Files
MobileGL/docs/Disaggregated/notes/p3a/p3a-results/espryt-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

38 KiB
Raw Blame History

P3a package C — Espryt verification round (espryt-v3)

Branch p3a/espryt, worktree /home/swung/w7/p3a-espryt, rebased onto refs/heads/feat/disaggregated (31e370be = contract + dev merge + wire + client + the two version bumps) and then three new commits. Not pushed. Every file touched is still in C.2's list; MultiDraw.cpp and Utils.cpp are untouched.

This is the round in which the handle path executed for the first time against real emitters and real applier bodies. It found three defects, all in C.2's files, all fixed and all with a measured before/after. It also found one red that is not this package's — 26 BufferTest unit cases — with a proof of ownership and the exact one-line fix for the owner (§7).

rebase clean, no conflict, 16 commits replayed
new commits 13b380fe, 75ea7ee2, 3c55e027
HEAD 3c55e027
default-mask integration lane 920 / 920 green (was 204/463 red on the stub base, 13 red on the first run of this tree)
retrace 79/79 push, 79/79 verify-armed, 0 Fatal
still red 26 BufferTest unit cases — package-boundary, §7

1. The rebase

git rebase refs/heads/feat/disaggregated          Successfully rebased and updated refs/heads/p3a/espryt

No conflict, and this time no hidden trap either (the ID-9 three-way-merge trap of the previous round was about Managers.cpp and dev's d7655247; 31e370be adds nothing in MG_Backend). All sixteen commits replayed: 6a1a9bc7 → ed83424c at the tip, 9ae4ec09/4db6bf1e → 9951961d at the base. git diff --stat against the new base names exactly SlotTables.h, Managers.h, Managers.cpp, DirectGLES.cpp, SanityTest.cpp.

One environmental repair was needed before the retrace could run: this worktree's tools/trace_replay/fixtures were LFS pointers (7.7 MB of 131/133-byte stubs against 803 MB in ~/w7/pipe), so the first retrace invocation reported passed 2 / 79 with rc=1 0s ssim=None on every Minecraft case — a missing-fixture failure that reads exactly like a total regression. They were materialised by copying ~/w7/pipe/tools/trace_replay/fixtures/. over them; the 92 assume-unchanged entries were already set, so git status stayed clean. Recorded for the integrator: a package worktree's fixtures are not materialised by wsl_tree.sh, and a retrace run against pointer files is a false red.


2. The two cross-package items

B-M3 — the hostBytes convention (client passes shadow base + offset)

Already conformant; no change was needed, and the report that said otherwise was wrong about the code.

client-v2.md §5 item 1 asks espryt to move its latch because the client sends base + at on resource_subdata and base + offset on resource_flush_range. The handle arm has always subtracted:

Ops_H_SubData    (Managers.cpp:1921):  if (bytes != nullptr) resource->hostBytes = static_cast<const Uint8*>(bytes) - offset;
Ops_H_FlushRange (Managers.cpp:1969):  if (bytes != nullptr) resource->hostBytes = static_cast<const Uint8*>(bytes) - start;

git log -S puts both lines in e2 (4e446501, the original switch-over commit), not in any later one. The subtrahends are the right ones: offset is MG_Pipe::MGPipeSubDataBufferOffset(record), which MGPipeTypes.h:661 defines as the record's UnionBox.X, i.e. exactly the at the client adds (PipeFill.cpp:650 MGPipeApplyResourceSubData(record, base + at) beside MGPipeBuildSubDataRecord(handle, at, …)); start is record.Offset, which the client adds at PipeFill.cpp:701. Ops_H_Respecify takes the pointer verbatim because the client sends the true base there (PipeFill.cpp:640), and Ops_H_ResidentSubData records nothing (D10), which matches the client's base + (at - offset) staging pointer.

So espryt-v1.md §1's phrase "latches the client's shadow base" described the value after the subtraction, which is what it is. B does not change and C did not need to. The behavioural half is now also observed rather than argued: CrossFrameBufferScenario's non-zero-offset flush and StreamedArenaScenario's two recycle cases — the shape client-v2 named as the one that would catch it — are green in the default-mask lane (§4, L1/L4), and they were green with the same code before and after this round, so nothing here is a new claim.

ID-13 — one FlushPendingRangesNow, and the handle arm's separate ladder

Confirmed, with the residual difference from ID-13's literal text stated rather than glossed:

  • Defined exactly once. grep -n FlushPendingRangesNow Managers.cpp finds one definition (:1316) and two call sites (:1723, :2872), plus two comments. Package D's script exits 2 on a second definition, and it exited 0.
  • Byte-identical to 5cb826b0. Its extracted-body sha is c65570022a5856c1afccfd0cb42fefac707689f5bb3154c3a8e17d2495ccdf8e at both refs.
  • FlushPendingRangesFrom is the separate ladder (:1051, under #if MOBILEGL_PIPE_PUSH), with four call sites (:1721, :2044, :2732, :2870).
  • No hunk lands inside any of the ten. git diff 5cb826b0 HEAD -- Managers.cpp is 16 hunks (-U0), and the gate below proves none of them touches a protected body.

