Files
MobileGL/docs/Disaggregated/notes/p4a/p4a-results/clientsp-v3.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

29 KiB
Raw Blame History

P4a package C — clientsp. Rework result, v3

Branch refs/heads/p4a/clientsp, head 92dffb81. Not pushed. Working tree clean. Rebased onto refs/heads/feat/disaggregated = 712c9467 (c0 + c0b + c0c + c0d + c0e + wire), then four v3 commits. git diff --stat 712c9467..HEAD: 11 files, +3398 / −68; the v3 work alone (820d1a60..HEAD) is 6 files, +448 / −55.

The v2 re-review's verdict was REWORK on one major (C2-M1), three minors and three nits. All seven are closed; the rebase turned up one further defect of its own (§2.4), fixed in the same round.


0. Commits

# hash message (exact, single line, no attribution)
c1'' 1534cf37 [Feat] (Pipe): content-address sampler states, mint one sampler view per texture and push the three unit sets behind their own content hashes
c2'' 5a3e9f08 [Feat] (Pipe): publish a program's per-stage SPIR-V and reflection archive as a shader CSO and its default uniform block as global constants
c3'' bc2aad21 [Feat] (Pipe): resolve a program pipeline into one shader CSO out of the reserved composite band and release it exactly once
c4'' ccde2906 [Test] (Pipe): wire the sampler and program subsystems and pin the padding-proof CSO identity, the program-resolved unit sets, the linked-snapshot stage mask and the composite band's single release
c5' b2edca27 [Fix] (Pipe): give the sampler family the birth-hook entry points the contract declares, take the publication latch where its creates go out, and reference-count cache entries so an LRU eviction can never take a handle a standing record still names
c6' f0f8cab6 [Fix] (Pipe): give the program family its birth-hook entry point and publication latch, invalidate the global-constants key whenever a create_shader_state is re-issued, and count a truncated module tail instead of asserting it
c7' 820d1a60 [Fix] (Pipe): stop a make-current disarming the composite resolver's release path - the memo's freshness and the slot's release obligation were one flag, so after the first Reset no signature move ever spoke a delete
c8 ac09e5b3 [Fix] (Pipe): key the composite resolver's memo on the context as well as the pipeline's GL name - the resolver is a process singleton while GL names are per context, so a make-current released the other context's live composite
c9 d6e52f75 [Fix] (Pipe): count the sampler cache's unaccountable releases and its referenced evictions rather than asserting them, and state what a shared reference count, an over-capacity cache and the per-unit reconciliation really cost
c10 79b58198 [Fix, Test] (Pipe): read a forked refusal drive's log from the offset it had instead of unlinking the file - once a case initialises the library the log is already open, so the child was writing into a deleted inode and the parent read nothing
c11 92dffb81 [Test] (Pipe): drive the sampler cache's two unaccountable-release counters - a handle it never handed out and a second release of a reference only one holder owed

c1''..c7'' are v1+v2 replayed. As at v2, ccde2906 and b2edca27 do not build on their own (the wired constants and c0b's if constexpr seam are set one commit before the entry points exist); the tree builds again at f0f8cab6 and at every commit after it. The integrator squashes or merges the branch anyway.


1. The rebase onto 712c9467

Seven commits replayed; four conflicted, all in MobileGL/MG_Test/Pipe/, and all in the one replayed commit ccde2906. The cause is that the pipe's ae1a1c50 ("pin every P4a record's lifecycle…", one of wire's twelve) filled the same four stub files C owns with the applier's half of the same families: CompositeResolverTest.cpp, ImageEmitTest.cpp, ProgramEmitTest.cpp, SamplerEmitTest.cpp. MG_Impl/Pipe/* and MG_State/** rebased clean — c0d's Tracker.h, c0e's Named and wire's PipeApply.* are files C does not touch.

