Files
MobileGL/docs/Disaggregated/notes/p5/p5-results/t1-review-v1.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

27 KiB
Raw Blame History

t1-review-v1 — adversarial review of package t1 (test and CI plumbing for the SPLIT arm)

Reviewed p5/t1 5175f9cc..8d0850bd (4 commits, 13 files) + ~/w7/notes/tools/wsl_p5_{gate,rss}.sh. Worktree under review ~/w7/p5-t1; a detached review worktree ~/w7/p5-t1rev was built at 8d0850bd for the mutations and removed afterwards. No fix, no commit, no device, no LFS fetch.

Counts: 0 BLOCKER / 7 MAJOR / 8 MINOR / 3 QUESTION. Gates observed actually going red: 5 of 13 (see the table in §3). Every number t1 reports that I could re-run, re-ran identical.


1. The MAJOR findings

M-1. The arming conjunction's negative half is a grep for a message string the owning packages control. Renaming it arms 11 lanes and 8 of them go GREEN having run monolith.

Evidence (mutation, observed). In ~/w7/p5-t1rev I changed nothing but the message text in the six c0 stub files — sed -i 's/Fatal{Unimplemented/Fatal{NotYetImplemented/' over MG_Remote/Client/{CapsMirror,ClientSession,EmitTables}.cpp and MG_Remote/Server/{PipeApplier,ServerLoop,ServerSession}.cpp. Every entry point still ends in std::abort(); RemoteEmitTable() still aborts on sight; ImplementedVerbCount() still returns 0. Reconfigured and rebuilt build-split, then ctest -L integration-split:

91% tests passed, 1 tests failed out of 11
  8 x Passed      (5 Split.ClearThenReadPixels, 2 Split.Triangle, 1 Split.PersistentCoherentMap)
  1 x Skipped     DirectGLES.Split.PersistentCoherentMapScenario.TheMapLandsInTheArmItsLaneDeclares
  1 x Failed      DirectGLES.Split.PersistentMapArm.PersistentCoherentMapScenario...
                  "declared Which is: emulated / std::string(arm) Which is: adopted"

(~/w7/p5-t1rev-mutA-isplit.log.)

So the exact defect t1 discovered and reports as fixed in §3(a) is still reachable, by one word. The CMakeLists comment claims the negative half "is layout-independent in the way the house rule demands — c1, s1 and v1 have to delete those aborts to implement anything". That is not what the probe tests. mgl_itest_probe_for_symbol is file(GLOB_RECURSE …) + file(STRINGS … REGEX) (MG_IntegrationTest/CMakeLists.txt:408-420), so it tests only whether a line of text matching Fatal.Unimplemented survives under two named directories. Any of these disarms it without implementing anything:

  • renaming the message (measured above);
  • hoisting the MGP5_C0_STUB macro into a shared header outside Client/ and Server/ — the stubs are per-file macros today (MG_Remote/CapsCodec.cpp:32 is already that shape);
  • an implementation that keeps the abort but words it differently.

Concrete failure scenario. c1's author tidies the stub message while wiring the first three verbs. On the next CI run the 11 Split. entries arm and 8 of them report green under names that say they tested the split path, on an emit table that is not wired.

What saves it today, and how thin that is. The integration-split job's negative-control step does catch this: in the fake-armed tree both controls correctly reported "entries stayed GREEN" and would emit ::error (verified — §3 rows 7/8). But that control is two-state, it is a CI-only step, and the wsl_p5_gate.sh copy of it has its own hole (M-4).

Fix. Make arming a statement about this process, not about the source tree: assert MGITEST_SPLIT_EXPECT_TRANSPORT per case against the library's own resolved transport. t1 already carries that marker and names the gap as debt §6.3 — it is the fix for this finding, and it should land before the lanes can arm rather than after.

M-2. The same probe can be kept permanently DARK by a comment, and the only notice is a configure STATUS line.

Evidence (mutation, observed). With M-1's rename applied (i.e. "the packages have landed"), adding one comment line to MG_Remote/Client/CapsMirror.cpp:

// NOTE: this file used to raise Fatal{Unimplemented...} before c1 landed.

reconfigure prints

-- Integration tests: MG_Remote still carries c0 signature stubs (.../Client/CapsMirror.cpp ) -
   packages c1/s1/v1 have not landed, so every DirectGLES.Split. entry stays registered and SKIPS

file(STRINGS … REGEX) cannot tell code from comments. Every Split. entry and both negative controls then skip forever, in a fully implemented tree, and nothing fails or warns — the "test quietly measuring nothing" outcome the probe's own comment block says it exists to prevent.