The one deviation from ID-13's wording, and it is espryt-v2's declared D12, not a regression: ID-13 says "defined exactly once, outside any #if". The definition sits in the #else of #if MOBILEGL_PIPE_PUSH (verified by walking the preprocessor nesting: guard stack at :1316 is #else of #if MOBILEGL_PIPE_PUSH), so a push build contains no FlushPendingRangesNow at all and its legacy arm calls FlushPendingRangesFrom instead. Consequences, stated plainly so the integrator can rule:

  • both gates ID-13 names are satisfied — G1 (the pull build's text and symbol set are untouched) and G5 (the extractor finds one definition and it is byte-identical);
  • a push build contains one tier ladder, which is what ID-11's row exists for;
  • the price is that in a push build the MOBILEGL_PIPE_PUSH=0 / 0x7f arms run FlushPendingRangesFrom rather than the frozen body. The only textual difference between the two ladders is espryt-v2 m4's added if (hostBase == nullptr) return;, unreachable on the legacy arm. Both legacy arms are 920/920 green (§4).

Taking ID-13 literally instead would compile two ladders in every push build. This round did not change the shape; it is recorded here because the task's wording ("defined exactly once, byte-identical … and the handle arm's FlushPendingRangesFrom is the separate ladder") is met exactly, while ID-13's "outside any #if" is not.

Package D's G5 script, and whether it accepts a foreign cwd

It does not, and an invocation that looks like it worked is a false pass. scripts/p3a_untouched_regions.sh self-locates: REPO_ROOT=$(cd "$(dirname "$0")/.." && pwd); cd "$REPO_ROOT". Running bash /home/swung/w7/p3a-gates/scripts/p3a_untouched_regions.sh 5cb826b0 HEAD from /home/swung/w7/p3a-espryt exits 0 and prints "byte-identical between 5cb826b0 and HEAD" — but it has cd-ed into p3a-gates first, so HEAD is gates' HEAD, not espryt's. The authoritative form, which needs no copy of D's file in this tree, is to run it from gates' worktree against espryt's ref (the worktrees share /home/swung/w7/pipe/.git, so git show <ref>:<path> reaches it):

cd /home/swung/w7/p3a-gates && bash scripts/p3a_untouched_regions.sh 5cb826b0 p3a/espryt      rc 0
  [p3a-untouched] the 10 pool / deferred-release / ring / flush-drain functions are
  [p3a-untouched] byte-identical between 5cb826b0 and p3a/espryt
bash scripts/p3a_untouched_regions.sh --self-test                                              rc 0
  positive control: 10 bodies extracted / an untouched copy compares equal / an edit outside the ten is invisible
  negative control: a perturbed ClearBufferPool body is reported, and named
  negative control: a perturbed FlushPendingRangesNow body is reported, and named

The ten shas are unchanged from espryt-v2 §5.3.


3. The three defects this round found, and the fixes

The first execution of the handle path under real traffic left 13 failures in the default-mask integration lane (920 entries). Two of the thirteen (DirectVulkan.PrimGenReroute.…ActuallyArmed… and DirectVulkan.PointSizeDemotion.…ActuallyArmed…) are the shared-log-file race client-v2 §6 item 6 names and appear only in a -j 8 run; they are absent from every -j 4 run and from the serial reruns, and neither touches a file this package edits. The other eleven were real, and a mask bisect split them cleanly:

mask failures
0x7f (both subsystems off) 0
0xff (bit 7 resources only) 5 — XfbCaptureBufferReuse ×4, LargeArenaAdoption.RespecifiedIndexArenaKeepsVaoBinding
0x1ff (both) 11 — the five above plus VertexAttribBinding ×6

i.e. one defect in the vertex-input family and one in the resource family, plus a third that only the third of them exposed.

F1 — 13b380fe — the vertex-buffer entry was resolved by the attribute's GL binding point, not by its attribute index

Managers.cpp:3647 (the helper, renamed VertexBufferForBindingIndex → VertexBufferForAttributeIndex), :4070 (the call site).

