readable at all: the summary line's window is "since the previous line", so the case brackets its workload with two swaps and parses the last line.
  • DriverBenchStateToggle is a "does this case still run" gate against whatever $DRIVERBENCH_EGL_LIB names, not the measurement. D.4.4's numbers still come from run_driver_bench.sh {native,espryt,magma}.
  • §3.6's -j flake is worth a follow-up on its own; the fix shape is in that section.

7. Unfinished

  • D.2's "red before" log cannot be produced by this package (D-1). As D.2 specifies it, it would today record two skips and a pass, not a red.
  • scripts/g7_negative_control.sh has never completed its real path, only the refusal and --verify-patch-only. The first end-to-end run needs package A's RenderStateSpansTest.SetterConsistency; the script exits 2 rather than reporting a pass when that test is absent, so this cannot be mistaken for a green.
  • G3 / G4 not run here. This package changes no source the library is built from, so the 40-trace SSIM corpus and the verify divergence lane would be measuring the contract commit. They belong to the integrator's D.3.
  • G6 / G10 / G11 unreachable here — A's tests, and G11 needs two devices.
  • The Android build of the CPU series is compile-unverified. trace_benchmark.cpp and trace_replay_core.cpp are shared with the Android trace app; only the desktop CLI was built and run. The change is additive and the clock is guarded for non-POSIX, but apk.yml is where the NDK build of it first runs.
  • The two device profiles are unverified by construction (D-6) and cannot be finished without the devices. No device lock was taken by this package, and no on-device measurement (D.4.2, D.4.3, D.4.5) was attempted.
  • ctest -L benchmark was run only for the two DriverBench entries. The other five benchmark targets report Not Run because only the DriverBench target was built in build-bench; they are untouched by this package.