Files
MobileGL/docs/Disaggregated/notes/p4a/p4a-results/esprytdraw-v3.md
T
swung0x48 aba329cc53 [Docs] (Disaggregated): P6's plan and contract draft; the working notes move from a local directory into the repository
Two things land together: the plan and contract draft P6 needs before it starts, and the
working notes that until now existed on exactly one machine.

1. P6's plan and contract draft (neither started; not one line of code)

     docs/Disaggregated/P6-SPAWN-PLAN.md      the package plan
     docs/Disaggregated/P6-CONTRACT-DRAFT.md  the contract draft (moves to
                                              MG_Remote/CONTRACT-P6.md when c6 lands)

   The three findings that mattered, all written into the documents:

   a. P6 is SMALLER than its roadmap cell suggests. The transport primitives are already in
      the tree and already tested: SocketDoorbell including hangup detection and the two
      lost-wakeup fences (Doorbell.cpp:211-360, four cases in FdPassingTest), SCM_RIGHTS fd
      passing, ShmSegment::Adopt, and the SurfaceOp / WindowKind schema already defined in
      protocol.fbs. inproc uses a real second shm mapping rather than an alias, and
      SessionRings.h:22-28 says that was done so today is not the first time it runs. P6 is
      assembly plus a control plane plus a process, not "write a socket transport".

   b. The evidence for "P6 is only a transport swap" has EXPIRED. It was bought by P5c, whose
      audit found 59 direct memory accesses in a tree that already looked clean. P5e changed
      the tree again. So the first package, a6, is a read-only audit that writes no code, and
      every contract row that depends on it is marked PENDING a6 rather than decided in
      advance.

   c. P6 does not get a game running on the phone. ANativeWindow* is meaningless in the child,
      and protocol.fbs:186-196's own comment says Android transfers the window out of band -
      that band is P12. P6 lands pbuffer / surfaceless / offscreen, which is what HeadlessGL,
      the split-equivalent lane and trace replay need, and which is exactly what P6's own exit
      gate measures. A real window reaching the server is a named refusal. This gets its own
      section because "P6 is done" would otherwise read as "the game runs now".

   The contract adds rule G: a role may not name a process-local handle. It extends P5c's
   rule E (a role may not name the other role's memory) from memory to identity - EGLDisplay /
   EGLSurface / EGLContext / ANativeWindow* / fd number / pid / mapped address either cross as
   a token this contract defines or they do not cross, and a descriptor crosses only through
   SCM_RIGHTS, where the kernel renumbers it.

   What P5e left for P6 gets its own section. The latch semantics are already designed
   (ARCHITECTURE.md:504) and are copied unchanged. What is new is that after run-ahead the
   client no longer waits while the peer process can DIE: the client holds records that were
   published and will never be applied, and Android's freezer suspends the child without
   closing anything. So slow and dead must be distinguishable WITHOUT a timeout threshold.
   That is P6's only new design; everything else is carriage. It shares a predicate with E1's
   outstanding re-specification (ID-122), and the two should be designed together.

   Two gates carried over from P5e's lessons: G2 requires integration-spawn's test-name set to
   match integration-split's name for name, because a differing set means the lane is quietly
   running less; and the arm-proof gate requires every entry to record the child pid and
   transport=spawn in its OWN private log - ID-124 established that this kind of false green
   really happens, and a spawn lane that actually ran monolith would pass every other gate.
   S2's shape must be settled before it is written, and the document says why: E1 is the
   cautionary case.

2. The working notes move into the repository (docs/Disaggregated/notes/, 415 .md files)

   ROADMAP / CURRENT_STAGE_PROGRESS / MEASUREMENTS cite ruling numbers, audit line numbers and
   the provenance of gate figures from ~/w7/notes/, a directory that exists on one machine and
   is under no version control. A record cited hundreds of times that lives only locally is not
   a record, so it moves in whole, with its directory structure intact.

   Only .md moves. Scripts, run logs, census and gate result JSON, patches, APKs and nm dumps
   stay where they are: they are evidence DATA rather than documents - large, regenerated per
   run, and mostly carrying absolute paths and device serials that mean something only on the
   machine that produced them. The mapping and this split are written in
   docs/Disaggregated/notes/README.md.

   Two exceptions are recorded there too: P6's BRIEF has been promoted to a first-class
   document and keeps no copy under notes/, so the two cannot drift apart; and recovered/ holds
   material salvaged from failed workflow runs, which may duplicate the like-named phase
   directories and is kept because a few of those were the only complete version at the time.

   CI's citation lint takes docs/Disaggregated/*.md only (non-recursive), and notes/ is
   deliberately outside it: those notes' file:line citations were written against their own
   code heads, and resolving them against today's HEAD would fail wholesale - which is not an
   error, it is their correct state as a historical record. To check one, use the revision it
   declares (--rev).

3. Inconsistencies fixed in passing

   - README.md's status line still said P5e was in progress while the stage table already said
     closed.
   - CURRENT_STAGE_PROGRESS's next-step item 1 still listed the exit gate as pending, though
     (a) and (b) were run in the previous commit; rewritten to closed plus the one item that
     actually remains (E1).
   - ROADMAP's P6 row still had the pre-phase plan in three of its cells; it now carries the
     package order, the scope boundary and the two new gates.

Checked with check_doc_citations.py: 125 citations across docs/Disaggregated/*.md, 0 problems.
2026-09-18 23:31:39 -04:00

45 KiB
Raw Blame History

P4a package E — esprytdraw. Verification round, v3

Branch refs/heads/p4a/esprytdraw, worktree ~/w7/p4a-esprytdraw, HEAD 0cd0d1ec, rebased onto refs/heads/feat/disaggregated = ee31944b. One file, MobileGL/MG_Backend/DirectGLES/DirectGLES.cpp. Nothing was pushed.

This is the round that runs on the first tree where every record this package reads is REAL (contract c0..c0e, wire v3's applier, clientsp v3, clientfb v2, esprytobj v2). It started on 17396216; the integrator moved the pipe to ee31944b mid-round (29d51ab9, which closes SD-2 below), so every number in this report was re-taken on ee31944b and every red one has a control run of ~/w7/pipe at the same commit, on the same box, in the same session.

