Files
MobileGL/docs/archive/notes/p4a/p4a-results/clientfb-v1.md
T

28 KiB
Raw Blame History

P4a package B — clientfb, the framebuffer and texture/renderbuffer client. Result, v1

Branch refs/heads/p4a/clientfb, three commits on top of the tag p4a/contract (08192d72). Working tree clean. Not pushed. No other worktree touched.

step hash message
b1 51d1760f [Feat] (Pipe, State): mint a {slot, gen} handle for every texture and renderbuffer and publish its create, respecify, parameters and accumulated sub-data as pipe calls
b2 a78eaf93 [Feat] (Pipe): push the bound framebuffers with a resolved read surface, inline attachment formats and a content hash that covers the draw-buffer array
b3 f7de87b5 [Feat, Test] (Pipe): wire the framebuffer subsystem bit and pin the resolved read surface, the sticky bind mask, the level-shadow strides and the union-box/region-list invariant

b3's message is not C.1's wording, because C.1's wording would have been false: only ONE of the two wired-subsystem constants is flipped. §5 (D4) is the whole reason and it is the single most important thing in this document for the integrator.

13 files, +2708 / −56 against the tag.


1. Verdict on the hard rules

rule result
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 renamed / 0 resized, .text 10806323 → 10806323 (+0, +0.000%), 27811 → 27811 defined symbols. Run after every commit. P4a's admitted-resize set stays EMPTY.
gen_pipe.py --check / --self-test rc 0 / rc 0
gen_pipe_dirty_surface.py --check / --self-test rc 0 / rc 0 (the two contract-marked UNDECIDED rows unchanged)
check_include_closure.py --mode both --compiler clang++ --self-test --require-all rc 0, 4 probes, 0 skipped, 0 problems, 6 controls tripped
git diff --exit-code -- MobileGL/MG_Pipe/generated rc 0
p3a_untouched_regions.sh 37da3c3a HEAD rc 0, the eleven byte-identical
three builds (pull / push / verify) all three, rc 0
ctest -L unit ×3 1665 / 1665 in each, 0 failed
ctest names pull == push diff empty
no name removed vs ~/w7/p4a-before-ctest-names.txt 0 removed, 43 added (§7)
ctest -L integration-gpu -j 8, default mask 966 / 966
same, MOBILEGL_PIPE_PUSH=0 966 / 966
same, MOBILEGL_PIPE_PUSH=0x1ff (P4a off, G12's T2 arm) 966 / 966
ctest -L integration-verify -j 4 under GLIBC_TUNABLES=glibc.malloc.tcache_count=0 844 / 844, 0 Fatal{ lines
MOBILEGL_PIPE_VERIFY=1 retrace_gate.py … -j 4 79 / 79; grep -l 'Fatal{' empty; grep -L 'MGPipe verify:' empty
only B's files (+ granted sites) yes, §6
no push, no attribution lines yes

2. Per-file summary

MG_Impl/Pipe/TextureEmit.h (+1024, B owns it)

The whole texture and renderbuffer client. Pure builders first, so every case is one EXPECT per field and G7's scripted control can name the field it dropped:

  • MGPipeTextureStorageKindForTarget — StorageKind derived from the TARGET rather than from GetStorageType(), because resource_create is emitted from TextureObjectBase's constructor where the derived object does not exist and that accessor is pure virtual. Calling it there is undefined behaviour, not a wrong answer; the target mapping is exact.
  • MGPipePackSubDataTarget / …ResourceTargetOf / …UploadTargetOf — the low byte is MGPipeResourceTarget (so a buffer record still compares == kMGPipeResourceTargetBuffer whole, unchanged from P3a) and the high byte is TextureUploadTarget, which is the only place a cube face can ride. §5 (D2) records it as a deviation.
  • MGPipeTextureExtentOf — the extent trio plus the layer count, which the frontend answers in three different axes by target (1D array in y, other layered targets in z, cube = six blobs).
  • MGPipeBuildTextureResourceDesc, MGPipeBuildRenderbufferResourceDesc, MGPipeBuildTextureParams, MGPipeLevelPitchOf, MGPipeBoxOfDirtyRegion, MGPipeBuildSubRegion.
  • MGPipeTextureEmitter — the handles and their inverse, the sticky BindMask with the ImageBindableHint transition arming ForceResync exactly once, the publication latch, the descriptor memcmp dedupe, the drain list, and DrainTextureSubData with the empty-list early-out.
  • Six MG_State entry points and the two step-1 death emitters (§5, D3).

