Files
MobileGL/docs/Disaggregated/notes/p4a/p4a-results/clientsp-v2.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

31 KiB
Raw Blame History

P4a package C — clientsp. Rework result, v2

Branch refs/heads/p4a/clientsp, head 00a86d2a. Not pushed. Working tree clean. Rebased onto refs/heads/feat/disaggregated = 17db7598 (c0 + c0b + c0c), then three rework commits. git diff --stat 17db7598..HEAD: 11 files, +3027 / −52; the rework alone (f5c3364c..HEAD) is 6 files, +695 / −94.

0. Commits

# hash message (exact, single line, no attribution)
c1' 2a076438 [Feat] (Pipe): content-address sampler states, mint one sampler view per texture and push the three unit sets behind their own content hashes
c2' 94daf8ed [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' a23a07ad [Feat] (Pipe): resolve a program pipeline into one shader CSO out of the reserved composite band and release it exactly once
c4' f5c3364c [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 02475fbb [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 5f577d16 [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 00a86d2a [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

c1'..c4' are the v1 commits replayed; the rebase was clean, no conflicts (c0b/c0c touched PipeFill.cpp, SlotAllocator.*, PipeMutation.h, MGPipeTypes.h, PipeFields.def, PipeCatalogueTest.cpp and check_include_closure.py, none of which C owns or edits).

The rework is grouped by file pair, so f5c3364c and 02475fbb do not build. That is not a new fact: f5c3364c sets both wired constants and c0b's if constexpr seam then instantiates forwards to entry points that did not exist, so the branch has not compiled since the rebase. The tree builds again at 5f577d16 (both entry-point commits landed) and at 00a86d2a. Splitting the reconciliation out of the two defect fixes would have meant splitting hunks inside one file; the integrator squashes or merges the branch anyway.


1. The three majors

C-M1 — MGPipeCompositeResolver::Reset() disarmed the release path → FIXED (00a86d2a)

Live and the memo's freshness were one flag. Reset() cleared it, the reuse branch returned before anything could restore it, and from the first make-current onwards ReleaseEntry saw !Live and returned: no delete_shader_state, no Free, no counter, and the old composite's handle overwritten out of the resolver.

Split into two flags, each with exactly one meaning:

  • CompositeResolver.h:181 Bool Live — the release obligation. Set when the entry takes responsibility for a composite's slot, cleared only by ReleaseEntry.
  • CompositeResolver.h:183 Bool Fresh — the memo's "does this entry describe the current context's pipeline of this name". Reset() (:143-145) clears only this; the reuse branch re-arms it (:105-113) and Observe's mint path sets both (:125). HandleFor (:151) now asks for both, which is the only place Fresh is observable.

The reuse branch deliberately does not consult Fresh before deciding not to release: the handle is minted off the composite ProgramObject's own lifetime id, so an identical handle is an identical object and there is nothing to release whatever the memo's freshness says. Making !Fresh fall through to ReleaseEntry would have emitted a delete for a live composite.

Case added (C-m9's gap): CompositeResolver.ASignatureMoveAfterAMakeCurrentStillReleases ThroughTheResolver (CompositeResolverTest.cpp:346-403). It drives the release through the resolver — bind a real pipeline, emit, MGPipeProgramEmitterInstance().Reset() (the make-current, which reaches the resolver's Reset()), emit again on the unchanged stage set (the reuse branch), then glUseProgramStages a different fragment stage so the signature really moves — and asserts Releases == 1 and that the first composite's slot is no longer live. Both existing "FreesTheSlotExactlyOnce" cases call the death helper directly and never touch the resolver; this one does not. Negative control: restoring entry.Live = false in Reset() turns it red.

C-M2 — the (Cso, Version) global-constants latch survived a re-issue → FIXED (5f577d16)

ProgramEmit.h:259-271: on the branch of AcquireShaderCso that actually re-issues create_shader_state, m_constantsCso is reset to kMGPipeNullHandle and m_constantsVersion to kMGPipeGlobalConstantsNeverUploaded whenever the re-issued handle is the one the constants key names. Invalidated rather than re-emitted, because AcquireShaderCso has no business deciding when the constants go out — the next EmitGlobalConstants sees an unlatched key and sends them.