-                    VertexBufferForBindingIndex(st, attrib.BindingIndex);
+                    VertexBufferForAttributeIndex(st, attribIndex);

MGPVertexBuffer::BindingIndex and MGPVertexAttribWire::BindingIndex are two different numbers. Espryt consumes resolved attributes, so the client emits set_vertex_buffers as one entry per attribute slot with MGPVertexBuffer::BindingIndex == the attribute index (VertexInputEmit.h:190-193, :229 — "Espryt consumes RESOLVED attributes, so the set is one entry per attribute slot with BindingIndex == the attribute index"), each carrying that attribute's already-folded buffer, stride and divisor. MGPVertexAttribWire::BindingIndex is the GL binding point the attribute was attached to by glVertexAttribBinding (VertexInputEmit.h:325: vao.GetAttributeBindingIndex(i)) and is the key of the binding-point view — a view this arm never reads, because the resolution already happened on the client. The helper's own comment already said the entries are keyed by attribute index; only the call site disagreed.

The two numbers coincide for every attribute configured through glVertexAttribPointer, which is why every other scenario stayed green. They differ exactly in KHR-GL43.vertex_attrib_binding's subject: glVertexAttribBinding(2, 3), glVertexAttribBinding(0, 5), glVertexAttribBinding(1, 0) — the six VertexAttribBindingScenario cases that failed, all of which read the disabled/zero default (point 0 attribute 2 is (0, 0, 0, 0)) because the lookup found the wrong entry or none.

Measured: the six went to three immediately after this fix alone (the three fp64 ones survived to F3).

F2 — 75ea7ee2 — the ensure path asked a descriptor no content call refreshes whether the shadow held bytes

Managers.cpp:2759-2761 (EnsureBufferResourceForHandle).

-            const void* initialData = record->Desc.HasDefinedContent != 0 ? hostBase : nullptr;
+            const Bool shadowHasContent =
+                bufferObject ? bufferObject->HasDefinedContent() : (record->Desc.HasDefinedContent != 0);
+            const void* initialData = shadowHasContent ? hostBase : nullptr;

MGPResourceDesc::HasDefinedContent states what was true at the last resource_respecify and nothing refreshes it afterwards: NotifySubData (BufferObject.cpp:85), NotifyFlushMappedRange (:101), NotifyContentWrite (:127), MarkGpuWritten (:365) and LandBytesIntoResidentStore (:441, :459) all set the frontend's m_hasDefinedContent without re-emitting a descriptor, and re-emitting one per glBufferSubData would be new wire traffic D-J forbids. The legacy arm never had the problem because RespecifyStorageNow (Managers.cpp:967-968) reads bufferObject.HasDefinedContent() and bufferObject.MappedData() — the source of the bytes and the statement about them come from the same object. The handle arm paired the frontend's pointer with the applier's descriptor, and the two disagreed after the commonest idiom in the corpus:

glBufferData(target, size, NULL, usage)     -> Desc.HasDefinedContent = 0
glBufferSubData(target, 0, size, data)      -> record.Serial moves; the twin queues the range
   (or: the twin does not exist yet, and the call is dropped by design - "lazy: the ensure path full-uploads")
draw -> EnsureBufferResourceForHandle: pendingRespecify -> RespecifyStorageWith(size, usage, NULL, serial)
                                       which CLEARS pendingRanges and STAMPS syncedChangeSerial

— an empty store, declared current, with the application's bytes dropped and no diagnostic anywhere. It is a silent wrong-pixels defect, not a crash.

Measured, with the instrumented build (temporary MGLOG_E traces, removed before the commits):

before  TRACE respecifyWith id=1 size=16 data=(nil)         serial=2   extentChanged=1
after   TRACE ensure {2,0} … pendRes=1 init=0 ranges=0 synced=5 serial=8 content=1 host=0x5b73c1bad3c0
        TRACE respecifyWith id=2 size=4096 data=0x5b73c1bad3c0 serial=8 extentChanged=1

That first line is XfbCaptureBufferReuseScenario's vertex buffer (glBufferData(16, nullptr) in the fixture, then glBufferSubData in SetVertex): its store was created empty, so the capture program captured zeros and every span read back component 0 is 0, expected 10. The second is LargeArenaAdoptionScenario.RespecifiedIndexArenaKeepsVaoBinding's final glBufferData(4096, nullptr) + glBufferSubData(elements) after the 24 MiB adopted arena is retired — the element store came up empty and the draw was black ("VAO did not fetch the replacement index store").

Closed by this commit alone: all four XfbCaptureBufferReuse cases and the LargeArenaAdoption index case, i.e. every 0xff failure.

The read is declared, exactly like C-1's IsMapped(): it is a frontend read this function keeps for P3a (the signature already carries the object for D-N's SyncPersistentMappedRange, and C-2 already re-reads MappedData() from it at every use) and it retires the moment the client publishes a live content flag beside the descriptor. With no frontend object — the handle-only drains — the descriptor is all there is, and that is what the ?: says. This is not the prohibited "re-read MGB_CTX to make a lane pass": no GL context is read, the object is the one this function is already given, and the alternative (a twin-side hostContentDefined mirroring six frontend setters, two of which — NotifyContentWrite on a resident store, and the write-map path at BufferObject.cpp:209 — emit no wire call at all) could not be proven equivalent today. That alternative is the P5/P8 shape and is named in the comment.