How they were resolved. Not by hand-picking hunks: both sides are additions to the same stub, so the resolution is a three-way merge that replays each side's insertions at its own base anchor. A script took :1/:2/:3 from the index, computed both opcode lists against the base, asserted that no base range is touched by both sides, and emitted the union in base order (the applier block first, then C's #if !MOBILEGL_PIPE_PUSH pull-skip list and its #else namespace, then main()). The overlap assertion passed on all four files. One duplicated #include <MG_Pipe/PipeApply.h> in CompositeResolverTest.cpp was removed by hand.

Nothing of wire's was lost. Every TEST( name present in 712c9467's version of the four files is present at HEAD: 4 / 5 / 7 / 8 wire cases, 0 lost. The pull-skip MGL_*_TEST_LIST macros carry only C's push-only cases, which is correct — wire's cases compile in both builds and GTEST_SKIP() internally — and G2 (pull and push ctest name sets identical) holds at 2683 == 2683.


2. C2-M1 — the composite resolver's key

2.1 What changed

what where
Entry gains a context half CompositeResolver.h:230 Uint64 ContextId, beside :231 Uint PipelineName
Find requires both halves :242-247
Observe takes the context id :132-133 Observe(Uint64 contextId, const ProgramPipelineObject&, const ProgramObject&, MGPipeHandle)
the one production call site supplies it ProgramEmit.h:118-126, inside EmitShaderState(GLContext& ctx): Observe(ctx.GetTextureContextId(), *pipeline, *drawProgram, drawCso)
Fresh and HandleFor() deleted (C2-m1) — see §2.3
Reset() reclaims instead of invalidating :188-199, counted by Counters::Sweeps :120
the file header states the key, the identity and the death path :42-69

The mismatch arm is now reachable only for this context's own entry: another context's pipeline of the same name is not found, therefore not matched and not released, and its obligation stays owed to the context that took it (:147-151).

2.2 The context identity, and why this one

GLContext::GetTextureContextId() (Core.h:156-159 → TextureState.h:135-141, TextureState::AllocateContextId, a process-wide atomic). It is not a new notion:

  • it is already the pipe's context identity — PipeFill.cpp:236-237 copies it into PipeInputs::m_textureContextId at seven fill points (FillPoints.def, kDraw / kDispatch / kClear / kBlitOrCopy / kTextureOp / kReadback / kProgramOp), and Coverage.def:229 describes it in exactly those words: "a context identity the backend keys its own tables on";
  • it is what the backend's own per-context memos key on today (DirectGLES.cpp's g_observedUnitBindingsContextId, g_unitTextureSyncListContextId, g_imageSweepContextId, g_unitSamplerWalkContextId, Managers.cpp's m_syncedShapeContextId).

Deliberately NOT the GLContext address, which is the other candidate in reach — MGB_CTX_IDENTITY (PipeInputsSwitch.h:21 / PipeInputs.h:191) and MGPipeTracker::m_context (Tracker.h:271) both compare it. Core.h:156-158 says why that would be wrong here, in its own words: "both restart at 0 in a new context, and a recreated context can land on the old heap address" — so an address key would put this same defect back one context destroy-and-recreate later, with entries of a dead context inherited by its successor.

The id travels as a Uint64 argument rather than a const GLContext& for two reasons: the resolver keeps its "depends on nothing of the family's" include set (no Core.h), and a unit case can drive two contexts in a process that only ever has one. The header says in capitals that the argument is GetTextureContextId() and nothing else, and there is exactly one production caller.

2.3 Where a destroyed context's entries are released — stated

In ProgramObject::~ProgramObject, and nowhere else. Destroying a GLContext drops m_programPipelines, which drops each ProgramPipelineObject, which drops the composite it cached in its one-slot draw-program cache; ~ProgramObject then runs MG_Pipe::MGPipeEmitShaderCsoDestroyAndFree(m_lifetimeId) (ProgramObject.cpp:58) — D-H7's second release path — which emits delete_shader_state, drops the publication latch and frees the band slot, exactly once.