The verdict, in three lines.

  1. The arms engage and the flips are taken. Every decline site that can now speak, does; the eight-family refusal census is 0 on seven of the ten shapes (§6.1), and the three residuals are seam defects against B/C/D with named evidence and patches.
  2. On both itest lanes, E changes nothing. Default arm 483/491 with E and 483/491 without, failing sets byte-identical; 0x7ff 438/491 both; DirectVulkan 475/475 both; integration-verify 842/844 both, zero Fatal{.
  3. On the retrace, E found one real regression of its own - in e3, from v2, not from this round's flips - and this round root-caused it, patched it and verified the patch. create-indirect.DirectGLES was FAIL ssim=0.887403 against the control's PASS 0.999961; the cause is a client-side dirty shutter that cannot see an image re-bind (Tracker.h:518), the four-line fix is in §SD-0, and with it applied E's retrace is 62/79 with a failing set identical to the control's. The patch is in another package's file and is therefore reported, not committed.

Two findings are larger than this package and are handed up: the tree's own 62/79 at the default retrace mask, seventeen shader-pack traces wrong with no refusal behind them (§7), and the sampler-record residual that costs two scenarios a correct picture (§SD-1).


1. The rebase, and the commit series

git rebase refs/heads/feat/disaggregated replayed all ten v2 commits clean - no conflict, one file - both onto 17396216 and again onto ee31944b. D's replay touched none of the lines this package touches. Four new commits on top:

commit what
ea3d664c A4 + the ID-19 accessor. The four temporary family latches (EsprytDraw*HandlesEnabled) deleted; all 12 consults now call D's public wrappers FramebufferSubsystemEnabled() / TextureResourceSubsystemEnabled() / SamplerSubsystemEnabled() / ProgramSubsystemEnabled(). BoundFramebufferRecord takes wire v3's landed shape.
90f0c35a A1 + ID-27 + F2/F4/F5/F7. The framebuffer twin resolved by handle; the Both test replaced by the applier's bound handles; the undescribed-binding declines made loud.
3d6fcc4c The remaining flips - T2, I2, S2, P2, P3, P4, P6 - and the M/K sites' comments updated to record that the round left them silent deliberately.
0cd0d1ec The one flip the first real-path run corrected: T2 and S2 scoped to a draw that actually touches a texture unit.

The final diff against pipe is 1371 +/38 -, one file.

The git lfs checkout trap, recorded because it cost two runs

The task said "git lfs checkout first". On this box that is wrong and destructive: LFS content is not local (Skipped checkout … content not local), so the checkout replaced every real trace fixture with its 130-byte pointer and the first retrace came back 1/79 with 0s ssim=None on 78 of them. The working recipe is wsl_tree.sh's: copy from ~/w7/pipe/tools/trace_replay/fixtures/ and re-apply git update-index --assume-unchanged. And git reset --hard undoes it again - --assume-unchanged does not survive one - which is how it happened a second time while the commit series was being rebuilt. Anyone re-running a retrace on a package tree must re-copy the fixtures immediately before it.


2. ID-27 — the Both question is now asked of the applier's own fact

DirectGLES.cpp:2983 (v2) tested record.Target == MGPipeFramebufferTarget::Both to decide whether one framebuffer is bound to both bindings and the DRAW pass's attachment work need not be repeated. Since ID-19(c) the stored Target is the last emission's target, and B v2 emits Named records at sixteen DSA sites, so:

  • a framebuffer genuinely bound to both bindings can carry Target == Named, and
  • a framebuffer bound to neither can carry a stale Target == Both.

Replaced by BoundFramebuffer[Draw] == BoundFramebuffer[Read] - the applier's own fact, a {slot, gen} comparison, with both handles already proved non-null by F2 three lines up. The old test degraded to a redundant read-buffer sync rather than to a wrong picture (the != Both path takes the full SyncToBackend, which is correct and merely repeats work), which is exactly why no lane caught it and the wire review had to read the emitter. grep 'record.Target' DirectGLES.cpp is now empty, and so is grep '.Target =='.


3. The seven assumptions, reconciled against D's and wire's ACTUAL landed code

# what the integrated tree actually says what E did
A1 BackendPtr* GetOrCreateByHandle(MGPipeHandle) at Managers.h:453 returns a pointer and declines three ways (legacy arm, slot past the sanity bound, generation behind the live entry's), silently, "for its release-build voice at the per-kind resolver". The twin is now resolved by handle only - the FindByHandle-else-GetOrCreate(currentFBO) pair is gone, so no twin on this arm can be minted against the bound frontend address. The null return is a loud whole-arm decline naming the handle and LiveGenAt(slot).
A1, two corrections to the review's sketch slot.GetBoundObject() stays: A2's #else never fired, so SyncToBackend(SharedPtr<FramebufferObject>, target) still takes the frontend object as its argument. FramebufferRecordMatchesBinding (F3) stays: §2's "keep verbatim; it is the wording §8.2 greps" governs, and with the twin now resolved from the record and configured from the bound object, that identity check is the only thing between a mis-keyed record and one framebuffer's attachments written into another's twin. recorded at the site
A2 Confirmed unchanged: Managers.h:1769/:1774 still take const SharedPtr<FramebufferObject>&. nothing
A3 The feared no-op does not exist, and this is settled by code rather than by a counter. D added m_syncedResourceSerial / m_syncedParamsSerial beside the legacy members and re-keyed the sync functions; IsDrawSyncClean (Managers.h:1565-1583) still compares the FRONTEND's GetTextureParamsVersion() twice, GetContentVersion(), and samplerObject->GetVersion(), and returns false outright when the texture has no sampler object at all. Nothing on the read path advances any of those, so a read-only attachment's first sync is still a miss and g_readFboTextureSyncList still pushes. nothing; D-E3's closure stays where it is
A4 D's four wrappers exist with the predicted names at Managers.h:650-665, holding the static const latch; the resolvers carry the refusals, including D-K2's fourth row (ResolveTextureResourceSubsystemArm, bit 10 requires bit 11, with both mirror pairs said out loud). E's four latches deleted; no dependency direction is restated in this file any more.
A5 Confirmed. nothing
A6 MGPImageView::Access now HAS an encoding - MGPipeEncodeImageAccess in C's ImageEmit.h - but nothing in the contract documents it as read-back-able, and e3 still does not read it. unchanged; still a C documentation item
A7 Confirmed. nothing
A8 (the one that had to be measured) Image half: MAJOR-1's union is PERMANENT, and C's own source says so. ImageEmit.h:63-80 derives the window from "the highest image unit the CURRENT PROGRAM names … unioned with a sticky mark of every unit this emitter has already described" - explicitly not every unit the driver holds. The five I2 hits in §6 are that gap, observed. Sampler half: C's window is [0, GetMaxTouchedTextureUnit()] (SamplerEmit.h:822-825), the same bound E's walk uses, so MAJOR-2's unbind is expected unreachable - but it is capped at kMGPipeMaxTextureUnits and gated by a content-hash suppressor, so it may only become an assertion once C documents the membership in SamplerEmit.h. Until then both stay load-bearing.
A9 D landed InvalidateFramebufferHandleArmMemos inside InvalidateFramebufferBindingCache. closed