Two accessors so a case can see it: GlobalConstantsCso() / GlobalConstantsVersion() (ProgramEmit.h:378-379).

Case added: ProgramEmit.AReIssuedCreateReSendsTheDefaultUniformBlock (ProgramEmitTest.cpp:318-381). Negative control: deleting the four-line invalidation turns it red.

Two corrections to the review's scenario, both written into the case's comment.

  1. The behavioural half is not the discriminator. BumpLinkObservableVersions() (ProgramObject.cpp:346-359) calls MarkUBOContentDirty() beside ++m_linkVersion, so the Version half of the key moves on its own for the one re-issue trigger that exists today. The case therefore pins the mechanism — after a re-issue the emitter must hold no key at all — which is the property the contract needs and the half that goes red.
  2. A failed relink does not leave the block populated in this frontend. Link()'s prologue assigns m_spirv = {} (ProgramObject.cpp:532), so globalUboScratch is empty after any link attempt and EmitGlobalConstants early-outs on GetUBOSize() == 0. The review's "the client's scratch still holds the live values" is true of GL and not of this implementation. The fix is still right and still required — it is what makes any future re-issue trigger that does not happen to move the content version safe, of which a recycled slot is one — but the reachable damage today is smaller than the review states.

Noticed while writing the case, not fixed, recorded for ID-19: EmitGlobalConstants reads GetUBOContentVersion() (ProgramEmit.h:164) before GetUBOSize() (:168) joins phase B, and that join bumps the counter — so the first emission after a link latches a version one behind the bytes it sent and the next call re-sends identical bytes once. Over-emission, one record per link, never under-emission. Left alone because reordering the two reads changes when the join happens on a path the phase has not measured.

C-M3 — LRU eviction could take a handle a standing record names → FIXED per ID-17 (02475fbb)

Reference-counted entries, exactly as ID-17 rules.

what where
Uint32 RefCount on the entry SamplerEmit.h:331
Acquire(params, payloadBytes) returns a handle and takes a reference :244-248 → AcquireCanonical :337-359 (++RefCount on a hit :355, RefCount = 1 on a mint :410)
Release(handle) drops one; unknown handle or double release is a no-op, never an underflow :267-277
LRU considers unreferenced entries only :367-374
every entry pinned ⇒ mint beyond 256 and count it :376-381, Counters::OverCapacityMints :223
Evict asserts its victim is unreferenced :434-437
RefCountOf(handle) for diagnostics and cases :279-286
the reference a bound sampler state holds while bound EmitSamplerStates :751-783, m_stateRefs :996
that reference released at a make-current (working state, not a record) Reset() :944-960
the static_assert restated :207-209

The static_assert is kept but no longer claims to be what makes eviction safe. Its message now says what it really guarantees — a cache that could not hold one whole emission pass would over-capacity-mint on every pass — and the header block at :167-200 states plainly that correctness against eviction is the reference count.

EmitSamplerStates' reconciliation is one extra pass over the 192-entry unit array: Acquire has already taken a reference for every unit in the new set, so a unit whose handle did not move gives that duplicate back, a unit whose handle moved gives back the one it used to hold, and a unit that fell outside the window gives back its own. One binding is one reference, not one per pass.

Three cases added (SamplerEmitTest.cpp):

  • SamplerEmit.AReferencedCsoIsNeverTheLruVictim (:526-557) — a pinned value survives a miniature CTS sampler sweep (512 distinct unreferenced values through a 256-entry cache), the LRU really runs (Evictions > 0), nothing over-capacity-mints, and re-acquiring the pinned value is a hit and not a re-mint. Negative control: dropping the RefCount != 0 skip in the victim choice turns it red.
  • SamplerEmit.AFullyPinnedCacheMintsBeyondItsCapacityAndCountsIt (:559-586) — 256 pinned entries plus one more distinct value: Size() == 257, Evictions == 0, OverCapacityMints == 1.
  • SamplerEmit.ABoundSamplerStateHoldsItsCsoUntilTheUnitMoves (:494-524) — the emitter half: one binding is one reference across repeated passes, and the reference goes when the unit stops naming the CSO (released is not evicted).