The resolver never speaks a second release for such an entry, and could not: no future Observe can carry a dead context id, and MGPipeSlotAllocator::Free has already erased the lifetime-id mapping, so a release would resolve nothing. What the resolver does instead is drop the stranded entry at the next Reset() — the make-current, i.e. the one moment PipeFill.cpp:2229's FreshlyPrimed arm tells the client a context boundary was crossed. An entry is dropped when !Live or when MGPipeSlots().IsLive(ShaderCso, entry.Handle) is false, which is precisely "the obligation was discharged elsewhere". This also restores the bound the name-only key used to have: the vector is now bounded by the (context, pipeline name) pairs whose composite slot is actually live, not by every pair the process ever used.

Fresh and HandleFor() go with it (C2-m1, disposition "give HandleFor a caller or delete it, and say which in the header"): Fresh existed only because a name-only key could not tell "my own entry after a make-current" from "another context's entry", and HandleFor was its only reader and had no caller anywhere in the tree. Both are deleted and Reset()'s comment (:170-177) and Size()'s (:203-206) say so, the latter adding why a HandleFor(name) accessor must not come back.

2.4 The cases, and the mutations

Two new cases, both in CompositeResolverTest.cpp and both in the pull-skip list:

  • CompositeResolver.TwoContextsHoldingOnePipelineNameKeepTheirOwnComposites (:660) — the reviewer's probe, as a case. Two ProgramPipelineObjects built with the same external index, two composites (MakeShared<ProgramObject>(0u), external index 0, the shape Core.cpp builds), two band handles with real create_shader_state records and publication latches, then Observe(A) → MGPipeProgramEmitterInstance().Reset() (the make-current) → Observe(B). Asserts Releases unmoved, and that A's slot is still live and still published; then goes back to A and asserts the unmoved signature is a Reuses and still releases nothing. It is driven at the resolver because one test process has one GLContext; every value it supplies is what the single production call site supplies.
  • CompositeResolver.ADestroyedContextsEntryIsDroppedRatherThanReleasedASecondTime (:716) — §2.3 as a case: Observe, then drop the composite's last SharedPtr, and assert the slot went back exactly once with the latch, that the resolver spoke no release, and that the next Reset() drops the stranded entry (Sweeps + 1, Size() back to where it was, Releases unmoved).
mutation case that goes red result
Find drops the ContextId half of the key (v2's key) TwoContextsHoldingOnePipelineNameKeepTheirOwnComposites red, and only that one (11/12)
Reset() stops reclaiming stranded entries ADestroyedContextsEntryIsDroppedRatherThanReleasedASecondTime red, and only that one (11/12)
Release stops counting a handle it never handed out SamplerEmit.AReleaseThisCacheNeverHandedOutIsCountedRatherThanAbsorbed red, and only that one (21/22)

All three applied, built, run and reverted; the tree is clean and 12/12 + 22/22 green again.