Fix. Mask comments before matching (the house already has MaskCommentsAndQuotedText for this class of pass), or replace the content probe with the runtime assertion of M-1.

M-3. The probe scans only MG_Remote/Client and MG_Remote/Server. Two other stub files under MG_Remote are invisible to it — so the conjunction can arm partially.

grep -rl 'Fatal.Unimplemented' MobileGL/MG_Remote returns eight files. Six are the ones the probe watches. The other two are MG_Remote/CapsCodec.cpp (Fatal{UnimplementedCapsCodec …} - P5 package w1 has not landed) and MG_Remote/Wire/PipeWireCodec.cpp — both package w1's, both outside the probed directories.

Concrete failure scenario. c1/s1/v1 land; w1 has not. All 11 entries arm while every wire encode aborts. That is loud rather than silent (the abort reds the entries), but it also makes both negative controls tautological — the entries are red whatever the knob says, so E1 and E3(a) "pass" having proved nothing, and the job's red comes from the wrong place.

Fix. Probe ${MGL_ITEST_ROOT}/MobileGL/MG_Remote recursively rather than two subdirectories.

M-4. wsl_p5_gate.sh's E1/E3(a) negative controls report "RED, as it must be" when their filter selects zero tests.

wsl_p5_gate.sh (the E1/E3(a) loop):

( cd build-split && env "$knob" ctest -L integration-split -R "$filt" --no-tests=error > /dev/null 2>&1
  if [ $? -eq 0 ]; then echo "*** FAIL negative control $knob left the split lane GREEN"
  else echo "negative control $knob: RED, as it must be"; fi )

--no-tests=error makes an empty selection exit non-zero, which this reads as success. Observed: ctest -L integration-split -R 'ThisMatchesNothingAtAll' --no-tests=error → rc=8 → the gate would print "RED, as it must be."

The CI copy of the same control is correct — it counts ctest -N … | grep -c first and ::errors on < 1. The gate script is the one a package author runs before pushing; the two should not differ on exactly the falsifiability question. Today both filters do select 7 and 3 entries, so the hole is latent, not active.

Fix. Port CI's matched= guard into wsl_p5_gate.sh.

M-5. The run-level transport guard matches a bare KEY=VALUE substring, so a DEBUG-level build of a pull library passes it.

run_trace_case.cmake:255 is string(FIND "${split_log}" "MOBILEGL_TRANSPORT=inproc" …), and the integration-split job's step is grep -q "MOBILEGL_TRANSPORT=inproc" "${log}". The intended source of that string is ConfigLoader::InitTransport's INFO line (ConfigLoader.cpp:333, "Config: MOBILEGL_TRANSPORT=inproc - the MGPipe record stream crosses a real ring to an apply thread"), which exists only under MOBILEGL_BUILD_DISAGGREGATED.

But ConfigLoader.cpp:71 unconditionally logs MGLOG_D("Config: Accepted env variable: %s=%s", key, value) for every MOBILEGL_* variable, in every build including the pull one. At MOBILEGL_LOG_ACTIVE_LEVEL=…_DEBUG that line reads Config: Accepted env variable: MOBILEGL_TRANSPORT=inproc and satisfies both greps.

Observed (with a hand-written log of exactly that shape):

-- VERDICT: GREEN - t1's run-level guard accepted this log at offset 167
CI step: PASSES

Default and CI are INFO, where MGLOG_D is compiled out, so this is not live today. It becomes live for exactly the person most likely to hit it: someone who rebuilds at DEBUG to debug a split failure. The report calls this guard "the falsifiable half" (§3(b)) and it is one grep away from being airtight.

Fix. Match the distinctive part of the INFO line ("Config: MOBILEGL_TRANSPORT=inproc - the MGPipe record stream"), not the substring an env dump also produces.

Related, same guard, unfixed either way: the string searched for is hard-coded inproc while the condition is NOT $ENV{MOBILEGL_TRANSPORT} STREQUAL "monolith", so MOBILEGL_TRANSPORT=spawn (which ConfigLoader deliberately refuses by name and stays on monolith) reds with a message about inproc rather than about the refusal it actually got.