F3 — 3c55e027 — a lazily twinned store never published a shadow base, so the fp64 narrowing refused every draw

Managers.cpp:2741 (EnsureBufferResourceForHandle).

+            if (hostBase != nullptr) resource->hostBytes = hostBase;

resource->hostBytes is what the readers that hold no frontend object read, and there are three: Ops_H_Readback's queue flush (:2048), Ops_H_FlushRange's kill-switch map arm (:2005), and SyncFloat64AttributeAsFloat32ByHandle (:4479-4482, whose legacy counterpart reads bufferObject->MappedData()). It is written by the content-carrying ops — but a twin is created lazily, because D-A2's row makes Ops_H_Create a no-op, so the very first resource_respecify of a buffer finds FindBufferResourceForHandle == nullptr and drops its base on the floor:

TRACE respecify {2,0} width=16 defined=1 bytes=0x5fdd854e5440 res=(nil) id=0 storeSize=0

For the ordinary static vertex array — glGenBuffers + glBufferData(size, data) and no further content call — hostBytes then stayed null for the object's whole life. The full re-upload never noticed (it uses liveHostBase()), but the fp64 narrowing did: sourceBase == nullptr → return false → the Adreno workaround disables the array → the shader reads the disabled default (0, 0, 0, 1). That is exactly what DoubleArrayIsFetchedAtFloat32Precision, NormalizedIsIgnoredForDoubleArrays and LongDoubleArrayIsFetchedAtFloat32Precision reported.

The refresh is placed where the base is already computed, which is the one place that both holds the object and runs before every draw that uses the store. It is reached only for a non-adopted resource (the adopted arm returns at :2683), so C-2's rule is intact: an adopted store's hostBytes stays null and its bytes are read through persistentPtr.

Closed by this commit: the last three VertexAttribBinding failures. After it the whole lane is 920 / 920.

Instrumentation hygiene

Five temporary MGLOG_E("TRACE …") lines (in Ops_H_Respecify, Ops_H_Readback ×2, RespecifyStorageWith and EnsureBufferResourceForHandle) were used to obtain the traces above. They were removed before any commit; the reconstruction of the three commits was diffed against the verified working tree and is code-identical (the only textual difference is one re-wrapped comment line). grep -n "TRACE " Managers.cpp is empty at HEAD.


4. Verification transcript

All in /home/swung/w7/p3a-espryt at 3c55e027, CCACHE_BASEDIR=/home/swung/w7. All three build directories were rebuilt before anything was judged. build-verify was configured with the COMMON flags of ~/w7/notes/tools/wsl_p3a_gate.sh plus -DMOBILEGL_PIPE_VERIFY=ON -DMOBILEGL_ITEST_REQUIRE_GPU=ON.

4.1 Builds and unit

cmake --build build-linux  -j 24    rc 0
cmake --build build-push   -j 24    rc 0
cmake --build build-verify -j 24    rc 0

ctest --test-dir build-linux  -L unit -j 12    100% passed, 0 failed out of 1619
ctest --test-dir build-push   -L unit -j 12     98% passed, 26 failed out of 1619   <-- §7, NOT this package's
ctest --test-dir build-verify -L unit -j 12     98% passed, 26 failed out of 1619   <-- the same 26

The 26 are the whole and only red left on this tree. They are BufferTest.* and they are a test-fixture collision at the client×espryt seam; §7 proves the ownership and gives the fix, which was measured green.

4.2 G1 — the pull build is still inert

python3 scripts/symbol_report.py --before ~/w7/p3a-before-libMobileGL.so --after build-linux/libMobileGL.so \
        --threshold 0 --fail-on-symbol-set-change --fail-on-added-bytes 0            rc 0
symbol-report: .text 10806323 -> 10806323 (+0, +0.000%)   .data +0  .bss +0  .rodata +0
symbol-report: Total 17226471 -> 17226471 (+0)
symbol-report: 27811 -> 27811 defined symbols: 0 added, 0 removed, 0 resized, 0 renamed

0 / 0 / 0 / 0. ~/w7/p3a-before-head.txt reads b9eaa474…, i.e. ID-8d's re-captured baseline (the pull build taken after the two version bumps), which is this branch's base.

4.3 Names — G2's name half, and G14

ctest -N:  build-linux 2539   build-push 2539   build-verify 3371
diff <linux names> <push names>                                     -> EMPTY (0 lines)
comm -23 ~/w7/p3a-before-ctest-names.txt <linux names>              -> 0 removed
comm -13 ...                                                        -> 20 added

0 removed. The 20 added are this package's two (DirectGLESBufferDrawProbe.ALiveHostMapKeeps…, DirectGLESSlotTable.AGenerationBehindTheLiveTwinIsRefused…, both running in build-push and skipping visibly in build-linux) plus 18 from the merged client/wire (ResourceEmit.* ×7, VertexInputEmit.* ×7, TrackerWalk.* ×3, TrackerShippedEmitter.* ×1). This round adds no test name.