2. The c0b reconciliation (review §6, items 1–8)

item disposition
1. MGPipeSamplerEmitter::EmitSamplerCso(SamplerObject&) Added, SamplerEmit.h:888-892. Body is the family's handle rule: MGPipeSamplerCsoCacheInstance().Acquire(sampler.GetAllSamplerParameters(), bytes), then Release of the reference it just took — a birth is not a standing record, so pinning the default-parameter entry for the life of the process would buy no reader anything. The comment says what the review asked it to say: minting at construction content-addresses the default parameters and the cache dedupes, so it is one create_sampler_state per distinct default-valued sampler, not one per object.
2. MGPipeSamplerEmitter::EmitSamplerView(ITextureObject&) Added, SamplerEmit.h:900-905. A wrapper over AcquireSamplerView, resolving the texture handle exactly as EmitSamplerViews does and discarding the byte count.
3. MGPipeProgramEmitter::EmitShaderCso(ProgramObject&) Added, ProgramEmit.h:292-295. A wrapper over AcquireShaderCso.
4. the publication latch, four call sites All four, and nothing else. MGPipeNoteHandlePublished(SamplerCso, cso) at SamplerEmit.h:419 (immediately after MGPipeApplyCreateSamplerState); MGPipeNoteHandleUnpublished(SamplerCso, …) in Evict (:450) and in ResetForTest (:313); MGPipeNoteHandlePublished(SamplerViewCso, handle) at :855; MGPipeNoteHandlePublished(ShaderCso, handle) at ProgramEmit.h:255. Both headers now include <MG_Pipe/PipeMutation.h> explicitly (SamplerEmit.h:47, ProgramEmit.h:44) rather than relying on CompositeResolver.h pulling it in.
5. RecordIsPublished / NoteRecordDestroyed stop being the death path's authority Kept, comments rewritten: SamplerEmit.h:288-293 (the cache's), :907-916 (the view latch's), :925-930 (NoteRecordDestroyed), ProgramEmit.h:301-308 and :317-327. Each says the death helpers read A's latch, and each says why it stays (the version-first skip is the same latch). The program emitter's bound-mirror clearing keeps its three lines and gains the "why it self-heals" the review asked for: the slot's Gen moves on reuse, so a stale record latch is refused by RecordGen == handle.Gen and the three mirrors hold a handle whose generation can never be handed out again, so the next EmitShaderState compares against a different handle and re-binds. Clearing them is the cheaper answer, not the load-bearing one.
6. the sampler CSO has no lifetime-id mapping, and D must be told Sentence in the header, SamplerEmit.h:201-205, and §3 below.
7. two now-false header comments Rewritten. SamplerEmit.h:67-83 and ProgramEmit.h:51-62 both now say the constant is part of the emission gate (four conditions in wants(), the same pair in FamilyIsLive) and is additionally a compile-time contract, and that the runtime A/B is still the mask.
8. CompositeResolver.TheCompositeBandHasExactlyOneDoor still passes unchanged Verified, unchanged and green.