The ID-19 accessor question, answered. Wire v3 landed DrawFramebuffer() / ReadFramebuffer() as member functions returning a nullable pointer - i.e. both halves of v2's DEV-18 prediction. Because every read in this file goes through BoundFramebufferRecord, the rebase was that one function plus a null check at each of its four call sites, and nothing else moved.

4. The flip table, as executed

grep -c 'P4a decline-site' DirectGLES.cpp = 30 (29 rows, F8 twice), unchanged.

id class v2 state flip taken this round
F1 M silent none; is FramebufferSubsystemEnabled() now
F2 S half-described loud, both-null silent both-null now LOUD - names both bound handles and StaleFramebufferRecordLookups, so "no record was written" and "the slot's generation moved" are told apart without a third message
F3 S loud kept verbatim (the stem §8.2 greps); only the dereference changed. Deliberately kept against the review's A1 sketch - see §3
F4 S loud, continue g_fboRecordsTrusted = false; return false + the record's handle in the message
F5 — silent MOBILEGL_ASSERT + a kept null check (not deleted: a release build must not deref null if the ordering ever changes)
F6 S loud kept - parity with the pre-handle arm
F7 — silent same treatment as F5
F8 ×2 K silent none; gained wire v3's null check, which is not a decline
F9 — fixed in v2 none
NEW S — A1's third decline: GetOrCreateByHandle returning null is a loud whole-arm decline naming the handle and LiveGenAt(slot)
T1 M silent none; is SamplerSubsystemEnabled() now
T2 S silent loud-once, then decline — scoped to maxTouchedUnit >= 0 (see §6.1a)
I1 M silent none
I2 S silent loud-once at the validate point when the unit holds an image texture - a strictly narrower and more informative condition than the review's "when the program declares images", and the flip that found SD-4
I3 M silent none, and deliberately not folded into I2: A8 says nothing pins the window's membership, so a unit outside it is not evidence the way an empty set is
I4 — silent kept as defence; not an assertion (that bound is the applier's, not this file's latch)
I5 S loud none
I6 M silent none - I2 speaks once from inside the sweep instead, where the unit can be named
I7 — fixed in v2 none
S1 M silent none
S2 S silent loud-once, then the frontend walk — scoped to maxTouchedUnit >= 0
S3 — fixed in v2 none
S4 M silent none
P1 M silent none; is ProgramSubsystemEnabled() now
P2 S silent loud-once, telling the two bands apart - an out-of-range ordinary slot says "no create_shader_state arrived" (client seam), an out-of-range composite index says the band base or the decode is wrong (contract bug)
P3 S silent loud-once, naming dead vs live-but-differently-generationed - ID-8's stale-generation refusal
P4 S silent folded into P2/P3: stays silent because both causes have just spoken with the band and the generation named
P5 S loud none
P6 M→S silent loud-once, and the review's condition is not retested here because the only caller already guarantees it (GetUBOSize() > 0 && backendProgram.HasGlobalUboBlock(), the global-UBO upload)
P7 S loud (Fatal-shaped) verdict confirmed - 0 Fatal{ProtocolCorruption} on the verify lane

4.1 Every arm engages

The default-arm run is not merely green, it is green through the new arms: with F2, F4, T2, S2, I2, P2, P3, P6 all wired to speak on a decline, the census in §6 shows zero hits for every framebuffer, program and sampler-window site over 491 scenarios. A declining arm cannot be silent any more, so "no message" now means "the arm ran".


5. The lanes — every number

All on ~/w7/p4a-esprytdraw at 0cd0d1ec on ee31944b, CCACHE_BASEDIR=/home/swung/w7, -j 8, GLIBC_TUNABLES=glibc.malloc.tcache_count=0. The box was shared (load ~45 on 28 cores) with the gate run and package D's round throughout. Every red number below has a control run on ~/w7/pipe at the same commit, on the same box, in the same session, and every one of them matches.

build-linux (pull)      rc 0      build-push  rc 0      build-verify rc 0
build-nolegacy (MOBILEGL_PIPE_LEGACY_MEMOS=OFF, PUSH=ON)   configure rc 0   build rc 0
   0 warnings in DirectGLES.cpp on any of the four

