Files
MobileGL/docs/Disaggregated/notes/p4a/p4a-results/clientfb-review-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

30 KiB
Raw Blame History

P4a package B — clientfb, focused re-review v2

Reviewed: refs/heads/p4a/clientfb @ d1215fbc against clientfb-review-v1.md (M1–M4, nine minors, the c0b/c0c reconciliation, D4), clientfb-v2.md, and the rulings ID-8 / ID-12 / ID-14 / ID-17 / ID-18 / ID-19(c) / ID-21 / ID-25 / ID-27 / ID-28 / ID-33 / ID-34 / ID-35.

Method — this round is a REPLAY, not a read. B's eight commits were cherry-picked onto the real pipe head 92dffb81 in a private worktree (p4a/bint-review), built (push, verify and pull), and the four lanes run. Every number below is off that tree. Every finding carries file:line, a failure scenario and the refutation I tried. The worktree and its three build dirs were removed afterwards.


Verdict: ACCEPT WITH MINORS

All ten items of the v1 review and every ruling are closed, and closed by code rather than by argument: the region tail reaches the applier and the case asserts on what the applier stored; the three GL_Texture.cpp hooks cover every path that writes the built-in sampler and the latch reads both counters; the clears are acceptance-gated with a refusal case that constructs a real refusal; the metadata respecify moves the hint on an immutable texture without eating its pending upload; the Named sites cover every DSA framebuffer entry point in the tree. Nothing new was introduced: G1's mechanism is intact (every MG_State and GLImpl statement is inside #if MOBILEGL_PIPE_PUSH, the pull build and its 1759-case unit lane are green on the replayed tree and G2 is 0 diff lines), the exit-order rule holds, MGLOG discipline holds, and the file list is B's plus the two ID-8 grants.

The five findings are minors and one of them is a documentation inaccuracy in B's own report that the integrator should not carry forward as a proved property.

Both of B's findings are confirmed and neither is a B defect — §5.


1. The lane numbers on the replayed tree

Replayed HEAD aa0fd5aa = 92dffb81 + B's seven cherry-picked commits (§4).

lane replayed tree B's report (its own base)
build-push configure + build rc 0 rc 0
build-verify configure + build rc 0 rc 0
build-linux (pull) configure + build rc 0 rc 0
ctest -L unit build-push 1759 / 1759 1756 / 1756
ctest -L unit build-verify 1759 / 1759 1750 / 1756 — the six vanish, §5(b)
ctest -L unit build-linux 1759 / 1759 1756 / 1756
G2 pull vs push ctest name sets 1759 == 1759, 0 diff lines 0 diff lines
emit suites (the eight *Emit + CompositeResolver) 152 / 152 148 / 148
integration-gpu -R DirectGLES -j 8, default mask 444 / 491 — 47 failures 444 / 491
the 47 by name vs clientfb-v2.md's appendix IDENTICAL, 47 == 47, 0 diff —
integration-gpu, MOBILEGL_PIPE_PUSH=0x1ff 491 / 491 491 / 491
integration-gpu, MOBILEGL_PIPE_PUSH=0x3ff 491 / 491 491 / 491

The three-count deltas (1756 → 1759, 148 → 152) are C v3's own new cases, which are on 92dffb81 and were not on B's temporary base. The default arm's shape and count reproduce exactly, which is what makes §5(a)'s ruling checkable rather than reported.


2. Findings

n1 — clientfb-v2.md §2(5)'s M3 sentence is not what the code proves. TextureEmit.h:1064-1097

The report states "dispatched is the if constexpr itself, so a discarded call never clears either." The code is

Bool accepted = false; Bool dispatched = false;
if constexpr (MGPipeTextureRecordsReachTheApplier()) { dispatched = true; accepted = MGPipeApplyResourceSubData(...); }
...
if (dispatched && !accepted) { ++m_refusedSubDatas; MGLOG_E_ONCE(...); return false; }
mipmap->MarkStorageDirty(uploadTarget, level, false);

so with the constant back at 0 dispatched == false, the guard is false, and the flag is cleared. The same shape is in PublishCreate (:1050, if (dispatched && !accepted) return; — a discarded create still takes MGPipeNoteHandlePublished) and NoteRespecified (:1069, the discarded arm always advances LastDesc).