MG_Impl/Pipe/FramebufferEmit.h (+425, B owns it)

MGPipeBuildSurface (one statement per field), MGPipeResolveAttachmentUploadTarget, MGPipeDrawBufferIndex, MGPipeFramebufferStateContentHash (field-wise into a zero-initialised staging copy, ContentHash left 0), and MGPipeFramebufferEmitter::EmitFramebufferState — the per-target discrimination, the Target = Both collapse, the suppressor slot fed the combined hash with per-target latches deciding which record goes out, Complete from CheckCompleteness(), and the client half of D-C3's over-wide-attachment refusal.

MG_State/GLState/TextureState/TextureObject.{h,cpp} (+34 / +142)

Three non-virtual, push-only protected helpers on TextureObjectBase (PipePublishDescriptor, PipePublishParams, PipeNoteLevelDirty) declared in the header and defined in the .cpp, which is the ONE MG_State translation unit that sees the client emitter (§5, D1). ~TextureObjectBase becomes the three-step death path. Hooks: the constructor (mint + create), SetInternalFormat, the three border-colour setters, both swizzle setters, SetBaseLevel, SetMaxLevel, SetImmutableLevels, SetSamples, SetFixedSampleLocations, SetDepthStencilTextureMode (header, inline), and — after the storage mutation, never before BumpShapeVersion — TextureObjectWithOneMipmap::{AllocateStorage, TruncateMipmapLevels, MarkStorageDirty, MarkStorageDirtyRegion}.

TextureObject2DCube.cpp (+17), TextureObjectView.cpp (+12), TextureObjectBuffer.h (+9)

The cube's four storage overrides call the inherited helpers; the view's constructor publishes ViewOf explicitly (because SetImmutableLevels early-returns on a degenerate view); the buffer texture's SetBufferRange publishes the BufferForTexBuffer / BufOffset / BufSize trio at the first statement at which it is complete.

MG_State/GLState/RenderbufferState/RenderbufferObject.{h,cpp} (+7 / +61)

One push-only private PipePublishDescriptor, the constructor's mint + create, the three storage setters (D-D2's hole closed by EMISSION, with no new member and no widened shutter), and the three-step death path.

MG_State/GLState/FramebufferState/FramebufferObject.cpp (+25)

~FramebufferObject now calls MGPipeEmitFramebufferDestroyAndFree — steps 2 and 3 only, because D-I2 gives this kind a handle and no wire lifetime.

MG_Test/Pipe/{TextureEmitTest, FramebufferEmitTest}.cpp (+484 / +367)

14 + 8 emitter-side cases, appended before main() and disjoint from the wire package's applier-side cases. All 24 registered names exist in a pull build as visible skips.

MG_Test/State/ObjectLifetimeIdTest.cpp (+69, granted)

ProbeLifetimeIdAcrossAddressReuse extended to TextureObject2D, FramebufferObject, SamplerObject and ProgramObject (8 new cases, all green). The texture case doubles as a 64-round mint/free churn check under push.


3. EmittedCallSuppliesTheWholeField, per field

PipeFill.cpp is the contract package's for the whole phase and was not touched; the six P4a rows already sit in its false arm. This package's job was to confirm each answer against what it actually emits, and it concurs with all six. Only two of them are this package's families:

field emitted by supplies the whole field? why
GetFramebufferBindingSlot SetFramebufferState (this package) NO The field's storage is BindingSlot<FramebufferObject> — a frontend heap reference — and the call carries an eight-byte {slot, gen} Fbo plus a fully resolved descriptor. The applier has no way to produce the pointer and P4a deliberately does not give it one (a payload never contains a pointer). Skipping the pull would leave the mirror null on every draw of every push build. What retires it is the phase where the backend stops reading a frontend FramebufferObject, not a better applier. This is P3a's GetBoundVertexArray precedent, one for one.
(none) the texture-resource family n/a kMGPipeSubsystemTextureResources names NO emitted field and cannot: resource_create/_respecify/_subdata/_destroy and set_texture_params are dispatched at the GL call that causes them rather than filled into a PipeInputs field. PipeFill.cpp asserts this (NoDirtyBitOwnsTheTextureResourceSubsystem), and no Coverage.def row exists or may exist.

