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

20 KiB
Raw Blame History

P2 package D — p2/magma adversarial review (v1)

Reviewer ran everything below in ~/w7/p2-magma (branch p2/magma, 4 commits on top of 9c6a8a25 = tag p2/contract). Tree left exactly as found: no file edited, no commit, no git state change. Incremental rebuilds only (cmake --build, rc 0, no source changed); ctest wrote its usual Testing/Temporary output; scratch artefacts went to /tmp and to scratchpad/wf5/rev-magma/.

Verdict: NOT APPROVED — 4 majors.


0. What I re-ran, and what actually passed

Independent re-runs (not quotations of the result file):

gate command observed
G1 python3 scripts/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; the 4 resized are RenderState::RenderState() +148, RenderState::SetCapability +77, RenderState::IsCapabilityEnabled +29, _GLOBAL__sub_I_DirectGLES.cpp -9. Same four, byte-for-byte, against ~/w7/p2-contract/build-linux/libMobileGL.so → this package adds none.
G5 awk '/namespace RenderStateImpl \{/,/\} \/\/ namespace RenderStateImpl/' MobileGL/MG_Backend/DirectGLES/DirectGLES.cpp | sha256sum d8fd1c48…0efe27, equal to ~/w7/p2-before-syncrenderstate.sha
G14 ctest -N (corrected grep) vs ~/w7/p2-before-ctest-names.txt 2367 names; 0 removed; 4 added, all four the contract's MG_Test/Pipe placeholders
G2 (name half) diff of ctest -N names across build-linux / build-push / build-nolegacy identical in all three
unit ctest -L unit in build-linux, build-push, build-nolegacy 1489/1489 in each
integration ctest -L integration-gpu -j 4 -R DirectVulkan on build-push 432/432; with MOBILEGL_PIPE_PUSH=0 432/432; with MOBILEGL_PIPE_PUSH=0 MOBILEGL_PIPE_HANDLE_ABA_CONTROL=1 432/432
G3 (DirectVulkan half) python3 ~/w7/retrace_gate.py --tree ~/w7/p2-magma --lib build-push/libMobileGL.so --out … -j 4 --only 'DirectVulkan$' 40/40 PASS, min ssim 0.997009 (minecraft-1.21.4-main-menu)
G13 grep -rc pGLContext MobileGL/MG_Backend | grep -v ':0$' empty; check_include_closure.py → 4 probes, 0 skipped, 0 problem(s); gen_pipe.py --check / --self-test rc 0/0; the CI stdio alternation over MG_Backend+MG_State → clean all green
G9 gen_pipe_dirty_surface.py --check / --self-test rc 2 / rc 2 — package B's deliverable, not this one's; recorded so it is not read as green here

I also verified two load-bearing claims the package makes but does not prove, and both hold:

  • m1's "the CSO subset is a strict superset of ComputePipelineStateHash's inputs". I built an out-of-tree offsetof probe against MG_Pipe/MGPipeValueTypes.h + MG_Pipe/MGPipeRenderStateSpans.h (scratchpad/wf5/rev-magma/probe.cpp, compiled with build-push's own include/define set). Every field ComputePipelineStateHash reads (VulkanRenderer.cpp:4952-5041) lands inside a pipeline chunk — CullFaceEnabled@824, DepthTestEnabled@540, DepthMask@548, ColorLogicOpEnabled@880, MultisampleEnabled@885, PolygonOffsetFillEnabled@886, PrimitiveRestartEnabled@890, PrimitiveRestartFixedIndexEnabled@891, RasterizerDiscardEnabled@892, SampleMaskEnabled@896, SampleShadingEnabled@897, StencilTestEnabled@898, SampleMaskValue@760, MinSampleShadingValue@764, PatchVertices@264, PatchDefaultOuterLevel@268, PatchDefaultInnerLevel@284, PolygonModeFront@868, CullFaceModeSetting@828, DepthFunc@544, LogicOp@536, BlendStates@312, ColorMasks@549 — with StencilStates@768 straddling at exactly the documented sub-member granularity (Func/ops pipeline, Ref/masks dynamic).
  • "renderPassHash keeps colorAttachmentCount / rasterizationSamples in the key". True: VkRenderPassManager.cpp:617,626 fold the draw-buffer set and the valid draw-buffer count, :661,728 fold each attachment's sampleCount. So collapsing the state hash to the CSO handle loses no discrimination.
  • The chunk table itself. I re-derived all 45 RenderState::Set* definitions (brace-balanced body extraction, scratchpad/wf5/rev-magma/s38.sh) and cross-checked each against the probe's pipeline/dynamic verdict. No G7 violation found. Spot-verified by hand: SetSampleCoverage (RenderState.cpp:786, BumpVersions) → SampleCoverageValue@752 pipeline ✓; SetPolygonMode (:174, BumpVersions) → PolygonModeFront@868 / PolygonModeBack@872 pipeline ✓; SetStencilFunc (:637) bumps m_pipelineStateVersion only when Func moves and Func is the pipeline sub-member ✓; SetPrimitiveRestartIndex (:189) is ++m_version only and PrimitiveRestartIndex@876 is dynamic ✓; SetCapability(ClipDistance0..7) (:363-381) deliberately avoids BumpVersions and ClipDistanceEnabledMask is dynamic ✓; SetClearColor (:707) / SetLineWidth (:109) / SetBlendColor (:740) / SetViewport (:69) / SetScissorBox (:926) no bump, all dynamic ✓.