Failure scenario. The A/B arm the predicate exists to keep compilable — an operator or a bisect setting kMGPipeWiredTextureSubsystem = 0 — would clear a level's dirty flag with the record nowhere, i.e. v1's M3 data loss reintroduced by the very arm that is supposed to be inert, and MGPipeHandleIsPublished would answer true for handles the applier never received.

Refutation attempted, and it succeeds — for reachability, not for the sentence. PipeFill.cpp:1059-1092 gates all six birth hooks on FamilyIsLive(kMGPipeSubsystemTextureResources, kMGPipeWiredTextureSubsystem) and forwards through ForwardWhenWired<kWired>; with the constant 0 NoteLevelDirty is never reached, m_drain is permanently empty, and DrainTextureSubData (TextureEmit.h:786) returns at if (m_drain.empty()) before EmitOneLevel exists on the call graph. PublishCreate and NoteRespecified are unreachable for the same reason. So all three sites are dead code on that arm and the property holds — because the family is gated off upstream, not because of dispatched.

Not a defect. But the integrator should not carry the sentence forward as a local guarantee, and d1215fbc's comment ("the two places a discarded call may not be treated as an accepted one") names PublishCreate and NoteRespecified while the third site, the one M3 is actually about, reads the other way. One line — if (!dispatched || !accepted) return false; — makes the paragraph true at zero cost on the live arm; PublishCreate / NoteRespecified must keep their current shape, since on the A/B arm their "pretend it landed" bookkeeping is what keeps the dedupe coherent.

n2 — a bound-target record emitted from a DSA site latches on dispatch, not on acceptance. FramebufferEmit.h:319-360

EmitFramebufferByName writes m_lastEmitted[kDraw|kRead] (or m_named[slot]) from Emit(), which counts bytes. MGPipeApplySetFramebufferState returns void and can still refuse — a null handle, Fbo.Slot >= kMGPipeMaxFramebufferSlots, or an out-of-range DrawBuffers[i] (PipeApply.cpp:2157-2190). A refused record therefore latches and is never re-sent until its content moves: M3's class, one family over.

Refutation, and it holds today. Two of the applier's three refusals are unreachable from this client: HandleFor (:254) returns kMGPipeDefaultFramebuffer {0,1} or an allocator handle, and MGPipeSlotAllocator::Allocate documents and enforces "never returns slot 0 (reserved: null, and the default framebuffer for kind Framebuffer)", so the null-handle arm cannot fire; and m2's client-side gate (FramebufferEmit.h:160-172 + :479-495) refuses an over-wide draw-buffer token before the record is built. The third needs 65 536 live framebuffers and is a loud MGLOG_E on the far side. Pre-existing in v1, untouched by the rework, and out of B's scope — record it for wire's next round, since the resource family now has acceptance returns and this one does not.

n3 — the built-in sampler's cache reference is given back only when the texture's SLOT is recycled. TextureEmit.h:1017-1024

RetireIfRecycled (and ResetForTest, :911-918) are the only MGPipeSamplerCsoCache::Release callers on this side; I confirmed MGPipeEmitTextureDestroyAndFree (MG_Impl/Pipe/PipeFill.cpp:1197-1228) never reaches this emitter, so a destroyed texture leaves one reference standing.

Failure scenario. A workload whose live-texture count falls — a dimension unload, a resource-pack swap — parks (peak − live) references the LRU cannot evict; C's cache then mints over capacity and OverCapacityMints / referenced evictions climb, unattributably.

Refutation. NotifyAndFree returns the slot to the free list at death, MGPipeSlotAllocator::Acquire pops the free list before the high-water mark, and AcquireTexture runs RetireIfRecycled on the very next texture creation — so the residency bound is the texture-slot high-water mark, which is exactly the bound m_textures itself already carries. It is monotone-bounded, not a leak, and C v3 counts rather than asserts (92dffb81). Accepted as declared (B says "bounded by live texture slots"; "peak" is the exact word). The integrator should read C's two counters once D lands and the texture family runs real traffic.

n4 — the never-destroyed singleton's stated reason is now false. TextureEmit.h:1119-1124

The comment justifies "never destroyed" with "~TextureObjectBase reaches this emitter through the death helper's RecordIsPublished / NoteRecordDestroyed pair, which read and WRITE its tables". That pair was deleted in this very rework (the latch is A's), and the death helper does not touch the emitter at all. Documentation only: ID-8's rule is unconditional for every MGPipe process singleton, the object is still new'd and never deleted, and FramebufferEmit.h:631-637 states the unconditional reason correctly. Reword at the next touch of the file.

n5 — RepublishMask now runs inside the framebuffer record build. FramebufferEmit.h:539+ → TextureEmit.h:1084-1101

SurfaceOf calls NoteTextureBoundAs, which after M4 can emit a resource_respecify. So attaching a texture to an FBO emits a resource record from inside BuildFramebufferState's attachment walk.