3. The three minors and the three nits

# disposition
C2-m1 HandleFor() dead, Fresh unobservable, Reset() a no-op TAKEN, folded into the C2-M1 fix exactly as ruled: both deleted, Reset() given a real job (§2.3), and the header says which was chosen and why (CompositeResolver.h:170-177, :203-206).
C2-m2 Release silent on an unknown handle; "a double release is a no-op" is only true when nobody else holds the entry TAKEN, in the counted-refusal idiom C-m6 adopted. Counters::UnknownReleases (SamplerEmit.h:240) when the probe finds no entry, Counters::UnderflowedReleases (:241) when the count is already 0, both incremented in Release (:305-320). The comment above it (:288-304) now says in capitals that the count is per HANDLE, not per holder, that B's params record and this file's bind_sampler_states routinely name the same handle, and that a holder releasing twice therefore steals another holder's pin rather than harmlessly repeating itself. New case + mutation (§2.4).
C2-m3 the reconciliation's cost is understated and c0d makes it land far more often TAKEN as a code correction; the MEASUREMENTS.md half stays the integrator's (docs/ is not C's file). SamplerEmit.h:807-821 now states the real model: each Release of a non-null handle is a linear scan of the whole cache beside an Acquire probe that is another, so a pass with K units holding sampler CSOs costs 2 · K · O(cacheSize) comparisons — up to ~98 000 at K = 192 with a full cache, unbounded above that because the over-capacity path lets the cache grow past 256 — that Minecraft-shaped K is typically 0 and a CTS sampler sweep is the workload with both a large K and a large cache, that c0d's bit-13 shutter now mixes GetTextureBindGeneration() so it runs on every texture bind, that the set-hash suppressor stops the record but not the two scans, and that C-m2's per-unit latch removes both.
C2-n1 Evict's MOBILEGL_ASSERT is presented as an ID-17 mechanism but is off in every build that runs TAKEN, made a counter (the reviewer's second option). Counters::ReferencedEvictions (:245), incremented at Evict (:483-484); the comment says the sole caller structurally guarantees 0, that the eviction still proceeds (refusing would leave Mint without the room it asked for), and why an assert could not carry it. The new case asserts it is 0.
C2-n2 the cache never shrinks after an over-capacity peak TAKEN, one sentence beside OverCapacityMints (:224-232): read it as how far the cache permanently grew, not how often; the size stays at its high-water mark for the life of the process along with that many SamplerCso slots and that many applier records; bounded by the peak simultaneous pin count, so memory retention and not a leak.
C2-n3 AcquireWithForcedHashForTest is a public production method and §3's "full seam" is not full TAKEN, :279-285: the three members a consumer must not call are named — AcquireWithForcedHashForTest, ResetForTest and ResetCounters — and the sentence says the class is public because it is header-only, not because all nine members are for callers. B's seam, restated in §5 below, is the other six.

4. The defect the rebase itself turned up (79b58198)

Not in any review; found by the gate, and it is C's to fix because it exists only where the two halves meet. ctest -L unit in build-verify failed 6 of 1716 immediately after the rebase — all six of them wire's refusal cases, in the four merged files:

SamplerEmit.ARecordThatDoesNotDescribeItsOwnParametersIsRefusedNamingTheLength
SamplerEmit.AUnitWindowPastTheMergedUnitSpaceIsRefusedRatherThanTruncated
ImageEmit.AnImageWindowPastTheImageUnitSpaceIsRefusedRatherThanTruncated
ProgramEmit.ACreateWithNoArtefactsAnOversizedBlockOrACorruptSlotIsRefusedNamingTheProgram
ProgramEmit.TheDefaultUniformBlockLandsOnTheProgramsRecordAndTheSentinelIsRefused
CompositeResolver.ASlotAtTheShaderCsoLimitIsRefusedWhileTheLastBandSlotIsNot

Root cause. ExpectRefusedNaming's poison/verify arm forks a child, expects SIGABRT and then asserts the Fatal line was written to the log. RunInChild prepared for that by std::filesystem::remove(g_logPath) before the fork — correct while the library's log file was still closed, which it was in wire's own files. C's main() calls MobileGL::Initialize() (its cases need a GLContext), and Log.cpp:63-77 opens the log file with std::fopen(path,"w") once and caches the FILE*. fork() duplicates it, so after the remove() the child wrote its Fatal line into a deleted inode while the parent read a path that no longer existed: the child aborted correctly (DiedOfAbort passed) and child.Log was empty, so only the assertion on what it refused failed. It reproduced in build-verify only, because the push build's arm reads the log in-process and never unlinks it; and only in the four files C owns, because they are the only ones with both MobileGL::Initialize() and a forked drive.

Fix (all four files, identical): RunInChild remembers the log's end (LogEnd(), SamplerEmitTest.cpp:96) instead of unlinking the path, and ReadLog(from) reads back only what the child appended (:133). The comment at the call site says why the unlink cannot come back. 6/6 go green; build-verify unit is 100%.

This is worth the wire reviewer's eye as a contract between the two halves of these files: any future forked drive in MG_Test/Pipe must assume the library's log is already open.