The other four (GetImageTextureBinding, GetTextureUnitObject, GetProgramForDraw, GetProgramForDispatch, plus GetMaxTouchedTextureUnit) belong to package C and are untouched.

Consequence, and it is the one that matters for the verify lane: this package retires no pull at all. Every field it could have affected keeps coming through the residual fill, which is why the 79-case verify retrace reports zero divergences rather than "no divergence we looked for".


4. The wired-bit commit

b3 (f7de87b5) sets kMGPipeWiredFramebufferSubsystem = kMGPipeSubsystemFramebuffer in MG_Impl/Pipe/FramebufferEmit.h. PipeFill.cpp's OR and its four static_asserts were not touched; the assert that the constant is either 0 or its own bit still holds.

kMGPipeWiredTextureSubsystem stays 0. §5 (D4) is the full argument and the one-line edit the integrator makes after the rebase.


5. Deviations

D1 — MG_State reaches the client emitter through TWO translation units, not zero. P3a's buffer family declares its emission points in MG_Pipe/PipeMutation.h and defines them in MG_Impl/Pipe/PipeFill.cpp, so BufferObject.cpp sees a declaration and never the client's tracker. P4a cannot copy that: both of those files belong to the contract package for the whole phase (C.7 gives them to A, and the "no file touched twice" rule is what discharges the bb2a236d merge trap structurally), and neither carries a texture emission declaration. The next-best arrangement keeps the same property one level in:

  • every texture emission point is a protected, non-virtual, push-only member of TextureObjectBase declared in TextureState/TextureObject.h and defined in TextureState/TextureObject.cpp — the one MG_State TU that includes MG_Impl/Pipe/TextureEmit.h. The cube's, the view's and the buffer texture's TUs call the inherited helper and still see only a declaration;
  • the renderbuffer half has no base class to hang helpers on and exactly one .cpp, so RenderbufferState/RenderbufferObject.cpp is the second and last such TU.

So the client's tracker reaches exactly two MG_State translation units instead of zero. check_include_closure.py's mutation-header probe is unaffected (it asserts MG_Pipe/PipeMutation.h's own closure, which is unchanged) and is green. The integrator can fold the six free functions into PipeMutation.h and PipeFill.cpp in one mechanical commit once the phase's file ownership relaxes; nothing else has to move.

D2 — MGPSubData::Target is PACKED for a texture: low byte the resource target, high byte the upload target. D-D3's table asks Target to carry "the owner's upload target (MGPipeResourceTarget from D-A3, plus the cube-face upload target in MGPSurface-style terms)". The payload has one Uint16 for two facts the server needs and no other field can hold either: which kind of storage the destination is (the applier branches on it — a buffer target dispatches into MGPipeResourceOps, every other target stores and returns) and which cube face the level belongs to, which is not derivable from the resource target. The buffer half is preserved exactly: a buffer record's upload byte is 0, so Target == kMGPipeResourceTargetBuffer still holds for the whole field and w1's branch is right whether it tests the whole Uint16 or only the low byte. A static_assert pins that. If A would rather own the encoding, moving these three constexpr helpers to MGPipeTypes.h beside MGPSubData is a copy-paste.