The three birth hooks still have no call site, deliberately, and this is the one place the rework declines what the review implies. The entry points exist and are instantiated (the build proves it), but MGPipeEmitSamplerCsoCreate / MGPipeEmitSamplerViewCreate / MGPipeEmitShaderCsoCreate are called by nothing, so no constructor emits. Reasons, per hook:

  • MGPipeEmitShaderCsoCreate — D-H4 and C.7 put create_shader_state at the validate point ("nothing today — create_shader_state is emitted from the validate point"). A construction-time create would publish an empty archive for a program that has never linked and would force a phase-B join at glCreateProgram. Adding it would be a design change, not a reconciliation.
  • MGPipeEmitSamplerViewCreate — its call site is TextureObject's constructor, which is package B's file, and D-F2 mints the view at first use behind the version-first skip.
  • MGPipeEmitSamplerCsoCreate — the only honest call site is SamplerObject's constructor (C's file, one line). It is safe — the cache dedupes to one record for all default-valued samplers and the entry point releases its reference — but it is not needed: every consumer of a sampler CSO acquires it from the cache at the point of use. It would also make DirectGLES.cpp:209-214's backend-side SamplerObject(0) (D-F3's raw-depth-fetch sampler) emit into the pipe from inside MG_Backend during backend init.

Integrator's call. If a live birth hook is wanted for this kind, it is one line in SamplerObject::SamplerObject (MG_State/GLState/SamplerState/SamplerObject.cpp:26) and C will add it on request; the entry point behind it is already correct.


3. THE EXACT API PACKAGE B MUST CALL — quote this

In MG_Impl/Pipe/TextureEmit.h, inside EmitTextureParams, replacing MGPipeSlots().Acquire(MGPipeKind::SamplerCso, sampler->GetLifetimeId()):

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

// ... in EmitTextureParams, where MGPTextureParams::BuiltinSampler is filled:
Uint64 samplerBytes = 0;
const MGPipeHandle builtinSampler =
    MGPipeSamplerCsoCacheInstance().Acquire(sampler->GetAllSamplerParameters(), samplerBytes);
// ... and, when this texture's params record STOPS naming a previously acquired handle
//     (a parameter change that re-acquires, or the texture's own death):
MGPipeSamplerCsoCacheInstance().Release(previousBuiltinSampler);

The full seam, all in namespace MobileGL::MG_Pipe (no qualification needed inside TextureEmit.h), declared in MG_Impl/Pipe/SamplerEmit.h:

MGPipeSamplerCsoCache& MGPipeSamplerCsoCacheInstance();                                   // :469

class MGPipeSamplerCsoCache {
    MGPipeHandle Acquire(const SamplerParameters& params, Uint64& payloadBytes);          // :244
    void         Release(MGPipeHandle handle);                                            // :267
    Uint32       RefCountOf(MGPipeHandle handle) const;   // diagnostics / cases           :279
    Bool         RecordIsPublished(MGPipeHandle handle) const;                             // :294
    SizeT        Size() const;
    const Counters& GetCounters() const;   // Mints/Acquisitions/Hits/Collisions/
                                           // Evictions/Releases/OverCapacityMints
};

Types: ITextureObject::GetSamplerObject() returns const SharedPtr<SamplerObject>& (TextureObject.h:34); SamplerObject::GetAllSamplerParameters() returns const SamplerParameters& (SamplerObject.h:64). Acquire takes a non-const Uint64&, so B needs a real lvalue; fold samplerBytes into whatever B returns as payload bytes rather than discarding it, or the csob-blob accounting under-reports 100 bytes per mint.

Four rules B must honour.

  1. Delete B's own MGPipeSlots().Acquire(MGPipeKind::SamplerCso, …) line. Leaving it mints a second, record-less slot that ~SamplerObject then frees.
  2. Acquire takes a reference; B owes exactly one Release for it. The reference is what stops the LRU taking a handle a published MGPTextureParams still names — the applier does not resolve BuiltinSampler, and an eviction is not a parameter change, so nothing would refuse and nothing would re-emit.
  3. Release the previous handle whenever B re-acquires (the params moved, so the content-addressed handle moved) and when the texture's record stops standing. If B has no texture-death hook of its own — the texture death helper is A's and does not forward to the emitter — the pin survives until B's per-texture latch on that slot is overwritten by the next texture at that slot. That is bounded by the number of live texture slots, not unbounded, and the over-capacity path handles it; but a real release at texture death is better and cheap if B's latch already holds the handle.
  4. B's params gate must be keyed on GetTextureParamsVersion(), not on something a glSamplerParameter* write leaves alone — the review of B must check this. glTexParameter* moves both GetTextureParamsVersion() and BumpSamplingResolutionGeneration() (TextureObject.cpp:53); glSamplerParameter* goes through SamplerObject.cpp:65.