5. The B-facing API — unchanged from v2, with C2-n3's addition

Nothing about B's seam moved in v3. Restated so v2's §3 need not be re-read:

#include <MG_Impl/Pipe/SamplerEmit.h>          // add to TextureEmit.h's include block

Uint64 samplerBytes = 0;
const MGPipeHandle builtinSampler =
    MGPipeSamplerCsoCacheInstance().Acquire(sampler->GetAllSamplerParameters(), samplerBytes);
// ... and, when this texture's params record STOPS naming a previously acquired handle:
MGPipeSamplerCsoCacheInstance().Release(previousBuiltinSampler);

The six members that are B's: Acquire (SamplerEmit.h:281), Release (:305), RefCountOf (:322), RecordIsPublished (:340), Size, GetCounters — the counter set has grown to Mints / Acquisitions / Hits / Collisions / Evictions / Releases / OverCapacityMints / UnknownReleases / UnderflowedReleases / ReferencedEvictions. The three B must not call are AcquireWithForcedHashForTest, ResetForTest and ResetCounters (C2-n3).

The four rules are unchanged, and C2-m2 sharpens rule 2: B owes exactly one Release per Acquire, and a second one now shows up as UnderflowedReleases (or UnknownReleases) rather than silently stealing bind_sampler_states' pin. B's review has a number to assert on.


6. Gate numbers (head 92dffb81, working tree clean)

CCACHE_BASEDIR=/home/swung/w7, -j 8, GLIBC_TUNABLES=glibc.malloc.tcache_count=0 exported.

gate result
three builds (pull build-linux / push build-push / verify build-verify) rc 0 / rc 0 / rc 0
G1 symbol_report --before ~/w7/p4a-before-libMobileGL.so --after build-linux/libMobileGL.so --threshold 0 --fail-on-symbol-set-change --fail-on-added-bytes 0 0 added / 0 removed / 0 resized / 0 renamed, rc 0. .text 10806323 → 10806323 (+0, +0.000%); 27811 → 27811 defined symbols, 27072 normalised names unchanged
gen_pipe.py --check / --self-test rc 0 / rc 0 (7 negative controls tripped)
git diff --exit-code -- MobileGL/MG_Pipe/generated rc 0
gen_pipe_dirty_surface.py --check / --self-test rc 0 / rc 0; 2 UNDECIDED (the contract's D6, unchanged by c0d); 27 negative controls all tripped
check_include_closure.py --mode both --compiler clang++ --self-test --require-all --expect-probes 4 rc 0, 4 probes, 0 skipped, 0 problems, 6 negative controls tripped
scripts/p3a_untouched_regions.sh 37da3c3a HEAD rc 0, the eleven byte-identical
ctest -L unit -j 8, build-linux 100% passed, 0 failed out of 1717
ctest -L unit -j 8, build-push 100% passed, 0 failed out of 1717
ctest -L unit -j 8, build-verify 100% passed, 0 failed out of 1717
ctest names pull == push diff 0 lines, 2683 == 2683
no baseline name removed (LC_ALL=C comm -23 vs ~/w7/p4a-before-ctest-names.txt) 0 removed; +95 (2588 → 2683). Of those 95, 62 are the pipe's own between 37da3c3a and 712c9467 (ID-28), so 33 are C's
the emit suites, -R 'SamplerEmit\.|ImageEmit\.|ProgramEmit\.|CompositeResolver\.' 57/57 in build-push, build-verify and build-linux (all three 100%)
TrackerWalk.* (A's c0d cases, build-push) 18/18
ctest --test-dir build-push -L integration-gpu -R DirectGLES -j 8, default arm 100% passed, 0 failed out of 491; grep -c 'Fatal{' = 0
same, MOBILEGL_PIPE_PUSH=0x1ff 100% passed, 0 failed out of 491; grep -c 'Fatal{' = 0
ctest --test-dir build-push -L integration-gpu -j 8 (whole label), default arm 100% passed, 0 failed out of 966
ctest --test-dir build-verify -L integration-verify -j 4 100% passed, 0 failed out of 844; grep -c 'Fatal{' = 0
python3 ~/w7/retrace_gate.py --tree /home/swung/w7/p4a-clientsp --lib build-push/libMobileGL.so --out ~/w7/retrace-out/p4a-clientsp-v3 -j 4 passed 79 / 79, failed [], rc 0; 0 logs with Fatal{, 0 with ssim=None