So the package's arithmetic is right. What it does not survive is the four findings below.


MAJOR 1 — MOBILEGL_PIPE_LEGACY_MEMOS=0 is now unusable at bit granularity: a clear bit 0 aborts every DirectVulkan draw, from a per-draw site, and that is the exact lever G8's .Handles arm pulls

ResolveBoundRenderStateCso() (MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h:891-903):

        MG_Pipe::MGPipeHandle ResolveBoundRenderStateCso() const {
            if (!MagmaPipeSubsystemOn(MG_Pipe::kMGPipeSubsystemRenderState)) {
                MagmaPipeRequireLegacyArm("GetOrCreatePipeline");      // :893
                return MG_Pipe::kMGPipeNullHandle;
            }

MagmaPipeRequireLegacyArm (MagmaPipeArms.h:50-62) is MGLOG_F + std::abort() when MG_Config::Features.PipeLegacyMemos is false.

Two things are wrong with routing bit 0 through that lever.

  1. Bit 0 is not a Track-H subsystem. D14's runtime row says: "false: the legacy arm is never entered, and a Track-H subsystem whose bit is clear is a startup Fatal{PipeLegacyMemosDisabled}." MG_Pipe/MGPipe.h:77-78 labels only bits 5 and 6 // Track H, Espryt 0b and // Track H, Magma subsystem 4; bit 0 (kMGPipeSubsystemRenderState, :72) carries no such label and is the render-state CSO, not a memo re-key. The implementation makes the lever global, and makes the Fatal per-draw inside GetOrCreatePipeline rather than at startup.
  2. It breaks the per-subsystem A/B the lever exists for. Reproduced:
$ MOBILEGL_PIPE_PUSH=0x60 MOBILEGL_PIPE_LEGACY_MEMOS=0 \
    ctest --test-dir build-push -L integration-gpu -j 4 -R 'DirectVulkan.*CrossFrameBuffer'
    1911 - DirectVulkan.CrossFrameBufferScenario.VertexOrphanAndReupload (Subprocess aborted)
    … 9 of 9 aborted …

$ grep 'Fatal{PipeLegacyMemosDisabled}' /tmp/rev-magma-0x60.log
[FATAL]: MGPipe: Fatal{PipeLegacyMemosDisabled} GetOrCreatePipeline wanted the pre-handle arm
         but MOBILEGL_PIPE_LEGACY_MEMOS=0 forbids entering it

$ MOBILEGL_PIPE_PUSH=0x20 MOBILEGL_PIPE_LEGACY_MEMOS=0 \
    ctest --test-dir build-push -L integration-gpu -R '…IndexBufferSubData'
    0% tests passed, 1 tests failed out of 1      (Subprocess aborted)

0x60 is D18's .Handles environment written literally ("bits 5 and 6 set"), and 0x20 is Espryt's Track-H bit alone — with bit 6 clear Magma's own vertex-input site is on the legacy arm legitimately, yet the run still dies, from an unrelated subsystem's site. So package C cannot run its own Track-H A/B with the legacy lever either.