Refutation, and this one is right by construction. The ordering is the correct one — the resource record precedes the surface that names it — there is no re-entrancy (RepublishMask touches only m_textures and the applier; BuildFramebufferState holds no reference into either), and NoteTextureBoundAs's if (now == before) return; makes every attachment after the first free. Not a defect; recorded because it is a new emission on the validate-point path whose cost will show up in D's payload histogram, and because on the 0x3ff arm RepublishMask's own MGPipeTextureSubsystemEnabled() gate is what keeps it silent (verified: 0x3ff is 491/491).


3. The per-item closure table

item verdict evidence on the replayed tree
M1 region tail reaches the applier CLOSED TextureEmit.h:1068-1069 passes m_regions.empty() ? nullptr : m_regions.data(). TextureEmit.TheApplierStoresTheRegionListTheEmitterBuiltAndNotAnEmptyOne (TextureEmitTest.cpp:731-765) asserts on MGPipeApplier().TextureResources[slot]'s PendingUpload::Regions, field by field (X, Y, W, H, SrcOffset, SrcRowStride, SrcSliceStride) — the STORED list, and it first asserts RefusedSubDataCount() == 0 so a declared-but-missing tail fails loudly.
M2 three GL_Texture.cpp hooks CLOSED All 36 GetSamplerObject()->Set* writes in the file (:1262-1320, :1379-1422, :2083-2142) are inside one of the three hooked functions TextureParameterObject_State (hook :1365), TextureParameterObjectf_State (:1467), TexParameterf_State (:2181). Every other family funnels in: TexParameteriv/IIv/IUiv → TexParameteri_State:2196 → TextureParameterObject_State, TexParameterfv:2233 → TexParameterf_State, DSA :6334/:6340 → the two Object forms. Every return; before the hook in all three switches is a RecordError arm (checked at :1296 :1331 :1341 / :1433 :1443 / :2115 :2157); every success arm breaks and reaches the hook. The four border-colour helpers at :44-70 are outside the three, and correctly so: MGPTextureParams (MGPipeTypes.h) carries Res, BuiltinSampler, BaseLevel, MaxLevel, Swizzle[4], DepthStencilMode, ForceResync, SamplerResync, MinLod, MaxLod, LodBias and no border colour, and MGPipeBuildTextureParams (TextureEmit.h:303-331) reads only those three floats off the sampler object. So the record has no field a path outside the hooks can move.
M2 both-versions latch CLOSED TextureEmit.h:660-673 reads GetTextureParamsVersion() and sampler->GetVersion() and ForceParamsResync. ALodWriteOnTheBuiltinSamplerRepublishesTheParams asserts the texture's own version did not move before driving the sampler write — the case cannot pass for the wrong reason. MinLod/MaxLod/LodBias all asserted.
M2 no double Release, no leak CLOSED (with n3) :675-700: one Acquire per emission; the duplicate reference is handed straight back when the value did not move, the previous handle released when it did; SamplerEmit.h:305 makes Release(null) a no-op. Exactly one reference per entry. Death: n3.
M3 clear on acceptance CLOSED (with n1) TextureEmit.h:1090-1097; ARefusedUploadLeavesTheLevelDirtyAndOnTheDrainList constructs a real refusal via MGPipeApplierReleaseObjectRecords() and asserts flag + drain entry + counter, then drives the self-heal. The if-constexpr-discarded case is n1.
M3 refusal survives CLOSED as above; plus the second finding B made itself (a refused respecify republishes the create and retries once, :600-621) is pinned by the same case's second half.
M4 metadata respecify CLOSED RepublishMask (:1084-1101) starts from entry.LastDesc and writes only BindMask + ImageBindableHint. AnImmutableTexturesImageBindableHintReachesTheApplierAfterItsAllocation (:819-853) allocates immutable storage, stands a pending upload, then image-binds, and asserts on the applier's record: hint 1, mask set, Width unmoved, serial advanced, pending upload survived.