4.4 G5 — the ten byte-identical functions

rc 0, see §2. Self-test rc 0, both negative controls tripped and named.

4.5 Purity

grep -rc 'pGLContext' MobileGL/MG_Backend | grep -v ':0$'        -> empty
grep -c  'MG_State'   MobileGL/MG_Pipe/PipeApply.h               -> 0
git status --porcelain                                            -> clean

g_pendingFetchBaseInstance / ScopedFetchBaseInstance still have hits and every one is inside a MOBILEGL_PIPE_LEGACY_MEMOS arm (Managers.h:1191-1200, Managers.cpp:3523-3530, :3745, DirectGLES.cpp:5293) — D2, unchanged.

4.6 The integration lanes — every arm green

Run serially, no competing load (the protocol client-v1 established for the …ActuallyArmedWhenTheEnvironmentPinsItOn family).

# lane result
L1 ctest --test-dir build-push -L integration-gpu -j 8 (DEFAULT mask 0x1ff, both backends) rc 0 — 100%, 0 failed of 920
L2 the same with MOBILEGL_PIPE_PUSH=0 rc 0 — 100%, 0 of 920
L3 the same with MOBILEGL_PIPE_PUSH=0x7f rc 0 — 100%, 0 of 920
L4 the buffer/VAO family, -j 4 -R 'LargeArenaAdoption|StorageBufferRegrow|VertexAttribBinding|MultiDraw|PrimitiveRestart|CrossFrameBuffer|ResidentIndex|BufferTexture|AtomicCounter|XfbCaptureBufferReuse|PackedWordReadback|DoublePrecision|VertexArrayEnableDisable|DrawParameters' rc 0 — 100%, 0 failed of 184
L5 MOBILEGL_ESPRYT_DISABLE_INVALIDATE_FLUSH=1 … -j 8 (D.3's named kill-switch arm) rc 0 — 100%, 0 of 920
L6 ctest --test-dir build-verify -L integration-verify -j 4 rc 0 — 100%, 0 failed of 832, Fatal{ lines: 0
L7 MOBILEGL_PIPE_PUSH=0xff (bit 7 alone, the supported one-family A/B) rc 0 — 100%, 0 of 920
L8 MOBILEGL_PIPE_PUSH=0x17f (bit 8 without bit 7) rc 0 — 100%, 0 of 920, and the refusal fires (below)

DoublePrecisionScenario — the declared watch item — is green in L1 and L4 and needed no fix. D7's SyncGpuWrites gap on SyncFloat64AttributeAsFloat32ByHandle is still real and still declared; it is only observable when the fp64 source buffer is shader-written and not yet pulled back, which no scenario in the corpus does. What DoublePrecisionScenario and the three VertexAttribBinding fp64 cases did expose is F3, which was a different defect on the same function (a null source base, not a stale one) and is fixed.

L8 / espryt-v2 §6 item 10 — 0x17f refuses, by name, and runs the legacy vertex-input arm:

[ERROR]: MGPipe: kMGPipeSubsystemVertexInput (bit 8) is set but kMGPipeSubsystemResources (bit 7) is clear;
         the vertex-input handle arm resolves attribute buffer ids through the resource slot table, which only
         bit 7 populates - REFUSING bit 8 and running the legacy vertex-input arm. Set bit 7 as well, or clear both

4.7 ID-10 — the base-instance control, both directions

The brief asks for the base-instance scenarios green on this tree and red with the three MG_Impl/GLImpl/Drawing/GL_Drawing.cpp call sites stubbed out. The plain form of that control is not falsifiable on this host, and the reason is a capability, not a defect:

[INFO]:  base instance (EXT_base_instance; emulated by attribute offsets when absent): yes

llvmpipe exposes GL_EXT_base_instance and all three entry points, so BackendUsesNativeBaseInstance() (Managers.cpp:3569 = g_GLESCapabilities.SupportsBaseInstance) is true and Managers.cpp:4044 forces fetchBaseInstance to 0 — the attribute-offset emulation, which is the only consumer of MGPipeApplierState::VertexFetchBaseInstance anywhere in the backend, never runs. First experiment, measured: with the three call sites stubbed out, the base-instance scenarios were 10/10 green and the whole lane was 920/920 green.