Include closure re-run after the seam is added is B's, but C re-ran the gate on this tree with the new flag: --mode both --compiler clang++ --self-test --require-all --expect-probes 4 → rc 0, 4 probes, 0 skipped, 0 problems, 6 negative controls tripped.

And the sentence for D (SamplerEmit.h:201-205): a content-addressed sampler CSO has no lifetime-id mapping, so NotifyStateObjectDestroyed(SamplerCso, lifetimeId) arrives from ~SamplerObject with no handle behind it, every time. A backend must not key a sampler twin on a SamplerObject's lifetime id; the twin's life is create_sampler_state → the LRU's delete_sampler_state and nothing else.


4. The ten minors and the two nits

# disposition
C-m1 dispatch-through-a-pipeline window Declared, not taken. Closing it means keying the image window and EmitGlobalConstants on GetProgramForDispatch() as well as GetProgramForDraw(), which changes what set_shader_images describes for an ordinary graphics draw too (the union of two programs' windows) and needs a dirty-bit shutter that can see the dispatch program — Tracker.h, A's file for the phase. Reach: compute through a program pipeline only, which no fixture in the phase exercises. Recorded for the integrator.
C-m2 no version-first skip in front of the CSO acquire Declared, ID-19 budget item. The refcount reconciliation adds one more pass over the unit array per firing of bit 13, so the cost the review measured grew rather than shrank. The latch that closes it is, per unit, (sampler lifetime id, ctx.GetSamplingResolutionGeneration()) — and it is now sound, which it was not before: a bound CSO's handle cannot move under the reference count. Not taken in this round to keep the change surface auditable. The integrator should record it in MEASUREMENTS.md under ID-19 — docs/ is not C's file.
C-m3 the collision branch's second eviction path TAKEN, SamplerEmit.h:337-359. On a memcmp rejection the probe now counts the collision and keeps going instead of evicting and giving up. Both halves close at once: there is no second eviction path for the reference count to police, and a later entry with the same hash whose bytes really do match is found instead of a duplicate being minted.
C-m4 dead Forget(), recyclable key TAKEN: Forget() deleted. The key-recycling reason is written into the Entry comment (CompositeResolver.h:164-175) — a ProgramPipelineObject has no lifetime id (ComputeDrawProgramSignature reads the stage programs' ids), so GetExternalIndex() is the only available key; a recreated pipeline name inherits the entry, the first Observe finds a mismatched signature and releases it, and that release resolves nothing because the allocator erases the lifetime-id mapping on Free. It costs one redundant, idempotent death notice — the same shape the composite's own second release path already has.
C-m5 one-entry MGPipeProgramOpaqueUnits memo Declared, not taken. A two-entry memo would fix it without giving up the shared instance, and it is the same object C-m1 would have to widen; both belong in one change with a measurement behind it. Perf only.
C-m6 compiled-out MOBILEGL_ASSERT on the module count TAKEN, ProgramEmit.h:233-240. The assert is gone; m_moduleTruncations counts it and TruncatedModuleCount() (:382) reports it — D-J3's counted refusal, not an assertion that vanishes at INFO.
C-m7 ARedundantRebindOfTheSameSamplerEmitsNothing has no EmitterScope REFUTED. The case has had one since v1 — SamplerEmitTest.cpp:459, the first line of the body. Nothing to fix; a sentence was added saying why the scope is load-bearing. The case also now clears the sampler it bound to unit 2 on the way out, so it cannot leak state into the new reference-count cases.
C-m8 the collision case cannot go red for its name TAKEN, and closed with the #if-free test seam the review named as the honest option: MGPipeSamplerCsoCache::AcquireWithForcedHashForTest(params, forcedHash, payloadBytes) (SamplerEmit.h:256-262) puts two genuinely different values on one hash. AHashCollisionDoesNotAliasTwoSamplerStates (SamplerEmitTest.cpp:290-328) now asserts the two handles differ, Collisions == 1, Mints == 2, and — the part the old body could not reach — that the probe keeps going past the rejected entry so re-asking for the first value is a hit rather than a third mint. It keeps the one-ULP half as the ordinary-path control. Negative control: deleting the memcmp confirm turns it red.
C-m9 no case drives the resolver's own release TAKEN — see C-M1's new case.
C-m10 no DirectVulkan arm for the leak cases Not C's. contract-v2 §4.3 assigns the leak cases to package F. The integrator must confirm F's G8b cases run on the DV lane rather than assume it. Recorded, unchanged.
n1 Evict's copy-assignment does not carry padding TAKEN, SamplerEmit.h:452-458: the sentence the review asked for, saying the assignment is safe only because every entry's padding was already zeroed by Mint's memcpy and Vector growth memmoves a trivially copyable type.
n2 SanityTest.cpp:3898 is a second producer of external index 0 Not C's file (D's). Recorded here so D is told; harmless, it never reaches the emitter, but it is a second producer of the value D5's discriminator reads.