Measurement points. The three integration-gpu runs, integration-verify and the retrace were measured at 79b58198. The only change since is 92dffb81, one unit case in MG_Test/Pipe/SamplerEmitTest.cpp: build-linux/libMobileGL.so is byte-identical across it (sha256 prefix 2625db7a19a54d3a before and after) and G1, p3a_untouched_regions, unit ×3, the emit suites, the name pair and all three static scripts were re-run at 92dffb81. Nothing that those five lanes exercise changed.

The LFS precondition, stated exactly. git lfs checkout in this worktree still prints Cannot checkout LFS objects, Git LFS is not installed — git-lfs exists but git lfs install has never configured the smudge filter for this user. It does not matter: the fixtures are the real files restored by C's v1 (deviation D11) and hidden with --assume-unchanged. Measured immediately before the retrace: 94 files, 803 MB, find … -size -1k | wc -l = 0, git status --porcelain empty. Every case reported a real ssim (1.0 / 0.999x), never None, so the passed 2 / 79 false red did not occur.


7. The c0d / c0e re-run list, and the deltas

The re-review's §9 said the numbers do not carry over from v2 because c0d changes how often C's code is entered: bit 13's shutter now mixes GetTextureBindGeneration() (so EmitSamplerStates runs on every texture bind at any unit), and the else if (pipeline) arm drives bits 6/7/8/14 under a bound separable pipeline (so EmitShaderState and the composite resolver are exercised for the first time). With c0d and wire on the base, everything named was re-run:

lane v2 (base 17db7598, stub applier) v3 (base 712c9467, real applier + c0d + c0e) delta
emit suites, build-push 34 / 34 57 / 57 +23 = 20 wire applier cases + 3 of C's own
emit suites, build-verify (not reported apart) 57 / 57 —
TrackerWalk.* not run (c0d not on base) 18 / 18 new
ctest -L unit ×3 1668 1717 +49
ctest names pull == push 2634 2683 +49
names added vs baseline +46 +95 +49 (62 pipe − 16 already counted + 3 C)
integration-gpu -R DirectGLES, default 491 / 491 491 / 491, 0 Fatal unchanged
integration-gpu -R DirectGLES, 0x1ff 491 / 491 491 / 491, 0 Fatal unchanged
integration-gpu, whole label, default 966 / 966 966 / 966 unchanged
integration-verify -j 4 844 / 844, 0 Fatal 844 / 844, 0 Fatal unchanged
retrace 79 / 79 79 / 79, 0 Fatal, 0 ssim=None unchanged
G1 0/0/0/0 0/0/0/0 unchanged

(The emit-suite arithmetic, checked at the source: the four stub files each carried one case at 17db7598 and carry 4 / 5 / 7 / 8 at 712c9467, so wire added 20; C's own three are the two composite-resolver ones and the sampler release-counter one. 34 + 20 + 3 = 57. The same 46 names the pipe gained between the two bases plus C's 3 account for the +49 in the unit and name rows.)

The scenarios the re-review named as the only things that can see the paths c0d opened, all on the default arm of build-push:

scenario set result
-R 'ProgramPipeline' (the sixteen SSO conformance cases and their siblings) 22 / 22
-R 'ImageLoadStoreSso|RelinkStageSet|PostLinkAttach' 16 / 16
-R 'IntegerBorderColor|SampledSetStaleness|PixelStoreSweep' (a different sampler object rebound to a unit) 18 / 18
-R 'CsoContentAddressing|ObjectSubsystemControl|ResourceSubsystemControl' 10 / 10

What this means, stated honestly. On 712c9467 the applier is real but Espryt is still v1, so there are no Espryt handle twins at all: the default arm's shape is "the client emits, the server pulls" — every record C sends lands in MGPipeApplier() and is validated there (the Start + Count bounds, the Parameters.Size gate, the link/spirv non-null gate, the composite band's table split, the verify build's program-archive round trip all ran for the first time, which is what the 6 refusal cases of §4 exercise), while the backend still reads the pulled PipeInputs block for its drawing. So a green integration-gpu here proves the records are well-formed and nothing the client emits perturbs the pull path; it cannot yet prove Espryt consumes them. No Fatal on any arm, 0 in integration-verify, and no refusal counter moved.