M-6. retrace-split is a clone of retrace-verify with the negative control removed, and the exit gate's own stated control (E2's "drop one Clear emission") is neither implemented nor named as skipped.

retrace-verify carries an always-on step, "Negative control - a corrupted snapshot field must red this retrace" (MOBILEGL_PIPE_VERIFY_CORRUPT=GetRenderStateParameters on OpenRA/DirectGLES), written precisely so "40 traces, zero divergences" is not a statement about a comparator nobody watched. retrace-split has no equivalent step. BRIEF §7 E2's control is "把 Clear 的发射器改 成丢弃一次(临时补丁)必须让 SSIM 掉下阈值"; the report substitutes the transport-resolution assertion (§3(b)) without saying the charter's control was dropped.

Those two are not equivalent. The transport assertion proves ConfigLoader printed a line; it does not prove a single record crossed the ring. Once c1 lands, an emitter that resolves the transport and then emits nothing still passes retrace-split at ssim 1.000. t1 did verify the pull-library control by hand (I reproduced it, §3 row 4) — but by hand, once, locally, not as a CI step.

Fix. Add the pull-library control to retrace-split as an always-on step (swap in build-linux/libMobileGL.so, assert the case reds), and record E2's emitter-drop control as an explicit debt on the commit that lands c1.

M-7. Exit gate E3(e) — "以上五条再跑一遍 ring 小到足以至少发生一次背压等待的配置" — is absent from the package and absent from the report.

git diff 5175f9cc..8d0850bd | grep -E 'IPC_RING_MB|IPC_STAGE_MB|背压' is empty; no lane sets a small ring; t1-v1.md §5 (brief/contract disagreements) and §6 (debts) do not mention E3(e) at all. The brief's reporting rule is "每一条 brief/契约没做的裁定(带证据 + 推翻它需要什么)" — this one is silently missing rather than argued away. It may well be un-runnable before s1 lands a ring that can backpressure; that is the argument that should have been written down.


2. MINOR and QUESTION

N-1 (MINOR). The skip reason the 11 entries print is factually wrong. Verbatim from ctest -V:

MobileGL/MG_Remote/Client does not exist yet - no source there names kRemoteEmitSlotCount / BackendObject_Remote / CapsMirror, so package c1 has not landed…

MG_Remote/Client does exist and does name those symbols — the CMake probe's positive half is TRUE today. What disarms the lane is the new negative half (the surviving stubs). SplitLane.h's SkipReasonForSplitOnlyAssertions() was not updated when the conjunction was added in CMake, so the message points an investigator at the wrong missing thing. The report's acceptance row ("The skip names MG_Remote/Client and the three packages, verbatim") quotes this as evidence of quality. The CMake message(STATUS …) is correct; only the runtime text is stale.

N-2 (MINOR). Three spellings of the same symbol list. CMakeLists probes kRemoteEmitSlotCount|RemoteEmitTable|CapsMirror; its own comment above says kRemoteEmitSlotCount|BackendObject_Remote|CapsMirror; SplitLane.h says kRemoteEmitSlotCount / BackendObject_Remote / CapsMirror. Exactly the drift this package is supposed to be against.

N-3 (MAJOR-adjacent, filed MINOR because the local gate covers it). The three-arm name+status identity is not checked in CI. integration-split's step "The same integration entries under monolith and under inproc" writes /tmp/names-split.txt, asserts only that the count is >= 1, and never reads the file again. It then runs both arms with --output-on-failure, under which a skipped test is not a failure. So the failure W-3 says is the one worth catching — "an inproc arm that SKIPPED forty entries the monolith arm ran" — is invisible to this job. Only wsl_p5_gate.sh compares JUnit name+status. Fix: move the gate script's statuses() diff into the CI step, or delete the dead file so nobody reads the step as doing the comparison.

N-4 (MINOR). github_split_matrix can silently become empty on a key typo, and a comment in test.yml claims it cannot. Measured against trace_cases.py: "ci": false → ValueError; ci_backends without DirectGLES → ValueError; "split": "yes" → ValueError; but "splitt": true (key typo) → loads clean, the split subset becomes [], the matrix becomes {"include":[]}, and GitHub skips retrace-split with no red. test.yml says "trace_cases.py refuses a split case that is not in CI or does not run DirectGLES, so the subset cannot silently become empty." Fix: assert the matrix is non-empty in the trace-cases job (one line), or reject unknown keys in the manifest.