ID-19(c) the Named sites CLOSED All 23 Named*Framebuffer* symbols in the tree enumerated. 17 call sites over 12 entry points, not sixteen — B's table counts glBlitNamedFramebuffer's two objects as one row. Mutators, published after the mutation: :1461 :1490 (Texture), :1516 :1546 (TextureWithUploadTarget, i.e. the 1D/2D helper), :1603 :1708 (TextureLayer), :1780 :1792 (Renderbuffer), :1994 (DrawBuffers), :2011 (DrawBuffer), :2023 (ReadBuffer) — 11. Consumers, published before MGP_FILL: :672 :673 (blit, both objects), :688 :702 :716 :730 (the four clears) — 6. Nothing is missed: every ->AttachTexture / ->AttachRenderbuffer / ->Detach / draw- or read-buffer mutation in a Named* entry point is followed by a publish. NamedFramebufferTexture3D is a pure RecordUnsupported… stub; InvalidateNamedFramebuffer{Data,SubData} validate only; the three getters and CheckNamedFramebufferStatus are queries. NamedFramebufferParameteri is declared-not-hooked and that is safe by a stronger argument than B gives: it moves only the no-attachment defaults FillGeometry reads, and every consumer republishes before handing the object over, so a stale record can never reach the server.
ID-19(c) default framebuffer CORRECT, not wrongly emitted Every framebuffer == 0 path reaches pDefaultFramebufferInfo->defaultFBO, HandleFor returns kMGPipeDefaultFramebuffer {0,1} and IsDefault = 1 with the colour surface under Color[0] (:504-512). That is the record ID-19(c) wants for glClearNamedFramebufferfv(0, …). No collision in m_named[0]: the allocator never returns slot 0 for kind Framebuffer.
ID-19(c) ReadSurface CLOSED out.ReadSurface = SurfaceOf(fbo, fbo.GetReadBuffer()) (:520); BuildFramebufferState lost its readFbo parameter entirely, so the v1 shape is unrepresentable.
ID-19(c) sticky mask + content hash CLOSED MGPipeCopySurfaceForHash copies the new TextureTarget (:205-209); Target is a hash input, so a Named record can never be suppressed against a bound one; the per-object m_named table with the generation checked buys the other direction; Reset() clears it in the safe direction.
ID-19(c) no Named record for a null handle CLOSED PipePublishFramebufferByName returns on !fbo; HandleFor cannot produce null.
c0b no MG_State TU includes an emitter CLOSED grep -rn "MG_Impl/Pipe/" MG_State/ returns three comment lines and no #include. RenderbufferObject.cpp now includes MG_Pipe/PipeMutation.h.
c0b helper/latch deletions CLOSED all eight free functions, Entry::Published, both RecordIsPublished, both NoteRecordDestroyed, PublishedIn, Retire, ArmForTest/m_armed gone; both destructors call one helper.
c0c helper/constant deletions CLOSED MGPipePackSubDataTarget/…ResourceTargetOf/…UploadTargetOf, kMGPipeDepthStencilModeDepth/Stencil, kMGPipeSurfaceKind* all deleted; MGPipeDepthStencilModeByte kept.
c0c TextureTarget filled, 0xFFFF elsewhere CLOSED MGPipeEmptySurface() (:113-118) sets kMGPipeSurfaceNoTextureTarget (= 0xFFFF, MGPipeTypes.h:533) and TextureUploadTarget::Unknown, used on the empty point and both SurfaceOf early returns; the texture arm writes GetTarget().
D4 flip verified, expected-failing set empty CLOSED kMGPipeWiredTextureSubsystem = kMGPipeSubsystemTextureResources; the five named cases all pass in the 1759/1759 push lane, and again in verify and pull.
the six minors fixed CLOSED m1 targets.size() == 1 (:88-91); m2 the second refusal loop (:479-495); m3 the owner acquired before any Entry& (:544-590); m4 RetireIfRecycled in both Acquire*; m5 dissolved by c0e's per-object rule; m6 the Unknown sentinels. Each has its own named case.
the three declared JUDGED SOUND §3 below.
exit order (ID-8) HOLDS (n4 on the comment) both emitters new'd and never deleted; Entry::Texture is a raw pointer; the new Entry::BuiltinSampler is a POD handle — no new static holder of a frontend SharedPtr.
MGLOG discipline / no leftovers CLEAN B's two headers contain exactly four log statements, all MGLOG_E_ONCE on error paths (TextureEmit.h:623 :1092, FramebufferEmit.h:472 :489); B's added lines in the two GLImpl files contain no log statement at all; no TODO/FIXME/printf.
files vs ownership CLEAN 15 files: B's six MG_State + two MG_Impl/Pipe + two MG_Test/Pipe + the granted MG_Test/State/ObjectLifetimeIdTest.cpp + the two ID-8 grants GL_Texture.cpp and GL_Framebuffer.cpp. Nothing outside.
clientsp v3 API shapes CONFIRMED B calls MGPipeSamplerCsoCacheInstance().Acquire(params, Uint64& payloadBytes) and Release(handle) — SamplerEmit.h:266 / :305 on 92dffb81 — plus RefCountOf / RecordIsPublished, which is exactly the seam C v3 names as B's. Nothing else. It compiles and the emit suites are 152/152, so the v2→v3 move did not touch it.
wire v3 acceptance returns CONFIRMED PipeApply.h:786 :834 :863 all return Bool on the pipe head; B reads all three.