Failure scenario on the integrated tree: HandleRecycleScenario.Handles (G8, an always-on ctest entry the acceptance gate names) runs with MOBILEGL_PIPE_LEGACY_MEMOS=0. It is green only if bit 0 happens to be set and a render-state CSO is bound on every draw of the scenario. ResolveBoundRenderStateCso's second guard (VulkanRenderer.h:897-901) turns the "no CSO bound" case into the same hard abort rather than the documented fallback — and MGPipeApplyDeleteRenderState clears BoundRenderStateCso (MG_Pipe/PipeApply.cpp:179-180), so the null state is reachable, not hypothetical.

Not declared: §4 of magma-v1.md lists ten deviations and none of them is this; §7 mentions only the compile-time build-nolegacy case.


MAJOR 2 — the all-pull control arm is contaminated: MOBILEGL_PIPE_PUSH=0 still mints client handles on every draw, so D.4.3's T2 no longer means what the brief defines it to mean

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp:6603 and :7187:

#if MOBILEGL_PIPE_PUSH
        snap.vaoHandle = ResolveVaoHandle(vao);
#endif

Guarded by the compile-time switch only — not by MagmaPipeSubsystemOn(kMGPipeSubsystemMagmaVertexInput), unlike every other re-keyed site in the package (:3638, :6290, VertexInputStateFactory.cpp:59,108,128,157). ResolveVaoHandle (VulkanRenderer.h:1377-1389) calls MagmaPipeHandleOf → MGPipeSlots().Acquire (MagmaPipeArms.h:75-77) on first sight of each VAO and a Uint64 compare on every later draw. The stored value is dead in that arm: vaoMoved at :6288-6296 takes the (address, lifetimeId, configVersion) branch when the bit is clear, and snap.vaoHandle is read nowhere else.

D14 states the contract this breaks verbatim: "MOBILEGL_PIPE_PUSH=0 in the environment is the all-subsystems-pull control that reproduces P1's behaviour exactly." It does not: the control arm now pays a hash-map insert per VAO and a memoised compare per draw, and grows the allocator (see Major 3).

Failure scenario: D.4.3 defines T2 = ns_per_op(push, MOBILEGL_PIPE_PUSH=0) − ns_per_op(pull) as "P1's residual fill alone", so that T1 − T2 "isolates exactly what P2 added and what it removed". With this code, T2 also carries part of what P2 added, so T1 − T2 under-reports P2's cost by an unknown amount — on the number the day-43 GO/NO-GO reads (G11, ROADMAP.md:49,56). Two lines (&& MagmaPipeSubsystemOn(...)) would fix it. Not declared.


MAJOR 3 — Magma mints handles from the client-side allocator and nothing ever frees them: permanent, unbounded growth per VAO and per vertex/index buffer on the default push path

MagmaPipeArms.h:75-77 is the only caller of MGPipeSlots() in the whole tree, and MGPipeSlotAllocator::Free / ::Reset have no caller at all:

$ grep -rn "MGPipeSlots()" MobileGL --include=*.cpp --include=*.h | grep -v MG_Impl/Pipe/SlotAllocator
MobileGL/MG_Backend/DirectVulkan/Renderer/MagmaPipeArms.h:76:  return MG_Pipe::MGPipeSlots().Acquire(kind, lifetimeId);
$ grep -rn "MGPipeSlots().Reset\|\.Free(MGPipeKind\|Slots().Free" MobileGL   # (empty)

MGPipeSlotAllocator::AllocateFor (MG_Impl/Pipe/SlotAllocator.cpp:83-95) appends a SlotState to KindState::Slots and inserts into ByLifetimeId for every distinct lifetime id; lifetime ids are process-global monotonic atomics that are never reused (MG_State/GLState/VertexArrayState/VertexArrayObject.cpp:17-21, MG_State/GLState/BufferState/BufferObject.cpp:18-24). Magma acquires:

  • MGPipeKind::Buffer, once per enabled vertex attribute inside ComputeHash (VertexInputStateFactory.cpp:59-62),
  • MGPipeKind::VertexElementsCso, per VAO (VertexInputStateFactory.cpp:88, VulkanRenderer.h:1379-1385).