G1  symbol_report.py --threshold 0, before ~/w7/p4a-before-libMobileGL.so
    .text 10806323 -> 10806323 (+0, +0.000%)
    27811 -> 27811 defined symbols: 0 added, 0 removed, 0 RESIZED, 0 renamed
G2  pull-vs-push ctest name diff                      0 lines
G14 vs the P4a before-set                             0 removed, +141 added, 2729 total
G5  scripts/p3a_untouched_regions.sh 37da3c3a HEAD    rc 0 (P3a's eleven byte-identical)
G5  ~/w7/p4a-gates/scripts/p4a_untouched_regions.sh   rc 0 (the seventeen)
      DepthStencilSamplingReadImpl  822ca3b0ece4e662f9b462945f2697f6b14547d841d926ae94c2b35139f16e53
      ShouldUseCaveatTextureFormat  266710e0fdc19756f31e27dcee536149c99fc9fe583752042135b3a9cde26366
      == the two shas D's report prints; these are the two D-N rows that live in E's file
    RenderStateImpl sha  d8fd1c48716056c53675 == ~/w7/p4a-before-syncrenderstate.sha
    pGLContext in MG_Backend  MG_Backend/MGPipe/PipeInputs.h:1  == pipe's, unchanged
    MGPipeUnmigratedEmulation in DirectGLES.cpp  4
    record.Target / .Target == tests in DirectGLES.cpp   0   (ID-27)

ctest -L unit          build-linux 1763/1763   build-push 1763/1763   build-verify 1763/1763
emitter/CSO suites (build-push)                168/168

5.1 -R DirectGLES -L integration-gpu, six arms, with the control beside each

mask E (0cd0d1ec) pipe without E (ee31944b) verdict
default 0x1fff 483/491, 8 failed 483/491, 8 failed failing sets byte-identical
0x1ff 485/491, 6 failed — the six armless .Handles aborts
0 485/491, 6 failed — same six
0x3ff 485/491, 6 failed — same six
0x7ff 438/491, 53 failed 438/491, 53 failed identical — SD-5, a D-side half-run
0xfff 483/491, 8 failed — same eight as the default
C.4 family regex (default) 495/497, 2 failed — the 2 are SD-1's pair
-R DirectVulkan 475/475 475/475 unchanged, and clean

The eight default-arm failures are six + two:

  • six DirectGLES.HandleRecycle.Handles.HandleRecycleScenario.* (Subprocess aborted) — category (a), F's, expected until F lands. Positively identified this round rather than assumed: their logs carry exactly one line, Fatal{PipeLegacyMemosDisabled, "MOBILEGL_PIPE_PUSH leaves kMGPipeSubsystemSamplers (bit 11) clear (or refuses it) and MOBILEGL_PIPE_LEGACY_MEMOS …"} — the itest's 0x1ff pin meeting D's armless stop.
  • IntegerBorderColorScenario.SamplerParameterIivBorderColourSurvivesToAnIntegerSampler and SampledSetStalenessScenario.AQueuedClearIsMaterialisedWhenASamplerObjectCompletesTheTexture — category (b): a NAMED REFUSAL, seam defect SD-1, §6.2.

5.2 -L integration-verify (build-verify, tcache_count=0)

842/844 passed, 2 failed, 0 Timeout, 0 Subprocess aborted, 0 Not Run
Fatal{  lines: 0
the 2 = DirectGLES.Verify.{IntegerBorderColor…, SampledSetStaleness…}  = SD-1, again

Zero Fatal{ is P7's verdict confirmed and the exit-order lane clean on both backends.

6. The census, and the seam defects

6.0 The census could not be taken the obvious way, and the first attempt was a false ZERO

MOBILEGL_LOG_ENABLE_CONSOLE is 0 (MobileGL/Defines.h:71), so the library's log has exactly one sink: the file named by MOBILEGL_LOG_FILE_PATH. Grepping ctest -V output - which does capture every test's stdout - therefore returns zero of everything, always, and looks exactly like a clean census. Setting one MOBILEGL_LOG_FILE_PATH for the whole run does not work either: Log.cpp:71 opens it with "w", so 491 processes truncate each other and only the last survives.

The recipe that works, and that the integrator and package D should both use. Move build-push/MobileGL/MG_IntegrationTest/MobileGLIntegrationTest aside to ….real and put a four-line shell shim in its place that gives each process its own log file, named after its --gtest_filter so a residual can be traced back to its scenario, and that leaves an already-set MOBILEGL_LOG_FILE_PATH alone so the four scenarios that manage their own (PoisonOmission, PointSizeDemotion, StorageBufferRegrow, PrimitivesGeneratedNoXfb) are unaffected.

One trap inside the trap: any cmake --build after installing the shim RE-LINKS over it, after which a naive re-install sees ….real already present and shims the stale binary. Delete both and rebuild. One census run was thrown away to this.

Positive control, so the zeros below mean something: the same harness at MOBILEGL_PIPE_PUSH=0x9ff on one scenario produces ERROR]: MGPipe: kMGPipeSubsystemSamplers (bit 11) is set but kMGPipeSubsystemTextureResources (bit 10) is clear … - D's dependency refusal, captured.

6.1 The census — all TEN shapes, not eight

ID-34's N-2 corrected the census from six shapes to eight. It is still short by two, and by its own method: grep 'applier record' Managers.cpp cannot see :7584 and :7895, because there the phrase straddles a string-literal line break ("… has no applier " / "record …"). Both are the built-in sampler CSO family, and one of them fires on this tree.

-R DirectGLES -L integration-gpu, 491 scenarios, 477 processes that logged:

# shape (Managers.cpp) default 0x1fff 0x1ff
1 :6310 texture … storage and uploads cannot be driven 0 0
2 :7564 texture … built-in sampler cannot be pushed 0 0
3 :7778 texture … parameters cannot be pushed 0 0
4 :8866 framebuffer … its read buffer cannot be pushed 0 0
5 :9083 framebuffer … cannot be synced from the pushed record 0 0
6 :11925 program … no shader-CSO applier record 0 0
7 :12174 sampler … its parameters cannot be pushed 2 → SD-1 0
8 :12418 renderbuffer … storage cannot be allocated 0 0
9 :7584 built-in sampler CSO … declining the built-in sampler push 3 → SD-3 0
10 :7895 built-in sampler CSO … border colour cannot be pushed 0 0
— E's three seam lines (does not describe the binding it names) 5 → SD-4 0
— Fatal{ProtocolCorruption} 0 0
— Fatal{PipeLegacyMemosDisabled} 6 (= the six armless aborts) 6

0x1ff is zero for every row by construction (all four E families take their M sites), and is recorded because it is the arm the six aborts are pinned to.