3.1 The three declarations, judged

  • m8 (SetInternalFormat's zero-extent respecify) — accept. Gating it needs a TextureObjectBase member, which resizes the pull build's object and breaks G1; every applier body tolerates a zero-extent respecify, and it is not byte-equal to the create so it is a real record rather than a duplicate.
  • m9 (no DSA-named path in G6's case) — accept, and it is now stronger than a declaration: ID-19(c)'s two new cases (AFramebufferHandedOverByNameGetsANamedRecordWithoutMovingABinding, ANamedRecordIsSuppressedPerObjectAndNeverAgainstABoundRecord) give the DSA paths coverage one level up.
  • D14 (the three GL_Texture.cpp sites unpinned by a unit case) — accept. A MG_Test/Pipe binary cannot drive glTexParameterf; the unit case drives the exact call the sites make, and the wiring rides the integration lane's texture family. Note this is exactly the class of thing the default arm will exercise once D lands.
  • D15 (the ten mutators publish after the mutation) — accept. FillPoints.def is A's and has no verb for any of them; "the point at which the object is final for this call" is the only available reading, and the five consumers — the calls that actually hand the object over — do publish before MGP_FILL.
  • D12 (no clean arm on the drain list) — accept, verified. EmitOneLevel's third statement is if (!mipmap->IsStorageDirty(uploadTarget, level)) return true; and true drops the entry from both lists; ALevelMarkedCleanIsCollectedAtTheNextDrain pins the collection and the re-append after a later write, which is the half that would have made this a texel-loss bug.
  • D13 (clientsp merged as a fourth head) — dissolved by the replay. On 92dffb81 the cache is in the tree; the merge disappears with the other three.
  • D16 / D17 — recorded; D17 (the dedupe can suppress a respecify the applier no longer holds) is honest and its blast radius is bounded by the loud refused-upload path that heals it.

4. THE VERIFIED REPLAY RECIPE

Verified end to end: it applies with exactly one conflict, builds in all three configurations, and reproduces B's lane numbers.

cd /home/swung/w7/pipe
git worktree add -b p4a/bint-review /home/swung/w7/p4a-bint refs/heads/feat/disaggregated   # = 92dffb81
cd /home/swung/w7/p4a-bint
git submodule update --init --recursive
cp -a /home/swung/w7/pipe/3rdparty/glslang/External 3rdparty/glslang/
cp -a /home/swung/w7/pipe/tools/trace_replay/fixtures/. tools/trace_replay/fixtures/
git ls-files -m tools/trace_replay/fixtures | xargs -r git update-index --assume-unchanged

# B's three v1 commits, as they sit on the temporary base:
git cherry-pick -x 5e87d3d2     # mint {slot,gen} for texture + renderbuffer        -> clean
git cherry-pick -x 5fac004a     # push the bound framebuffers                       -> clean
git cherry-pick -x e2bf183a     # wire bit 9 + the emitter cases                    -> CONFLICT, see below