D3 — step 1 of the death order is emitted from the destructor, one statement before the contract's helper. MG_Impl/Pipe/PipeFill.cpp's MGPipeEmitTextureDestroyAndFree and …RenderbufferDestroyAndFree hard-code published = false at the tag, with the note "the texture emitter publishes nothing yet". That file is A's for the whole phase, so this package supplies the answer from the object's destructor instead:

~TextureObjectBase / ~RenderbufferObject
  -> MGPipeEmit<Kind>ResourceDestroy(lifetimeId)   // 1. the wire delete, published-gated
  -> MGPipeEmit<Kind>DestroyAndFree(lifetimeId)    // 2. the notice, 3. the slot

which is exactly PipeMutation.h's fixed three-step order and not a variation on it. Both new functions are published-gated rather than slot-gated, for the reason PipeFill.cpp states in full. The integrator can fold them into the helpers' bodies when D1 is folded.

D4 — kMGPipeWiredTextureSubsystem is still 0, and that is a BLOCKED flip rather than an unfinished one. THIS IS THE ONE THING TO READ. Unlike the other three P4a families, the texture family gets no fresh apply entry points: the catalogue is closed and a texture rides P3a's OWN resource_create / resource_respecify / resource_subdata / resource_destroy rows. On a base without w1 those four have P3a's buffer bodies, and two of their properties make a texture record actively harmful rather than merely ignored:

  • MGPipeApplierState::Resources is ONE vector indexed by SLOT (D-B2 makes it three, one per resource kind). Buffer, Texture and Renderbuffer slot spaces are independent, so a texture create at slot 12 overwrites the buffer record at slot 12 and the next write to that buffer is refused against the texture's extent — a dropped content write with no diagnostic beyond the refusal counter;
  • SubDataBoxFault validates every record as the buffer half of MGPSubData, so a texture sub-data record is Fatal{ProtocolCorruption} on record.Level != 0 alone.

This was found, not argued: the first verify-build unit run with the bit flipped went red with Fatal{ProtocolCorruption} resource_subdata {slot=12, gen=0, glName=8}: the buffer half carries a mip level, and it took TextureTest.GetTexImageReadsALevelWhoseLowerLevelsWereNeverDefined — a pre-existing case, nothing to do with P4a — down with it. That is the applier's trip wire doing exactly its job, and it is evidence that D.1's stated order (wire before clientfb) is not merely tidy.

The package is therefore complete and gated but not switched on:

  • kMGPipeWiredTextureSubsystem = 0, so kMGPipeWiredSubsystems does not carry the bit, the validate point never calls DrainTextureSubData, the drain list never grows, and nothing this file builds reaches an applier that cannot hold it;
  • MGPipeTextureRecordsReachTheApplier() is that constant asked as a constexpr predicate and guards the four MGPipeApplyResource* calls with if constexpr. set_texture_params is deliberately NOT behind it — MGPipeApplySetTextureParams is one of P4a's own fifteen entry points and is a stub at the tag, so sending it keeps that seam exercised;
  • MGPipeTextureEmitter::ArmForTest(Bool) (beside ResetForTest()) is initialised from the constant and is what the two suites flip, so the whole conversion — descriptors, params, the sticky mask, the drain list, the sub-data record and its strides — is gated by tests that assert on the emitted records rather than on any applier state. The flip cannot land untested.

The integrator's edit after the rebase onto p4a/wire is one line: change that 0 to kMGPipeSubsystemTextureResources. §8 lists what to re-run.

D5 — the drain list's membership flag lives on the CLIENT, not on MipmapStorage. D-D4 asks for "a Bool m_onDrainList beside the level's existing m_isDirty, inside #if MOBILEGL_PIPE_PUSH". It is a per-slot Vector<Uint32> of packed (uploadTarget, level) keys in the emitter instead, for D-D5's own inversion: the emission cursor is client state and MG_State gains no member at all this way (D-D2 makes the same argument for the renderbuffer). The cost is a linear scan of ≤ 96 shorts on the glTexSubImage path, which has just memcpy'd texels. TextureEmit.ABailedLevelStaysDirtyAndStaysOnTheDrainList pins the behaviour either way.