So the control was run in the configuration where the pushed value is observable — the emulation — by forcing both native-base-instance gates false (Managers.cpp:3569 and DirectGLES.cpp:5276's UseNativeBaseInstance(); forcing only the first double-applies the shift, which is itself a positive proof that the pushed value reaches the fetch, and was measured: 2/10 red with the call sites present, 10/10 green with them stubbed — the inverse, exactly as a double shift predicts).

experiment (temporary, in place, never committed) base-instance scenarios whole lane
F1 both native gates forced false (emulation live), the three call sites present 10 / 10 PASS 920 / 920
F2 both native gates forced false, the three call sites STUBBED OUT 8 / 10 — 2 FAILED —
restored (git checkout -- on all three files) 10 / 10 PASS 920 / 920

The two that go red are VertexAttribBindingScenario.BaseInstanceMovesTheInstancedArraysStartElement and BaseInstanceLeavesPerVertexArraysWhereTheyWere. DrawParametersScenario's eight cases stay green in both directions, and that is correct rather than a hole: they observe gl_BaseInstance / gl_DrawID, which come from SetCurrentBaseInstance — "orthogonal and runs either way — it feeds the shader's gl_BaseInstance, not the fetch" (DirectGLES.cpp:5272-5274). The ID-10 gate is the two VertexAttribBinding base-instance cases, not the DrawParameters eight, and the integrator should record that correction.

git status was verified clean and the tree rebuilt after every experiment; grep -rc 'ID-10 CONTROL, TEMPORARY' is 0 in all three files.

4.8 The retrace, both sweeps

python3 ~/w7/retrace_gate.py --tree ~/w7/p3a-espryt --lib ~/w7/p3a-espryt/build-push/libMobileGL.so \
        --out ~/w7/retrace-out/p3a-espryt -j 4                                     rc 0
    passed 79 / 79; failed: []
    logs 79, with Fatal{: 0

MOBILEGL_PIPE_VERIFY=1 python3 ~/w7/retrace_gate.py --tree ~/w7/p3a-espryt \
        --lib ~/w7/p3a-espryt/build-verify/libMobileGL.so \
        --out ~/w7/retrace-out/p3a-espryt-verify -j 4                              rc 0
    passed 79 / 79; failed: []
    logs 79;  armed (MGPipe verify: present) 79;  NOT armed 0;  with Fatal{: 0

Every case at or above its own SSIM threshold on both backends (rd12-odinlite 1.0 / 0.999991, 1.21.4-in-world 0.999976, 1.21.4-main-menu 0.997009, OpenRA 1.0/1.0, startup 1.0/1.0). This is the first verify-armed retrace of the whole corpus with the backend switched over: the compare-at-read re-reads every migrated field from the live context at every backend read and found zero divergence and zero unmigrated reads. No Fatal{UnmigratedPipeInput}, so no FillPoints.def row is owed. Both output trees were deleted afterwards (retrace-out/p3a-espryt* → 0 left).

4.9 D.4.6 — peak RSS

Not available, so skipped. ~/w7/retrace_gate.py has five arguments (--tree --lib --out -j --only) and contains no rss / maxrss / getrusage reference at all (grep -ci → 0). A push-vs-pull peak-RSS reading over the retrace would need the gate to grow the measurement; it is not something this round can report without changing package D's tool.

4.10 Cross-package confirmations owed by espryt-v2 §6

item result
12 — wire's C2: the applier reset must advance the two serials holds. PipeApply.cpp:645-646 (MGPipeApplierReset) and :658-659 (MGPipeApplierReleaseObjectRecords) are ++; grep finds no = 0 on either serial.
17a — OnGpuWritten sets both m_hasDefinedContent and m_gpuWritePending (m8) holds. ResourceTracker.h:614 calls buffer->MarkGpuWritten(), which sets both (BufferObject.cpp:364-367).
17b — MGPVertexBuffer::BindingIndex is the attribute index (D-H3) holds (VertexInputEmit.h:229) — and F1 is the fix for the arm that assumed the opposite.
B-DEV-7 — kMGPipeResourceTargetBuffer == 0 confirmed (ResourceTracker.h:160).
13 — D.3's MOBILEGL_ESPRYT_DISABLE_INVALIDATE_FLUSH=1 arm green, §4.6 L5.
10 — 0x1ff / 0x7f / 0x17f all three run; 0x17f refuses by name, §4.6 L8.

5. Track H census inputs

Unchanged from espryt-v1 §2 as amended by espryt-v2 §3, plus this round's three:

Members deleted from the handle arm and kept compiled under MOBILEGL_PIPE_LEGACY_MEMOS (a pull build forces that ON, so no sizeof moves — G1 is 0/0/0/0 at HEAD, which is the measurement):

  1. BackendVertexArrayObject::m_hasSyncedConfigVersion
  2. BackendVertexArrayObject::m_syncedConfigVersion
  3. BackendVertexArrayObject::m_syncedAttributeVersions (Array<VertexAttributeVersion, 32>) — the largest single deletion of the phase
  4. BackendVertexArrayObject::m_syncedIndexBufferVersion (wrapping Uint16)
  5. BackendVertexArrayObject::m_syncedIndexBufferObject (raw frontend pointer — the wrap-hole identity patch)
  6. ConvertedFloat64Stream::sourceLifetimeId (ConvertedVertexStreamKey's sourcePin)
  7. VertexArrayImpl::g_pendingFetchBaseInstance + SetPendingFetchBaseInstance + GetPendingFetchBaseInstance + ScopedFetchBaseInstance and its three scopes

Kept, NOT deleted (D-K): PipeResource::m_backend, SetBackendResource, ReleaseBackend, BackendBufferResource — under push they simply stop being written. Deleting them moves sizeof(BufferObject) in the pull build, which is a straight G1 break; they retire with the pull path at P13.