#   549ef30b IS SKIPPED — see §4.1

# B's four remaining rework commits:
git cherry-pick -x a390bb1a     # birth hooks, D4 flip, region tail, acceptance     -> clean
git cherry-pick -x d013b7ec     # the two granted GLImpl files                      -> clean
git cherry-pick -x 4ec43194     # the twelve new cases and the six changed          -> clean
git cherry-pick -x d1215fbc     # the docs refresh                                  -> clean

Configure with ~/w7/notes/tools/wsl_tree.sh's COMMON flags, CCACHE_BASEDIR=/home/swung/w7, -j 8: build-push (-DMOBILEGL_PIPE_PUSH=ON), build-verify (-DMOBILEGL_PIPE_VERIFY=ON -DMOBILEGL_ITEST_REQUIRE_GPU=ON), build-linux (no extra flag).

4.1 549ef30b must be SKIPPED, not resolved

It retires wire v3's local kMGPipeFramebufferTargetNamed for c0e's enumerator. Wire v3 already did that itself on the pipe (712c9467, ID-21/ID-27/ID-28): the constant does not occur anywhere in 92dffb81 — git grep kMGPipeFramebufferTargetNamed 92dffb81 is empty — and wire's own cases already spell static_cast<Uint8>(MGPipeFramebufferTarget::Named). Cherry-picking it conflicts in PipeApply.h against a hunk that no longer exists. B's own §7 item 7 anticipated exactly this. git cherry-pick --skip, or simply omit it from the list as above. Nothing in B's remaining commits references the retired constant.

4.2 The one conflict, and its recorded resolution

e2bf183a conflicts in MG_Test/Pipe/FramebufferEmitTest.cpp and MG_Test/Pipe/TextureEmitTest.cpp — wire's cases (already on the pipe, from ae1a1c50 / 5ab90dec / 8f1eaafa) and B's cases are appended at the same anchor.

One conflict hunk per file, and the base region of each is EMPTY (Framebuffer: ours 359 / base 0 / theirs 360; Texture: ours 620 / base 0 / theirs 477) — i.e. both sides are pure additions with nothing removed, the same shape B resolved its own wire merge with. Resolve as an overlap-checked union with B's block FIRST (theirs before ours), which reproduces the layout p4a/clientfb itself has: on B's branch its own cases sit above wire's (TextureEmitTest.cpp:319..991 are B's, 1074.. are wire's).

This ordering is load-bearing. Resolving ours-first also compiles, but the later 4ec43194 then fails to apply (4 hunks, non-empty bases) because its context assumes B's block comes first. With theirs-first every remaining commit applies clean.

Checks run on the union result, all green: 0 duplicate TEST( names in either file, exactly 1 MGL_*_CLIENT_TEST_LIST macro per file (the macro region merged automatically, outside the conflict, and carries both sets: 13 framebuffer + 21 texture client entries), and no overlapping helper names — B's side adds MakeTexture2D / MakeColorTexture / BindDrawAndRead, wire's adds TextureDesc / TextureParams / TextureUpload / Region / TextureRecordOf / FramebufferRecord, each inside its own namespace { … } under its own #if guard.

4.3 Fidelity of the replay

git diff d1215fbc <replayed HEAD> over B's own paths is empty: MG_Impl/Pipe/TextureEmit.h, MG_Impl/Pipe/FramebufferEmit.h, MG_Impl/GLImpl/Texture/GL_Texture.cpp, MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.cpp and all of MG_State/ are byte-identical. The only residue is wire v3's and C v3's own tails (CompositeResolver.h, SamplerEmit.h, ProgramEmit.h, PipeApply.{h,cpp}, C's four test files) plus 3 lines in the two conflicted test files: wire's Named-constant retirement applied to wire's own cases, and one blank line from the union. So the replayed tree carries B's work exactly.


5. B's two findings, checked

(a) The 47 default-arm failures are bit 10 with no server consumer — confirmed, and it is NOT a B defect

Reproduced exactly: 444/491, and the 47 names are identical to clientfb-v2.md's appendix, 0 diff. 0x1ff and 0x3ff are both 491/491 on the replayed tree, so bit 9 — every line of B's framebuffer work including the 17 Named sites — is independently clean and the cause is bit 10 alone.