5. Negative controls (each patch applied, built, run, reverted)

control case result
Reset() clears Live again CompositeResolver.ASignatureMoveAfterAMakeCurrentStillReleasesThroughTheResolver red
the constants-key invalidation deleted ProgramEmit.AReIssuedCreateReSendsTheDefaultUniformBlock red
the LRU stops skipping pinned entries SamplerEmit.AReferencedCsoIsNeverTheLruVictim red
the memcmp confirm deleted SamplerEmit.AHashCollisionDoesNotAliasTwoSamplerStates red

All four reverted; tree clean, rebuild rc 0, all 34 emit cases green again.


6. Gate numbers (head 00a86d2a, working tree clean)

gate result
three builds (pull build-linux / push build-push / verify build-verify), -j 8, CCACHE_BASEDIR=/home/swung/w7 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. .text 10806323 → 10806323 (+0, +0.000%), 27811 → 27811 defined symbols. rc 0
gen_pipe.py --check / --self-test rc 0 / rc 0 (7 negative controls tripped); 71 calls, 72 verify payloads, 63 PipeInputs fields, 69 verbs, 477 inventory rows, 0 UNMAPPED
git diff --exit-code -- MobileGL/MG_Pipe/generated rc 0
gen_pipe_dirty_surface.py --check / --self-test rc 0 / rc 0; 79 mutators all mapped, 0 COARSE, 2 UNDECIDED (the contract's D6, unchanged)
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 1668
ctest -L unit -j 8, build-push 100% passed, 0 failed out of 1668
ctest -L unit -j 8, build-verify 100% passed, 0 failed out of 1668
ctest names pull == push diff empty, 2634 == 2634
no baseline name removed (comm -23 vs ~/w7/p4a-before-ctest-names.txt, sorted) 0 removed; 46 added (2588 → 2634) = 13 c0 + 1 c0b + 2 c0c + 25 C v1 + 5 C v2
the emit suites, build-push, -R 'SamplerEmit\.|ImageEmit\.|ProgramEmit\.|CompositeResolver\.' 100% passed, 34/34 (29 at v1 + 5 new)
ctest --test-dir build-push -L integration-gpu -R DirectGLES -j 8, default arm 100% passed, 0 failed out of 491
same, MOBILEGL_PIPE_PUSH=0x1ff 100% passed, 0 failed out of 491
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 under GLIBC_TUNABLES=glibc.malloc.tcache_count=0 100% 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-v2 -j 4 passed 79 / 79, failed: [], rc 0; 0 logs contain Fatal{; every case reported a real ssim (1.0 / 0.999x), never None

The LFS precondition, stated exactly. git lfs checkout in this worktree prints Cannot checkout LFS objects, Git LFS is not installed — /usr/sbin/git-lfs (3.7.1) exists but git lfs install has never configured the smudge filter for this user. It does not matter here: the fixtures are already real files, restored from ~/w7/pipe by C's v1 (deviation D11) and hidden with --assume-unchanged. Measured before the retrace: 94 files, 803 MB, find … -size -1k | wc -l = 0, git status --porcelain empty. The retrace's per-case ssim=1.0 confirms it read real archives rather than the ssim=None false red.