Failure scenario: a Minecraft session cycles chunk VAOs and their vertex/index buffers continuously. Every one of them permanently costs a SlotState plus an UnorderedMap node in a process-lifetime singleton that has no reclamation path — on Android, in front of the LMK. Nothing in the package's evidence can see this: the retrace fixtures and integration scenarios are seconds long. Note this is not "a P2-wide fact" as §8.4 frames it: on this tree Magma is the only consumer of the allocator, and MGPipeKind::VertexElementsCso is not one of the kinds any package's explicit-destroy work covers (D13's death notification is Espryt's six twin kinds).


MAJOR 4 — the density premise the two new direct-mapped tables are justified by is false, and the package's own D-5 says so; the result is a hot-path memo that degrades in exactly the workload the code comment names

LookupVaoDrawMemo's handle arm (VulkanRenderer.cpp:3630-3661) argues:

"slots are dense by construction (the allocator has a free list plus a high-water mark), so consecutive VAOs land in consecutive entries and the collision the address hash existed to spread does not arise below the table size."

VertexInputStateFactory::MemosFor (VertexInputStateFactory.cpp:81-97) repeats the argument for its own 2048-entry table. But by Major 3 no VertexElementsCso slot is ever freed — which magma-v1.md D-5 asserts itself ("Nothing in P2 frees a VertexElementsCso slot"). So handle.Slot grows monotonically with the number of VAOs ever created, and "below the table size" holds only for the first 2048 VAOs of the process. After that the live working set is scattered over an unbounded slot range and handle.Slot & 2047 (VulkanRenderer.cpp:3646, VertexInputStateFactory.cpp:93) collides at birthday rates.

What a collision costs is worse than what it replaced, because the handle arm deleted the two mitigations the address-hashed table had:

  • no second candidate (m_vaoDrawMemoTable[index ^ 1u], :3688) — one entry, unconditionally claimed;
  • no victim choice by bindings.frameSerial (:3694-3697) — the incumbent is always evicted.

On a claim the arm resets contentHash, layoutFactsValid and the three bindings serials (:3650-3658), and MemosFor clears the whole VaoBackendMemos (:90-94). So a colliding pair of live VAOs re-hashes its configuration, re-resolves its vertex bindings and re-derives its layout facts on every draw — in the Minecraft chunk-cycling stream the comment at :6283-6287 names as the reason the VAO memo exists at all. The old table was keyed on live VAOs and could not do this.

Failure scenario, concretely: two live VAOs whose slots are 2048 apart, alternating draws; each LookupVaoDrawMemo / MemosFor call wipes the other's memo and returns a cold entry. Correctness is preserved (the {slot, gen} compare is exact); the entire per-draw win m3/m4 is claimed for is not. The package publishes no measurement that could detect it (§8.6 is explicit that nothing here is measured), and the desktop fixtures never exceed 2048 VAOs. D-5 declares the fixed size; it does not declare that the density argument its sibling comments rest on is contradicted by its own premise.


MAJOR 5 (folded into the verdict, listed separately for the integrator) — m1, the package's principal deliverable, is exercised by nothing on this tree, so every green it reports for the CSO re-key is vacuous

The package's own §3 probe is the proof: with MOBILEGL_PIPE_LEGACY_MEMOS=0 at runtime "the first Fatal names GetOrCreatePipeline". That is ResolveBoundRenderStateCso's MGPipeHandleIsNull(boundCso) branch (VulkanRenderer.h:897-901) — i.e. no CSO is ever bound on this tree, so pipelineStateHash is never 0, entry.renderStateCso is always the null handle, and the pipeline memo runs the legacy ComputePipelineStateHash arm on every one of the 432 integration cases and all 40 retraces I re-ran. The only m1 code path any test executes is its abort path — which is the path Major 1 shows is wrong.

The brief's own criterion is ROADMAP.md:7, 每个门必须能因它存在的理由变红: G2/G3 cannot go red for m1's reason, because m1 never runs. This was avoidable inside the package's own scope — branching or rebasing onto p2/tracker (D.1 orders tracker before magma anyway), or driving one MGPipeApplyCreateRenderState/…BindRenderState pair through the applier in a unit test, would have armed it. §7 and §8.1 disclose the gap honestly; disclosure does not make the package verified.