Re-keyed memos:

memo was is
BackendVertexArrayObject's sync gate config version + 32 per-attribute version triples {elementsHandle, elementsSerial} + VertexBuffersSerial
the twin's index memo wrapping Uint16 + raw BufferObject* m_syncedIndexSerial (Uint64)
ResolvedDrawBuffers::configVersion frontend VAO config version {elementsHandle, elementsSerial, buffersSerial}
ResolvedDrawBuffers::Entry raw BufferObject* + MGPipeHandle handle (the probe's identity); iboFrontend likewise + iboHandle
ConvertedFloat64Stream frontend lifetime id + frontend change serial buffer {slot, gen} + applier Serial
GLESBufferResource::syncedChangeSerial mirrors BufferObject::GetChangeSerial() mirrors MGPipeResourceRecord::Serial

Push-only members added (none of them in a pull build): GLESBufferResource::hostBytes, BackendVertexArrayObject::m_syncedVertexBuffersSerial / m_syncedElementsHandle / m_syncedElementsSerial / m_hasSyncedElements, ResolvedDrawBuffers::{elementsHandle, elementsSerial, buffersSerial, Entry::handle, iboHandle}, BackendSlotTable::kMaxHandleSlot + LiveGenAt (no member).

Unchanged on purpose (D-J / D-A4): g_bufferBackendIdGeneration, m_hasConvertedFloat64Attribute, m_syncedFetchBaseInstance (as the emitted record), BackendSlotTable's single-entry m_memoLifetimeId/m_memoHandle front memo, TempBufferTarget and its redundant-bind cache.

This round's additions to the census:

  • GLESBufferResource::hostBytes gains a publisher: EnsureBufferResourceForHandle republishes the live shadow base at every ensure (F3). Its lifetime rule is otherwise unchanged (refreshed by a content-carrying call, cleared by an orphaning respecify and by a successful persistent-map adoption, never read by a path that holds the frontend object).
  • IsBufferDrawCleanByHandle reads five things (four applier reads + the frontend's IsMapped(), C-1) and EnsureBufferResourceForHandle now reads two frontend things beyond SyncPersistentMappedRange: MappedData() (C-2) and HasDefinedContent() (F2). All three retire at P5/P8 and are named at their sites. D6's "no MG_State type in any new handle-arm signature" is unchanged: it holds for the nine op bodies and the object-free helpers, and its three named exceptions are EnsureBufferResourceForHandle, MarkBufferGpuWritten and IsBufferDrawCleanByHandle.
  • No memo was retired or re-keyed by this round; VertexBufferForBindingIndex → VertexBufferForAttributeIndex is a rename of a lookup helper, not of a memo.

6. Deviations

D1-D14 stand as espryt-v1 §5 / espryt-v2 §4 record them, with two notes:

  • D7 narrows again. The fp64 function had three problems, not one: the declared SyncGpuWrites gap (still open, still P8's), the freed shadow base (C-2, fixed last round), and the base that was never published at all (F3, fixed here). Only the declared one is left.
  • D12 is the live divergence from ID-13's letter — a push build contains no FlushPendingRangesNow. §2. It wants an explicit integrator ruling; it costs nothing today (both legacy arms are 920/920).

No new deviation is declared by this round.


7. Still red, and it is NOT this package's: 26 BufferTest unit cases

Symptom. In build-push and build-verify under the default mask, 26 of the 86 BufferTest cases fail; they pass in build-linux, and under MOBILEGL_PIPE_PUSH=0 and =0x7f in build-push. The failures are all EnsureGpuResidentStorage() returning false and mock.gpu.data() staying NULL — the mock backend is never called.

Mechanism, with the evidence. MG_Test/Buffer/BufferTest.cpp:1650-1655's ScopedBackendOps installs a mock through MG_State::GLState::SetBufferBackendOps(&kZeroCopyMockOps). The BufferTest binary brings up the real DirectGLES backend (its log: Initializing MobileGL… Renderer Name: Espryt), so BufferImpl::RegisterBufferBackendOps() runs and, per this package's D3, also installs the handle-shaped table (Managers.cpp:2454 MG_Pipe::MGPipeSetResourceOps(&g_glesResourceOps)). MGPipeResourceSubsystemEnabled() (PipeFill.cpp:593) is therefore true, and the frontend's push arms (BufferObject.cpp:602-608, :655-661, and six more) route the call to MGPipeEmitMapPersistent and never reach g_bufferBackendOps — the mock the fixture installed. The fixture scopes one table and not the other.