8. What the integrator must re-run

  1. B's seam first (§5, ID-14 / ID-17): B takes BuiltinSampler from MGPipeSamplerCsoCacheInstance().Acquire(...), holds the reference, releases the previous handle on re-acquire and at texture death, and deletes its own MGPipeSlots().Acquire(SamplerCso, …). With wire underneath, a null or stale BuiltinSampler is Fatal{ProtocolCorruption} on every set_texture_params. B's review should now also check UnknownReleases == 0 && UnderflowedReleases == 0 after its texture cases — C2-m2 made that assertable.
  2. check_include_closure.py … --expect-probes 4 again after B adds #include <MG_Impl/Pipe/SamplerEmit.h> to TextureEmit.h — the probe count moves.
  3. MEASUREMENTS.md under ID-19, which is docs/ and therefore not C's file: C-m2's per-unit (sampler lifetime id, GetSamplingResolutionGeneration()) acquire latch with the corrected cost model (C2-m3 — 2 · K · O(cacheSize), not one array walk, and c0d's bit-13 shutter makes it fire on every texture bind), and the one extra set_global_constants per link that C-M2's case turned up.
  4. The wire reviewer should see §4 — the forked-drive/log-file contract between wire's cases and C's MobileGL::Initialize(). The fix is in C's files; the rule belongs to both halves.
  5. A DirectVulkan arm for the death/leak cases (ID-8, C-m10) — still F's, still to be confirmed rather than assumed.
  6. After esprytobj + esprytdraw: MOBILEGL_PIPE_PUSH=0x5ff / 0x9ff must be refused at D's texture-family Resolve*SubsystemArm (ID-15's fourth D-K2 row), never half-run — C emits MGPBoundView::Texture (SamplerEmit.h) and MGPImageView::Res (ImageEmit.h:97-99) as Texture handles unconditionally.
  7. The open question C named and still declines: whether MGPipeEmitSamplerCsoCreate gets its one-line call site in SamplerObject's constructor. C's reasons for declining are in clientsp-v2 §2 and are unchanged; if the answer is no, write it into ID-26 rather than leave it as a loose end.
  8. ProgramPipelineScenario and the sampler-rebind scenarios on the integrated tree once Espryt v2 lands — §7 says why a green run here is necessary but not sufficient.

9. Per-item closure table

item ruling v3 disposition
C2-M1 key Entry on (context identity, pipeline name) using the identity the pipe client already uses; say where a destroyed context's entries are released; add the probe as a case + a mutation CLOSED — GetTextureContextId(); ~ProgramObject (§2.3); 2 cases, 2 mutations
C2-m1 fold into C2-M1: give HandleFor a caller or delete it, and say which CLOSED — deleted with Fresh, said in the header
C2-m2 two counters + one sentence CLOSED — UnknownReleases / UnderflowedReleases, the per-handle-not-per-holder sentence, a case and a mutation
C2-m3 record the corrected cost model CLOSED in code; the MEASUREMENTS.md half is the integrator's (§8.3)
C2-n1 say so in the report row, or make it a counter CLOSED — made a counter (ReferencedEvictions)
C2-n2 one sentence beside OverCapacityMints CLOSED
C2-n3 tell B the three it must not call CLOSED — in the header and in §5
rebase onto 712c9467 resolve faithfully DONE — §1, 0 of wire's 24 cases lost, overlap-checked union
the forked-drive log defect not in any review FIXED — §4, 79b58198