Minors

  1. G1's admitted resize set is short by one. _GLOBAL__sub_I_DirectGLES.cpp −9 is a fourth resized symbol; brief §A G1 says "empty or exactly the one mangled RenderState::RenderState()" and D15 admits three. Confirmed present on ~/w7/p2-contract too, so it is the contract's, not this package's — but the integrator must amend D15's set or explain it, or G1 is red by its own text after every merge.
  2. D-7's claim is inaccurate. "Both acquisition sites sit behind a one-entry lifetimeId -> handle memo" — there are three acquisition sites, and the third, VertexInputStateFactory.cpp:59-62 (MGPipeKind::Buffer), has no memo: it probes the allocator's hash map once per enabled attribute, up to 32 times per ComputeHash.
  3. ComputeHash is the one re-keyed site with no arm assertion. VertexInputStateFactory.cpp:57-72 has no MagmaPipeRequireLegacyArm call, so with Features.PipeLegacyMemos=0 and bit 6 clear it silently falls back to the pre-handle GetLifetimeId() key instead of declining — inconsistent with the other five sites and with D14's "the legacy arm is never entered".
  4. First MG_Backend → MG_Impl include, and the architectural conflict is not named. MagmaPipeArms.h:22 includes <MG_Impl/Pipe/SlotAllocator.h>; ARCHITECTURE.md:9 states the server process links MG_Backend + the MGPipe object table and not MG_State, MG_Impl, glslang. check_include_closure.py has no probe rooted at any MG_Backend header (its four probes are the value / artifacts / mutation / wire headers), so G13's green says nothing about this. D-11 declares the include; it does not name ARCHITECTURE.md:9.
  5. D-3's static_asserts are sound but do not discharge D19's test. I checked the coverage list at VulkanRenderer.cpp:404-448 against DynamicTailKey's inventory (:344-350): complete (Viewports, DepthRanges, BlendColor, PolygonOffsetFactor, PolygonOffsetUnits, LineWidth, ScissorBoxes, and Ref/ValueMask/WriteMask per face), and not vacuous — the inverted ScissorTestEnabledMask assert at :445-448 proves MagmaRenderStateRangeIsDynamic can return false. But D19 still lists DynamicChunksCoverMagmasDynamicTailKey among the four tests package A fills into RenderStateSpansTest.cpp; the integrator must ensure the outcome is not "neither".
  6. D-4 verified. ScissorTestEnabledMask@900 is in pipeline chunk P6 and SetCapability(ScissorTest) (RenderState.cpp:356-362) does call BumpVersions(), so the tree is right and D12.3 is wrong as written. The tail stays correct because BumpVersions moves both counters.
  7. D-2 verified. MGPipeApplyBindRenderState writes both versions (MG_Pipe/PipeApply.cpp:165), so GetRenderStateParametersVersion() does still move on a pipeline-only change. The tree wins over D12.3; the integrator's call whether to add a dynamic-only version to the wire.
  8. §5.5's correction to the brief is right and load-bearing. The brief's G2/G14 grep -E '^\s+Test #' matches only 4-digit ids (ctest -N right-aligns); I used the corrected grep -E '^\s+Test\s+#[0-9]+:' throughout and got 2367 vs the brief's form's 1368. Fix BRIEF-P2.md §A before wsl_integrate_p2.sh uses it.
  9. The reported flake did not reproduce. PrimGenReroute…TheRerouteIsActuallyArmed… and PointSizeDemotion…TheDemotionIsActuallyArmed… passed in every one of my four -L integration-gpu -j 4 -R DirectVulkan runs. Plausible as a load-sensitive lavapipe flake, but unconfirmed here.
  10. MAGMA_TAIL_INPUT_IS_DYNAMIC(Viewports) / (DepthRanges) / (ScissorBoxes) assert over the whole array while the tail reads only index 0. Stricter than needed and harmless, but a future per-viewport chunk split would break the build with a message that does not say why.
  11. MGLOG_D_ONCE at VulkanRenderer.h:898 sits on a per-draw path. Compiled out at MOBILEGL_LOG_ACTIVE_LEVEL=INFO (every build here), and _ONCE elsewhere, so it is not a ROADMAP.md:7 violation — recorded because §5.8 is right that the fallback leaves no trace by default, which is what made Major 5 easy to miss.

What I could not reach

  • G4 (verify): no build-verify on this tree, and C.3 asks for none.
  • G6/G7/G8/G10/G12: packages A, B and E own the artefacts.
  • G3's DirectGLES half and the 79-case total: this package is DirectVulkan-only by construction.
  • G11 / all of D.4: device work.
  • Any behavioural check of m1 (Major 5): impossible on this tree by construction.