D6 — the descriptor is deduped on its own bytes rather than on a version. glTexStorage2D allocates level by level and every parameter setter that touches a descriptor field publishes; an 88-byte memcmp against the last emitted descriptor is cheaper than the emission it avoids and is what keeps glRenderbufferStorage's three-setter sequence one record rather than three (TextureEmit.ARenderbufferRespecifyPublishesItsExtentWithoutAVersionCounter pins that). It is the "version-first skip before anything expensive" shape, with the descriptor standing in for a version the frontend does not have.

D7 — MGPSurface::Kind reuses MGPipeKind rather than minting a three-value enum. MGPipeKind already spells Texture and Renderbuffer, its None is 0, and a zero-initialised MGPSurface is therefore already the empty attachment point the contract's own sentence requires. A static_assert pins None == 0. If A wants a dedicated enumeration beside MGPSurface it is a rename, not a re-encoding.

D8 — MGPFramebufferState::DrawBuffers[] indexes THIS RECORD'S Color[], and the four default-framebuffer tokens narrow to 0. None → −1 and Color0..Color31 → 0..31 are the field's documented convention read literally. A default framebuffer keeps its one colour surface under BackLeft, this record carries it in Color[0], and IsDefault is what tells the server which framebuffer it is looking at. The distinction the narrowing loses is FRONT vs BACK and LEFT vs RIGHT, which MobileGL's frontend never gives a default framebuffer — FramebufferObject's constructor seeds BackLeft and nothing writes another. A phase that needs stereo has to widen the field, not re-encode this one.

D9 — MGPTextureParams::DepthStencilMode is numbered, not a GLenum. The frontend keeps GL_DEPTH_COMPONENT / GL_STENCIL_INDEX (0x1902 / 0x1901) and the payload byte cannot hold one. 0 is DEPTH_COMPONENT, which is also GL's initial value and therefore what a zero-initialised record already says.

D10 — MGPTextureParams::BuiltinSampler is IDENTITY-addressed on this branch. It is MGPipeSlots().Acquire(MGPipeKind::SamplerCso, texture->GetSamplerObject()->GetLifetimeId()) — the same key ~SamplerObject's death helper already resolves through, so the two agree by construction. Minting a handle is client state; the CALL that fills the record (create_sampler_state) is package C's. The cross-package seam to check at integration is whether C's content-addressed cache (capacity 256) keys its slots on the SamplerObject's lifetime id. If it does, nothing changes. If it mints on a synthetic id, this line has to become a call into C's cache, and TextureEmit.TwoTexturesWithIdenticalSamplingShareOneBuiltinCso — which SKIPs today with that reason printed — is its gate.

D11 — one SamplerCso slot per texture is not released on this branch, by construction. ~SamplerObject still raises the bare NotifyStateObjectDestroyed at the tag; D-I1's table gives package C the edit that turns it into MGPipeEmitSamplerCsoDestroyAndFree. Until clientsp lands, every texture that ever published its params leaves one SamplerCso slot behind. It is an ordered dependency, not a leak in this package: clientfb → clientsp is D.1's order, and TextureEmit.ADestroyedTextureReleasesItsResourceViewAndBuiltinSamplerSlots says so in the case body rather than asserting something it cannot yet prove.


6. Granted sites

Exactly one grant was taken.

file:line what
MobileGL/MG_Test/State/ObjectLifetimeIdTest.cpp:34-40 the five new includes
MobileGL/MG_Test/State/ObjectLifetimeIdTest.cpp:161-227 eight cases extending ProbeLifetimeIdAcrossAddressReuse / ExpectDistinctIdsWhileBothAlive to TextureObject2D, FramebufferObject, SamplerObject and ProgramObject

Grants NOT taken, and why:

  • MG_Impl/GLImpl/Texture/GL_Texture.cpp — not needed, including the one anticipated exception. The generated-mip storage grow at :528-539 calls AllocateStorage + MarkStorageDirty(false) + TruncateMipmapLevels + BumpContentVersion(), and the first and third of those are hooked on the OBJECT, so the respecify goes out from there. The same is true of both glTexBuffer entry points: they end in SetBufferRange + SetInternalFormat, both of which publish. No storage path was found that bypasses TextureObjectBase / TextureObjectMipmap.
  • MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.cpp — not needed; AllocateRenderbufferStorage_State goes through RenderbufferObject's own three setters.
  • MG_Test/Framebuffer/FramebufferTest.cpp, MG_Test/Texture/{TextureTest, TextureViewTest}.cpp — no payload edit broke a case (all three suites are green in all three builds), so nothing was appended and ID-8's collision warning does not apply.
  • MG_Test/Buffer/BufferTest.cpp — package D's read-and-confirm, not B's.