6.1a The flip the run corrected — T2 and S2, 333 false positives

Before 0cd0d1ec, T2 fired in 172 of 477 processes and S2 in 161, and every single S2 line said units 0..-1. maxTouchedUnit < 0 is "this draw touches no texture unit at all", for which an empty sampler-view / sampler-state window is not a seam defect but the only correct emission. Both are now scoped to maxTouchedUnit >= 0 and both are 0. This is the one place where the review's table, written before C landed, asked for a rule the real path disproved; the site comments say so and say how it was measured. The MINOR-4 gate tick still counts every decline, loud or not, so nothing became uncountable.


SD-0 — E's own regression, found and root-caused: create-indirect on DirectGLES

This is the one defect the round found in this package's own consequences, and the only difference between E's retrace and the control's.

retrace, default mask, 79 cases
  ~/w7/pipe   at ee31944b, WITHOUT E    62 / 79
  ~/w7/p4a-esprytdraw at 0cd0d1ec       61 / 79
  the single difference:
    MobileGLTraceReplay.minecraft-1.21.1-neoforge-create-indirect-in-world.DirectGLES
      E    FAIL ssim=0.887403      pipe  PASS ssim=0.999961     (deterministic, 4/4 runs each)

Bisected to the commit, by rebuilding at six points across E's fourteen commits (the pre-wire-v3 ones carry a mechanical accessor fix so they compile; it is behaviour-preserving - a null answer becomes a zeroed record, which is what a null Fbo already was):

ee31944b  base                                                   PASS 0.999961
78967961  e1  framebuffer records                                PASS 0.999961
6c63e017  e2  texture sync lists                                 PASS 0.999961
1e1a042f  e3  SHADER IMAGES FROM THE PUSHED SET                  FAIL 0.887403   <-- first bad
9aca1457  e4                                                     FAIL 0.887403
15ab31dc  … through 0cd0d1ec                                     FAIL 0.887403

Bisected to the field, with each of e3's four record-sourced fields put behind its own switch:

drop record->Layered  FAIL      drop record->Level   PASS  <-- this one
drop record->Layer    FAIL      drop record->InternalFormat  FAIL

Root cause, and it is NOT in DirectGLES.cpp. MGPipeDirty::NewShaderImages's shutter is Mix(Mix(textureContent, textureParams), programImages) (Tracker.h:518-519), and ImageEmit.h:20-25 states that as an invariant: "all three FRONTEND counters". None of the three moves when glBindImageTexture re-binds a texture that is already on that unit and changes only level (or layer, layered, format, access). So wants(NewShaderImages) is false, EmitShaderImages never runs, the applier keeps the PREVIOUS bind's Level, and e3 - the first code anywhere to read that field off the record - paints the wrong mip. The set-hash suppressor cannot save it: the emitter is never called, so no hash is ever taken. Flywheel/Create does exactly this, which is why one trace in 79 finds it.

Ruled out along the way (each rebuilt and re-run): trusting the record at the eager glBindImageTexture funnel rather than only at the validate point (FAIL either way, so DEV-14's latch is not implicated); ID-27's bound-handle comparison (FAIL with the old record.Target == Both restored); A1's GetOrCreateByHandle (FAIL with FindByHandle + GetOrCreate(currentFBO) restored).

The patch, written and VERIFIED, for the owner of MG_Impl/Pipe/Tracker.h (C / the contract - not E's file under C.7, so it is not committed on this branch):

// Tracker.h, in Update(), replacing lines 518-519
now[Index(MGPipeDirty::NewShaderImages)] = MGPipeMixShutter(
    MGPipeMixShutter(MGPipeMixShutter(textureContent, textureParams), programImages),
    ctx.GetTextureBindGeneration());

glBindImageTexture already calls NoteTextureUnitTouched (GL_Texture.cpp:6741), so that generation moves on every image bind and nothing new has to be counted - which matters, because ImageEmit.h:69 records that adding a counter to TextureState would resize the pull build's object and G1 forbids it. It is also the shape NewSamplers already uses two lines above, with Tracker.h's own argument for it: "MIXING THE GENERATION IN IS THE FIX RATHER THAN A SECOND GATE ON THE EMITTER". Measured with the patch applied to this tree:

create-indirect.DirectGLES     PASS ssim=0.999961   (was FAIL 0.887403)
DirectGLES integration-gpu     unchanged
DirectVulkan integration-gpu   unchanged
full retrace                   see §7

ImageEmit.h's invariant-2 paragraph must be reworded in the same commit: it currently says the shutter is three frontend counters and gives that as the reason the sweep's gate is sound. It becomes four, and the added one is still a frontend counter, so the property it protects survives - but the sentence as written is now the documentation of a bug.


SD-1 (category (b)) — the sampler family refuses, and it is the tree's only wrong picture on the DirectGLES itest lanes

Refusal, twice, one per scenario:

ERROR]: MGPipe: sampler 1 has no applier record on the handle arm, so its parameters
                cannot be pushed (handle {8, 0})   IntegerBorderColorScenario…IntegerSampler
                                (handle {6, 0})   SampledSetStalenessScenario…CompletesTheTexture

Reading side: Managers.cpp:12174, BackendSamplerObject::SyncToBackend - D's file. It resolves the sampler CSO record for the handle and declines when there is none; it does not fall back to the frontend, by design.

Emitting side: package C (clientsp, SamplerEmit.h). No create_sampler_state has been applied for the sampler object the draw binds. The generation in both handles is 0 while the slots differ (6 and 8): the handle was acquired (the slot exists) but never described.

The failure is visible, not silent, which is what makes it the most important residual: IntegerBorderColorScenario.SamplerParameterIivBorderColourSurvivesToAnIntegerSampler reads back component 0 = 0 instead of 255, 1 = 0 instead of -1, 2 = 0 instead of 7 and 3 = 1065353216 instead of 3 on 64 of 64 texels. 1065353216 is 0x3F800000 - 1.0f reinterpreted as an integer - so the integer border colour was not merely lost, the FLOAT form was applied in its place. That is SamplerParameters::borderColorForm (the field F's G7 negative control targets) not reaching the driver because the record it rides on was never created.