Whose it is. Not espryt's implementation choice: C.2's step e2 prescribes installing the table "at bring-up … beside RegisterBufferBackendOps", so any conformant espryt produces this. Not the client's either: the eight BufferObject.cpp dispatch branches are exactly what D-K specifies. It is the merge-seam case C.5 warns about in as many words — "Green on two branches separately is not green on the merge": at espryt-v2 these 86 were green because nothing emitted through the pipe; at client-v2 they were green because nothing registered a table; together they are red. MG_Test/Buffer/BufferTest.cpp is "nobody" in C.5's ownership table, i.e. the integrator's (or D's) to land, and it is not in C.2's file list, so this package may not commit the fix.

The fix, and it is measured, not proposed. ScopedBackendOps must scope the pipe table too, exactly as MG_Test/Pipe/ResourceEmitTest.cpp:139-153's ApplierGuard already does for the applier:

    struct ScopedBackendOps {
        explicit ScopedBackendOps(const MG_State::GLState::BufferBackendOps* ops) {
            MG_State::GLState::SetBufferBackendOps(ops);
#if MOBILEGL_PIPE_PUSH
            m_savedResourceOps = MG_Pipe::MGPipeGetResourceOps();
            MG_Pipe::MGPipeSetResourceOps(nullptr);   // these 86 cases are BufferBackendOps tests;
#endif                                                // the pipe-side dispatch is ResourceEmitTest's
        }
        ~ScopedBackendOps() {
#if MOBILEGL_PIPE_PUSH
            MG_Pipe::MGPipeSetResourceOps(m_savedResourceOps);
#endif
            MG_State::GLState::SetBufferBackendOps(nullptr);
        }
#if MOBILEGL_PIPE_PUSH
        const MG_Pipe::MGPipeResourceOps* m_savedResourceOps = nullptr;
#endif
    };

plus #include <MG_Pipe/PipeApply.h>. Applied temporarily in this worktree and measured:

B: BufferTest (default mask) : 100% tests passed, 0 tests failed out of 86
B: whole unit suite          : 100% tests passed, 0 tests failed out of 1619
(reverted; back to the committed tree: 70% passed, 26 failed of 86)

It weakens nothing: those 86 cases are BufferBackendOps dispatch tests, the pipe-side dispatch is covered by ResourceEmitTest's 25, and the fix restores exactly the arm they were written against. A larger alternative — giving the fixture a pipe-shaped mock so the same assertions run over the handle path — is worth doing, but it is a rewrite of 86 cases and belongs to whoever owns that file.

Also recorded, and not fixed here: the two DirectVulkan.…ActuallyArmedWhenTheEnvironmentPinsItOn cases that failed only in the very first -j 8 run of this tree and in no -j 4 run, no serial rerun and no later -j 8 run. client-v2 §6 item 6 already names this as the shared-log-file race at MG_IntegrationTest/CMakeLists.txt:1040; this package touches no DirectVulkan file and no MG_IntegrationTest file.


8. Open items carried forward (not this round's to close)

  • wire-review-v1 M-D's applier half is closed (PipeApply.h:112-113, bounded); the backend half was bounded by M-1 last round. Nothing left here.
  • The dangling-ResolvedDrawBuffers::Entry::resource invariant after ReleaseByHandle: both consumers still check identity through FindByHandle first, so nothing is dereferenced, and the invariant still deserves a sentence at the member in a later pass.
  • D7's SyncGpuWrites gap on the fp64 narrowing — P8's, unchanged.
  • The three GL_Buffer.cpp NoteBoundAs sites (client-v2 §6 item 2) and delete_vertex_elements / the VAO CSO slot free (client-v2 §6 item 3, which C.5 gives to this package but which C.2's ordered steps do not contain) — both still open, both invisible in monolith.
  • Package D's itest for C-1 (a persistently mapped, non-adopted buffer written only through its pointer, then drawn and read back) is still owed; SanityTest's unit case pins the probe's semantics but issues no GL.

Intermediate logs (~/w7/espryt3-*.log, ~/w7/espryt3-*.sh, ~/w7/espryt3-fixes*.patch) and the scratchpad's espryt3-* scripts are deleted; ~/w7/retrace-out/p3a-espryt and -verify are deleted; git status in /home/swung/w7/p3a-espryt is clean at 3c55e027. The trace fixtures are left materialised (803 MB, assume-unchanged intact), which is the state the next retrace run needs.