7. What was added, and what was not touched

43 ctest names added, 0 removed. 35 are the contract commit's; the 8 this package adds are the ObjectLifetimeIdTest extension. The 22 emitter-side cases in TextureEmit / FramebufferEmit land inside names the contract commit already registered.

Untouched, confirmed: everything under MG_Backend/, MG_IntegrationTest/, .github/, MG_State/GLState/{SamplerState, ProgramState}/**, MG_State/GLState/Core.{h,cpp}, every MG_Pipe/ file, every scripts/ file, every generated/*.inc, and MG_Impl/Pipe/{PipeFill.*, Tracker.h, SetHashSuppressor.h, SlotAllocator.*, CsoCache.h, ResourceTracker.h, VertexInputEmit.h, SamplerEmit.h, ImageEmit.h, ProgramEmit.h}.


8. Verification transcript

$ git log --format='%h %s' 08192d72..HEAD
f7de87b5 [Feat, Test] (Pipe): wire the framebuffer subsystem bit and pin the resolved read surface, …
a78eaf93 [Feat] (Pipe): push the bound framebuffers with a resolved read surface, …
51d1760f [Feat] (Pipe, State): mint a {slot, gen} handle for every texture and renderbuffer …

$ python3 scripts/symbol_report.py --before ~/w7/p4a-before-libMobileGL.so \
      --after build-linux/libMobileGL.so --threshold 0 --fail-on-symbol-set-change --fail-on-added-bytes 0
symbol-report: .text 10806323 -> 10806323 (+0, +0.000%)
symbol-report: 27811 -> 27811 defined symbols: 0 added, 0 removed, 0 resized, 0 renamed   rc=0

$ python3 scripts/gen_pipe.py --check                        rc=0
$ python3 scripts/gen_pipe.py --self-test                    rc=0
$ python3 scripts/gen_pipe_dirty_surface.py --check          rc=0
$ python3 scripts/gen_pipe_dirty_surface.py --self-test      rc=0
$ python3 scripts/check_include_closure.py --mode both --compiler clang++ --self-test --require-all
include-closure: 4 probes, 0 skipped, 0 problem(s), 6 negative-control trip(s)             rc=0
$ git diff --exit-code -- MobileGL/MG_Pipe/generated         rc=0
$ bash scripts/p3a_untouched_regions.sh 37da3c3a HEAD        rc=0  (the eleven byte-identical)

$ cmake --build build-linux  -j 12   rc=0
$ cmake --build build-push   -j 12   rc=0
$ cmake --build build-verify -j 12   rc=0
$ ctest --test-dir build-linux  -L unit -j 12   100% 1665/1665
$ ctest --test-dir build-push   -L unit -j 12   100% 1665/1665
$ ctest --test-dir build-verify -L unit -j 12   100% 1665/1665
$ ctest --test-dir build-push -R 'FramebufferEmit\.|TextureEmit\.'   100% 24/24 (1 SKIP, D10)

$ diff <(build-linux names) <(build-push names)                       (empty)
$ comm -23 ~/w7/p4a-before-ctest-names.txt <(build-linux names)        0 removed
$ comm -13 ~/w7/p4a-before-ctest-names.txt <(build-linux names)       43 added

$ ctest --test-dir build-push -L integration-gpu --no-tests=error -j 8
100% tests passed, 0 tests failed out of 966
$ MOBILEGL_PIPE_PUSH=0     ctest --test-dir build-push -L integration-gpu -j 8   966/966
$ MOBILEGL_PIPE_PUSH=0x1ff ctest --test-dir build-push -L integration-gpu -j 8   966/966
$ GLIBC_TUNABLES=glibc.malloc.tcache_count=0 \
      ctest --test-dir build-verify -L integration-verify --no-tests=error -j 4
100% tests passed, 0 tests failed out of 844        Fatal{ lines: 0