An independent negative control, stronger than B's probe-variable one. On this tree MGPipeResourceRecord::PendingUploads is read by MG_Pipe/PipeApply.{h,cpp} (the applier's own storage), MG_Test/Pipe/TextureEmitTest.cpp and MG_Impl/Pipe/TextureEmit.h — and by nothing in MG_Backend/ (VkTextureManager::FlushPendingUploads is DirectVulkan's unrelated internal staging, not the applier's set), and no file under MG_Backend/ reads MGPipeApplier().TextureResources[…] at all. So the accepted uploads land in a set that has no reader: package D is the consumer and is not in this tree. The client half of D-D5's inversion is present and correct and is destructive on its own, exactly as ID-35 rules.

Concur with ID-35 and with B's recommendation 1: the one line TextureEmit.h:98 is the whole switch, so moving the flip into D's landing commit keeps every intermediate tree green and keeps a bisect meaningful. If it stays in B's commit, the B→D window is knowingly red on the default arm and the retrace and must be gated at 0x1ff.

D-K2 gains the row B asks for, and the wording is right: bit 10 requires the server-side texture consumer, and it is the only P4a subsystem bit that is not independently switchable, because it is the only one whose emission takes state away from the legacy path rather than adding a record the server may ignore. (Bits 9, 11 and 12 are each 491/491 alone.)

(b) The six verify-lane unit failures vanish on 92dffb81 — confirmed

build-verify on the replayed tree: 1759 / 1759, zero failures. All six of B's named cases pass:

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

They were the wire × clientsp-v2 seam that C v3's 79b58198 fixed (the forked refusal drive unlinked the log while Initialize() held it open, so the child wrote into a deleted inode) — the empty-child-log shape B described. B's control run was right and its diagnosis was right. Nothing owed here.


6. B's post-integration checklist

  1. Decide the flip's home first (§5a). Recommendation: move TextureEmit.h:98 into D's landing commit. If it stays with B, gate the B→D window's default arm and the retrace at 0x1ff and say so in the log.
  2. Skip 549ef30b and use §4's cherry-pick list; resolve the one test-file conflict theirs-first per §4.2 — ours-first breaks 4ec43194.
  3. After the merge, re-run what this round could not: G1 (symbol_report --threshold 0 vs ~/w7/p4a-before-libMobileGL.so), G5 (scripts/p3a_untouched_regions.sh 37da3c3a HEAD), gen_pipe.py --check, gen_pipe_dirty_surface.py --check, and check_include_closure.py --mode both --compiler clang++ --self-test --require-all (B's new TextureEmit.h → SamplerEmit.h and GL_Framebuffer.cpp → FramebufferEmit.h edges). G2 is already verified on the replayed tree at 0 diff lines.
  4. After D lands, and these are the checks B's work is actually gated on:
    • the default arm and the retrace back to 491/491 and 79/79;
    • the residual-record sweep (ID-12): the 183 "no applier record" refusals to zero, counting all eight families per ID-34's N-2 — a residual one on a texture or renderbuffer handle is a B seam defect;
    • PipeStats::ClientTextureUploadEmissions against the server's TextureUpload{Emissions,Box,Rect,Jobs} on the 79-case retrace: a rect count of zero on rei/xaero-* means M1's tail is still not arriving, and it is the only place the region list is exercised at volume;
    • C's OverCapacityMints and referenced-eviction counters, for n3.
  5. integration-verify on the DirectVulkan lane under GLIBC_TUNABLES=glibc.malloc.tcache_count=0 for ID-8's per-kind leak test — still owed; this tree's integration-verify label runs DirectGLES.
  6. Package F's 0x5ff / 0x9ff dependency-refusal arms cannot be read off any tree before D (0x5ff is 444/491 for §5a's reason, not for D-K2's).
  7. Optional, and one line each: n1's EmitOneLevel guard, n4's comment. n2 goes to wire's next round.

Reviewer note. Replay worktree /home/swung/w7/p4a-rereview-clientfb on branch p4a/bint-review, its three build directories and all scratch logs were removed; git worktree prune run and the branch deleted. No device work. ~/w7/pipe and ~/w7/p4a-clientfb were read only.