Not E's: identical on ~/w7/pipe without this package, on both the gpu and the verify lane.

Patch, for package C. The refusal is at the first SyncToBackend of a sampler object whose CSO record was never emitted, so the fix is birth-side, not per-draw: every path that acquires a SamplerCso slot for a SamplerObject must emit create_sampler_state before any handle naming it can reach the applier - the "mint-and-create in one step" shape TextureEmit.h:943-949 already has. Two things to check first, in this order:

  1. glSamplerParameterIiv / glSamplerParameterIuiv are the entry points both failing scenarios use. If SamplerEmit.h's parameter hooks cover the f/i forms but not the integer Iiv/Iuiv forms, the object is never marked dirty and never emitted - which fits Gen == 0 exactly.
  2. SampledSetStalenessScenario binds a sampler object to complete an incomplete texture, i.e. the sampler is created and bound with no parameter call at all. If the emission hangs off a parameter hook rather than off the object's birth, that case can never emit.

Both point at one rule: emit on acquire, not on first parameter.


SD-2 (category (c)) — bit 10 cost the DirectVulkan lane 66 tests, and it was fixed under this round

Recorded because this round found it independently and its bisect is worth keeping.

On 17396216 (the tree this round started on), on ~/w7/pipe's own build-push, without E:

MOBILEGL_PIPE_PUSH=0     475/475   =0x3ff  475/475   =0x1000 475/475
                 =0x1ff  475/475   =0x7ff  409/475   =0x11ff 475/475
                 =0x1fff 409/475   =0xfff  409/475

Exactly bit 10 (kMGPipeSubsystemTextureResources), deterministic (DirectVulkan.EmptyScissorScenario.AnExplicitlyEmptyScissorBoxClipsEveryFragment reproduces serially and single-process), and identical with and without E. The integrator's 29d51ab9 ("gate the four P4a families on a backend having registered MGPipeResourceOps - Magma emitted, the applier accepted, and the acceptance-cleared dirty flags left its legacy path nothing to upload") landed while this round was running and closes it: on ee31944b the lane is 475/475 with and without E.

Two things from this round's own investigation are worth keeping in the record, because they would have sent the next person down the same two dead ends:

  • a negative control that refuted the obvious single-line culprit: putting only TextureEmit.h's mipmap->MarkStorageDirty(uploadTarget, level, false) (the sub-data acceptance clear) behind a switch left the lane at 66 failed either way. The clear that mattered was on a different path, which is why 29d51ab9 had to gate the whole family rather than one call;
  • the reading side, for the record: VkTextureManager::UploadDirtyMipLevels (VkTextureManager.cpp:2873, and :1828) skips any level whose IsStorageDirty is false, and no MG_Backend/DirectVulkan file reads MGPipeApplier().TextureResources or PendingUploads - so Magma could only ever lose by the emission, never gain by it.

SD-3 (category (b)) — the built-in sampler CSO family refuses, in the two shapes the census cannot see

ERROR]: MGPipe: texture N's built-in sampler CSO {slot, gen} has no applier record
                - declining the built-in sampler push                            x3
  DepthStencilReadbackMatrixScenario.ReadbackLeavesNoGLStateBehind
  UnboundImageDescriptorScenario.ABufferTextureWithNoAttachedBufferDoesNotLoseTheDraw

Both scenarios pass, so this is a silent fallback rather than a wrong picture - which is exactly why it has to be reported: the record for a texture's MGPTextureParams::BuiltinSampler handle was never applied, and D's twin declines the push instead of using it. Same owner and almost certainly the same root cause as SD-1 (C never emitted create_sampler_state for a sampler CSO whose handle it had already acquired); the difference is only which side holds the handle - here the texture record's built-in sampler rather than a bound SamplerObject.

Owed by D's report (documentation) and C (code). D: add :7584 and :7895 to the census in esprytobj-v3.md §4/§5 and say that the stem no applier record on the handle arm matches only seven of ten. The integrator's residual check must grep four patterns, not two: no applier record on the handle arm, no shader-CSO applier record, declining the built-in sampler push, border colour cannot be pushed.


SD-4 (category (b), found by E's own I2 flip) — set_shader_images is never emitted for a buffer-texture image

ERROR]: An image record does not describe the binding it names: image unit 0 holds a
        texture at a draw and no set_shader_images has ever been applied; running the
        pre-handle image bind.                                                        x5
  BufferTextureScenario.AnImageStoreIntoABufferTextureIsVisibleToTheCpu
  ImageTargetKindScenario.LoadsTextureBuffer
  ImageTargetKindScenario.StoresTextureBuffer
  NonCoreImageFormatScenario.ASplitBufferImageStillSamplesWholeTexels
  NonCoreImageFormatScenario.BufferImageAddressesTheApplicationsOwnTexels

All five pass, because E declines to the pre-handle bind - the fallback MAJOR-1's union exists to protect. The discriminator is the buffer flavour and nothing else: ImageTargetKindScenario runs one shader template over uimage1D, uimage1DArray, uimage2D, … and uimageBuffer, and only the Buffer kind appears here.

Emitting side: ImageEmit.h:81-90. The window is resolution.MaxImageUnit < 0 ? 0 : MaxImageUnit + 1, and count == 0 takes the zero early-out before any hash - so no set_shader_images is ever applied for the life of the process. MaxImageUnit comes from SamplerEmit.h:635-642, which counts a uniform location only when GetUniformSamplerOrImageUnitIndex(location) >= 0 && GetUniformTypeFacts(location).isImage. For an imageBuffer uniform one of those two answers wrongly; the unit index is the likelier - ES forbids glUniform1i on an image uniform, so it can only come from layout(binding=) at link, and ProgramLinkTask.cpp:57 sets isImage straight from glslang's type->isImage().

Reading side: DirectGLES.cpp's I2, the flip this round installed, which is the only reason this is visible at all.