$ MOBILEGL_PIPE_VERIFY=1 python3 ~/w7/retrace_gate.py --tree ~/w7/p4a-clientfb \
      --lib ~/w7/p4a-clientfb/build-verify/libMobileGL.so --out ~/w7/retrace-out/p4a-clientfb -j 4
passed 79 / 79; failed: []                                                                 rc=0
$ grep -l 'Fatal{'         ~/w7/retrace-out/p4a-clientfb/*.log     (empty)
$ grep -L 'MGPipe verify:' ~/w7/retrace-out/p4a-clientfb/*.log     (empty)
$ rm -rf ~/w7/retrace-out/p4a-clientfb

A trap worth recording for the integrator, because it cost a whole retrace round. The first retrace reported passed 2 / 79 with ssim=None on every case — the exact false red C.1 warns about. tools/trace_replay/fixtures in this worktree had been pulled back to 131/133-byte LFS pointers, with an mtime matching the "reset to the tag" the six sibling trees were given at 12:24; git update-index --assume-unchanged was still set, so git status was clean and nothing looked wrong. Restoring them from ~/w7/pipe/tools/trace_replay/fixtures (803 MB) and re-running gave 79/79. Every sibling tree reset at the same time is very likely in the same state.

Two other findings, both gates working:

  1. the Fatal{ProtocolCorruption} of §5 (D4), which is why the texture bit is not flipped;
  2. MG_Config::Features.PipePush defaults to 0 and only ConfigLoader sets the phase mask, so a unit case that does not arm it is green for the wrong reason. Both suites arm it explicitly and restore it, and the arming is called out in the scope's comment.

9. What must be re-run after the rebase onto p4a/wire

The integrator's edit is one line — kMGPipeWiredTextureSubsystem = 0 → = kMGPipeSubsystemTextureResources in MG_Impl/Pipe/TextureEmit.h — and it must be made in the same commit as the rebase, because w1 is what makes it safe. Then, in this order:

  1. ctest -L unit in build-verify, first and on its own. It is the cheapest place the two defects of §5 (D4) surface: TextureEmit's four drain cases and TextureTest.GetTexImageReadsALevelWhoseLowerLevelsWereNeverDefined were the exact five that went red before, and they are the named expected-failing set to re-state unchanged.
  2. ctest -L integration-gpu on build-push at the default mask AND at 0x1ff. The default arm is the first place a texture record meets a real applier record table; the 0x1ff arm proves the A/B still isolates it.
  3. ctest -L integration-verify under GLIBC_TUNABLES=glibc.malloc.tcache_count=0, for the two new death paths (~TextureObjectBase, ~RenderbufferObject) reaching a live applier.
  4. The 79-case verify retrace, which is the only thing that exercises the drain list on real texture traffic (rei, xaero-*, journeymap, modernui are the atlas fixtures D-D5 names). Check the fixtures are not LFS pointers first.
  5. G1, because the flip changes a constexpr in a push-only header and must still move nothing in the pull build.

Also re-check at integration, in this order of risk:

  • §5 (D10) — whether package C's sampler CSO cache keys its slots on the SamplerObject's lifetime id. If not, MGPipeTextureEmitter::EmitTextureParams's one Acquire becomes a call into C's cache and TwoTexturesWithIdenticalSamplingShareOneBuiltinCso stops skipping.
  • §5 (D11) — that clientsp really does replace ~SamplerObject's bare notice with the death helper, or every texture keeps leaking one SamplerCso slot into G8b.
  • §5 (D2) — that w1's Desc.Target branch reads MGPSubData::Target in a way the packing satisfies (it does if it tests either the whole Uint16 or the low byte against kMGPipeResourceTargetBuffer).
  • §5 (D1/D3) — whether to fold the six MG_State entry points and the two step-1 death emitters back into PipeMutation.h / PipeFill.cpp now that A's ownership window has closed.