7. c0d, and the two cross-package under-fires (review step 7)

c0d exists but is NOT on this branch's base. refs/heads/p4a/contract is at 9ea44389 ([Fix, Test] (Pipe): widen the two shutters that cannot see their own subject rather than gate their emitters on a second bit …), which is exactly C-M4 and C-M5. refs/heads/feat/disaggregated is still 17db7598, and the instruction was to rebase onto that and not to wait. So:

  • C-M4 (glBindSampler moves bit 12's generation, not bit 13's, so bind_sampler_states never fires for a sampler bind) and C-M5 (bits 6/7/8 read GetCurrentProgram(), null for the whole life of a bound separable pipeline, so a re-composited pipeline never gets a ShaderCso handle) are contract defects in Tracker.h, not C's, and c0d has taken them.
  • Nothing was re-run against c0d, because it is not on this branch's base. C's unit cases are indifferent to it by construction — they drive the emitters directly and never go through the dirty walk — but the integration-gpu arms are not: with the shutters widened, bind_sampler_states and the whole program family fire in cases where they previously did not, which is new emission traffic through C's code on the default and 0x1ff arms.
  • The integrator must re-run, after rebasing this branch onto a feat/disaggregated that carries c0d: the three integration-gpu arms, integration-verify, and the emit suites; and the sampler-bind and SSO scenarios named in review §10.6 (ProgramPipelineScenario's sixteen cases, and any scenario that rebinds a different sampler object to a unit) are the only things that can see C-M4/C-M5 at all.

8. What the integrator must re-run

  1. §3 first, before anything else — B's rework must take BuiltinSampler from MGPipeSamplerCsoCacheInstance().Acquire(...), hold the reference, and delete its own MGPipeSlots().Acquire(SamplerCso, …). With wire underneath, a null or a stale BuiltinSampler is Fatal{ProtocolCorruption} on every set_texture_params. Check B's texture cases show the handle coming from the cache, and that grep -c 'Fatal{' on an integration-verify run stays 0.
  2. Rebase onto a feat/disaggregated that carries c0d (9ea44389) and re-run §6 in full — §7 says why the integration-gpu arms in particular are not carried over.
  3. ctest --test-dir build-push -R 'SamplerEmit\.|ImageEmit\.|ProgramEmit\.|CompositeResolver\.' with wire underneath: on this branch every applier entry point is still a stub, so the Start + Count bounds Fatal, the Parameters.Size Fatal, the link/spirv non-null Fatal, the composite band's table split and the verify build's program-archive round trip all come alive for the first time.
  4. A DirectVulkan arm for the death/leak cases (ID-8, C-m10) — confirm F's G8b cases really run there, or say explicitly that they do not.
  5. After esprytobj + esprytdraw: MOBILEGL_PIPE_PUSH=0x5ff / 0x9ff (bit 10 without bit 11) must be refused at D's texture-family Resolve*SubsystemArm and logged, never half-run — C emits MGPBoundView::Texture and MGPImageView::Res as Texture handles unconditionally (SamplerEmit.h:721, ImageEmit.h:97-99), and ID-15 makes bit 10 require bit 11.
  6. The two open items C declined and named: whether MGPipeEmitSamplerCsoCreate gets a call site in SamplerObject's constructor (§2, one line, C will add it on request), and MEASUREMENTS.md's ID-19 entry for C-m2 (the per-unit acquire latch) and for the one extra set_global_constants per link noted in C-M2.
  7. retrace_gate.py after confirming tools/trace_replay/fixtures are not LFS pointers — the false red is passed 2 / 79 with ssim=None.