Patch, for package C, in preference order:

  1. Fix the reflection so an imageBuffer / uimageBuffer / iimageBuffer uniform contributes its unit to MaxImageUnit. If the gap is in uniformSamplerOrImageUnitIndex rather than in isImage, that is MG_State's file and outside P4a's set - the integrator has to grant it, and it is then a pre-existing frontend reflection gap that P4a's client emitter is the first code to depend on, worth saying in the phase's docs.
  2. Failing that, union the emitted window with the units the CONTEXT actually holds an image binding on - the same union E's sweep already takes server-side (MAJOR-1), which would make the two sides agree by construction.

Until one of them lands, A8's image half is settled: MAJOR-1's union in SyncImageTextureBindings is permanent and may not become an assertion.


SD-5 (category (c), pre-existing) — 0x7ff half-runs the framebuffer family against a refused texture family

53 DirectGLES scenarios fail at MOBILEGL_PIPE_PUSH=0x7ff, identically with and without E, where every other arm this round ran is 6 or 8. The shape is D's: ResolveTextureResourceSubsystemArm refuses bit 10 because bit 11 is clear (D-K2's fourth row, correctly), so the texture family runs the LEGACY arm - but ResolveFramebufferSubsystemArm tests the raw mask bit 10, not the texture family's resolved verdict, so the framebuffer family ARMS. That is precisely the half-run D-K2 exists to forbid, one level up: the dependency is expressed against the BIT and the bit lies whenever the depended-on family refused itself.

Patch, for package D (Managers.cpp, ResolveFramebufferSubsystemArm): make the row read the arm rather than the bit -

// bit 9 requires the TEXTURE-RESOURCE ARM, not merely bit 10: bit 10 can be set and still
// refused (bit 10 requires bit 11), and a refused family runs its legacy arm, which is
// exactly the state MGPSurface::Res cannot be resolved against.
refused = !TextureResourceSubsystemEnabled();   // logs its own reason

with the existing PipeSubsystemDependencyMissing line kept for the plain bit-10-clear case so the message still names both bits. 0x7ff is not a shipped mask, so this is not urgent - but it is a lane that reports 53 failures to anyone who sweeps the masks, and F's dependency-refusal scenario is the natural place to pin it.

7. The retrace, and what it says about the tree rather than about E

python3 ~/w7/retrace_gate.py --tree ~/w7/p4a-esprytdraw --lib build-push/libMobileGL.so --out ~/w7/retrace-out/p4a-esprytdraw-v3 -j 4, default mask, 79 cases.

~/w7/pipe   ee31944b, WITHOUT E                     62 / 79   17 failed, all DirectGLES
~/w7/p4a-esprytdraw 0cd0d1ec                        61 / 79   18 failed, all DirectGLES
~/w7/p4a-esprytdraw 0cd0d1ec + SD-0's Tracker patch 62 / 79   17 failed - THE SAME SEVENTEEN,
                                                              set-identical to the control

So the whole of E's retrace delta is one case, create-indirect.DirectGLES, and SD-0's four-line patch closes it exactly.

79/79 was not obtainable, and is not obtainable without this package. The integrated tree is at 62/79 at the default mask, and the seventeen are identical in name and in SSIM to six decimal places between the two trees:

improved-transparency-minecraft-26.3.DirectGLES              0.994435
minecraft-1.21.11-main-menu.DirectGLES                       0.936674
minecraft-1.21.4-fabric-iris-bliss-in-world.DirectGLES       0.856432
minecraft-1.21.4-fabric-iris-bsl-esc-menu-854.DirectGLES     0.922890
minecraft-1.21.4-fabric-iris-bsl-in-world.DirectGLES         0.851414
minecraft-1.21.4-fabric-iris-chocapic-v6-lite-in-world       0.855946
minecraft-1.21.4-fabric-iris-complementary-reimagined        0.815194
minecraft-1.21.4-fabric-iris-complementary-unbound           0.806136
minecraft-1.21.4-fabric-iris-derivative-main-d24.4.14        0.761700
minecraft-1.21.4-fabric-iris-iterationt-in-world             0.985983
minecraft-1.21.4-fabric-iris-iterationt-nodsa-in-world       0.985049
minecraft-1.21.4-fabric-iris-makeup-ultrafast-in-world       0.981368
minecraft-1.21.4-fabric-iris-nostalgia-in-world              0.845284
minecraft-1.21.4-fabric-iris-photon-v1.1-in-world            0.824615
minecraft-1.21.4-fabric-iris-photon-v1.3b-in-world           0.970720
minecraft-1.21.4-fabric-iris-sundial-lite-in-world           0.812143
minecraft-1.21.4-fabric-iris-super-duper-vanilla-in-world    0.610444

Fifteen of the seventeen are Iris shader packs, plus improved-transparency-26.3 and one main menu; every DirectVulkan half passes, and every non-Iris in-world and inventory trace passes. Nobody had run a retrace at the DEFAULT mask on an integrated P4a tree before this round - B's 79/79 was at 0x1ff, and D's arms were 189/6/6 on the itest lane only - so this is the first measurement of it, and it is an integrator-level finding independent of E: seventeen shader-pack traces are wrong on the DirectGLES handle arm, at SSIM 0.61 to 0.99, with zero named refusals and zero seam lines anywhere in the census. That is category (c): a silent difference in the client emission or the applier, not a declined arm. E's decline sites cannot speak about it because none of them is declining, and the fact that the failures cluster so exactly on shader packs (many programs, many framebuffer switches, many image and sampler units per frame) is the strongest lead the round can hand over.

The SD-0 patch is reverted on the branch (Tracker.h is not E's file); the working tree is clean.


8. The verification-round checklist (v2 §6), item by item

item result
6.0 do not run wsl_tree.sh; git lfs checkout first the git lfs checkout instruction is WRONG on this box - see §1. The branch was never reset; refs/heads/p4a/esprytdraw-backup untouched
6.1 take §4's reconciliations first done, in the order given; §3
6.2 apply the flips, assert every arm engages done; §4, §4.1 - and one flip was corrected by the run itself (§6.1a)
6.3 zero seam hits, and Fatal{ProtocolCorruption} on verify 5 image-seam hits (SD-4, a finding against C, not a reason to relax the check), 0 framebuffer, 0 program, 0 Fatal{ProtocolCorruption}
6.4 A8's measurement answered, and it is the image half that matters: ImageEmit.h's own text says the window is the program's named units unioned with a sticky mark, not the driver's set, and SD-4 is that gap observed. MAJOR-1's union is permanent. The sampler half's window is [0, GetMaxTouchedTextureUnit()], the same bound E walks, so MAJOR-2 is expected unreachable but stays load-bearing until C documents the membership in SamplerEmit.h (it is capped at kMGPipeMaxTextureUnits and gated by a suppressor)
6.5 A3 / D-E3: does the read-FBO list still do work? yes, and settled by code rather than by a counter: IsDrawSyncClean was NOT re-keyed onto the record's ParamsSerial - it still compares the frontend's params version twice, the content version, and the sampler object's version, and returns false outright when the texture has no sampler object. Nothing on the read path advances any of them. D-E3's closure stays in E
6.6 unit-list equivalence between 0x1ff and 0x1fff the two arms' failing sets differ only by the two SD-1 scenarios, and every family census row is 0 on both, so no rebuild class is missing; the per-frame rebuild COUNT is not instrumented on either arm and stays an optimisation-list item
6.7 named traces photon-v1.3b and improved-transparency-26.3 both run, both fail identically with and without E (§7); the canary that actually fired was create-indirect, and it caught a real one (SD-0)
6.8 the full C.4 verification §5. p4a_untouched_regions.sh run from F's tree, rc 0, and the two D-N rows that live in E's file carry D's own shas. g_rawDepthFetchSamplerState confirmed still the only file-static SharedPtr<SamplerObject> in this file, pre-existing at 37da3c3a, untouched (P3b/P4b's)
6.9 carried to other packages A9 closed by D; ID-12's DV-5 still the integrator's; G9-as-white-box landed in D
6.10 Track H census unchanged: 7 memos re-keyed, 0 retired, 1 bypassed, 1 new

9. What the integrator must re-run, and what is owed by whom

Before the push

  1. SD-0 → the owner of MG_Impl/Pipe/Tracker.h (C / contract). The four-line shutter patch in §SD-0, plus the ImageEmit.h invariant-2 rewording. Verified on this tree: it turns create-indirect.DirectGLES from FAIL 0.887403 into PASS 0.999961 and leaves both itest lanes unchanged. Until it lands, E's tree is one retrace case behind pipe's, and that case is the only difference between them.
  2. SD-1 → package C. Two scenarios red on the gpu AND verify lanes, with a named refusal and a demonstrable wrong picture (0x3F800000 where an integer border colour belongs). This is the residual ID-12's rule is about.
  3. SD-4 → package C (possibly reaching into MG_State's reflection under an integrator grant). Five scenarios silently on the pre-handle arm.
  4. SD-3 → package D's report and package C's code. The census is short by two shapes; the residual check needs four grep patterns.
  5. SD-5 → package D, when convenient. 0x7ff is not a shipped mask.

Re-runs the integrator owes after any of the above lands

  • the DirectGLES -L integration-gpu default arm and the census (the recipe is §6.0 - the obvious ctest -V census is a false zero and must not be used);
  • the full retrace at the default mask, which nobody had run on an integrated P4a tree before this round and which is at 62/79 without E (§7). Those seventeen are a category-(c) finding with no refusal behind them and they are not this package's; they need an owner before the phase closes.

Not owed by anyone, recorded as settled

  • the DirectVulkan lane: 475/475, with and without E, at every mask this round swept, once 29d51ab9 landed;
  • the six armless .Handles aborts: F's, and now positively identified by their Fatal{PipeLegacyMemosDisabled} line rather than assumed;
  • ID-27: record.Target has no reader left in this file.

10. Deviations of this round

  • DEV-19 — F3 was KEPT against the review's A1 sketch. The A1 cell says the seam check "can go" once the twin resolve takes the handle; §2's table says "keep verbatim; it is the wording §8.2 greps". The table wins, and the reason is stronger than precedence: with the twin now resolved from the record and configured from the bound frontend object, FramebufferRecordMatchesBinding is the only thing between a mis-keyed record and one framebuffer's attachments being written into another's twin. Removing it would delete the check that makes A1 safe.
  • DEV-20 — F5 and F7 became assertions PLUS a kept null check, not assertions alone and not deletions: MOBILEGL_ASSERT is inert in a release build, and a dereference of a nullable accessor is not something to leave to an ordering argument.
  • DEV-21 — I2's condition is narrower than the review's. "When the program declares images" is not testable at that site; "the validate-point latch is set AND this unit holds an image texture" is, is strictly stronger, and names the unit. The latch was extended to cover the pre-handle sweep as well, because I2 fires precisely when the record arm is NOT taken and a latch covering only the record arm would make the flip unreachable.
  • DEV-22 — T2 and S2 are scoped to maxTouchedUnit >= 0, against the review's table, on the evidence of the first real-path run (§6.1a). 333 of 333 hits of the unnarrowed rule were draws that touch no texture unit.
  • DEV-23 — P6 does not re-test the review's condition because its only caller already guarantees it; re-testing would have been a second reader of one rule.
  • DEV-18 (v2) paid for itself: wire v3 landed the accessors as member functions returning a nullable pointer - both halves of what v2 predicted - and the rebase was one function plus four null checks.

11. Logs

Everything this round wrote lives in ~/w7/p4ae3/ and is deleted at the end of the round per ID-5: the edit scripts, the four build-*.logs, the arm logs, the per-process mglogs/, the census extracts, the bisect logs and the retrace logs. ~/w7/p4a-esprytdraw keeps only build-linux, build-push and build-verify (build-nolegacy was deleted with the logs; it is a compile gate, rebuilt in one command). The shared scratchpad wsl directory was not touched (ID-7). Nothing was pushed; ~/w7/pipe and the other p4a-* trees were read-only except for the control ctest runs inside ~/w7/pipe/build-push and ~/w7/pipe/build-verify, which write no tracked file.