N-5 (MINOR). 10 of the 11 Split. entries have no runtime evidence of anything. Only DirectGLES.Split.PersistentMapArm. carries a MOBILEGL_LOG_FILE_PATH (measured: 11 entries carry MGITEST_SPLIT_LANE=1, 1 of them carries a log path). ClearThenReadPixelsScenario and TriangleScenario contain no split-specific assertion at all — once armed, their green is "the same GL workload passed", nothing more. This is t1's debt §6.3 and it is the same gap as M-1.

N-6 (MINOR). The "monolith" counting lane is not pinned to monolith. DirectGLES.PersistentMapArm. names MOBILEGL_LOG_FILE_PATH but not MOBILEGL_TRANSPORT, so a job-level export reaches it. Observed: with MOBILEGL_TRANSPORT=inproc ctest -R '^DirectGLES\.PersistentMapArm', its own log carries Config: MOBILEGL_TRANSPORT=inproc. In the inproc arm of the gate's own A/B, therefore, both halves of E3(c)'s "pair" are inproc. Nothing breaks (each half asserts the constant kExpectedRoundtrips = 1), but the pair is not the pair the name and §2c item 4 describe. The inverse is correct and deliberate: the Split. entries do name MOBILEGL_TRANSPORT in a property and a job-level monolith cannot reach them (observed), and MOBILEGL_ITEST_REQUIRE_GPU is correctly left to the job env.

N-7 (MINOR). Both new scenarios hardcode attribute locations 0 and 1 against shaders with bare in vec2 aPos; in vec3 aColor; — no layout(location=), no glBindAttribLocation. Location assignment for unqualified inputs is implementation-defined. The house has both patterns already (VertexArrayEnableDisableScenario binds explicitly, others use layout qualifiers), so this is a portability foot-gun that will surface as a red on a driver whose linker orders differently — a red for a reason that is not a defect, in the two scenarios whose whole job is the opposite.

N-8 (MINOR). RESOURCE_LOCK on the two counting lanes currently guards nothing. Each lane's TEST_FILTER selects exactly one case and each has its own log path, so there is no second holder of either lock. Harmless and defensive; the measured "3 runs at -j 8 → 4/0/2 spurious failures" must have been measured on a different shape. I ran the full integration-gpu label at -j 8 five times (2 A/B arms + 3 dedicated runs) with 0 failures — see §4.

Q-1 (QUESTION). What can the E3 A/B still claim? t1 measures, and I did not disturb, that on llvmpipe the monolith arm is adopted (mpr=1 pmap=0.00) while R-6 pins split at emulated. Two different buffer code paths. The only thing the A/B can then claim is mpr equality — and t1 implements that as "both lanes assert the same literal 1", which is not a comparison at all: two lanes asserting a constant would both stay green if the true value changed to 2 in both arms only because they'd both fail, and would both stay green if the workload changed shape in a way that kept the constant true for the wrong reason. The pmap half is the only real cross-arm statement and it is one-sided (EXPECT_GT on the emulated lane, RecordProperty on the other). This is argued honestly in §2c/§3(c); it is a question for the integrator, not a defect: is an E3 whose two arms run different buffer paths still the gate ROADMAP intends, or does D-3 need overturning with a device measurement?

Q-2 (QUESTION). The last MGPipe stats: line in a lane log is a teardown window. The counting case reads LastFromLaneLog() before teardown, so it reads the right window — but the file ends with frames=2 window=0 draws=0 … pmap=0 … mpr=0 (observed). wsl_p5_gate.sh greps the whole file (grep -hoE 'mpr=[0-9]+|pmap=[0-9.]+' … | sort -u) and therefore reports the union of the real window and the teardown zeros. Is the gate's quoted mpr=1 pmap=0.00 reading the window line or the union? Worth pinning before a MEASUREMENTS row is written from it.

Q-3 (QUESTION). MOBILEGL_LOG_ACTIVE_LEVEL admitting INFO is a load-bearing precondition of every split lane and is enforced nowhere. It is passed by hand in build-linux-split and by the preamble's $COMMON. At WARN or above every split lane reds loudly (the message even names the variable); at DEBUG the run-level guard silently passes a monolith library (M-5). A one-line if (MOBILEGL_BUILD_DISAGGREGATED AND NOT MOBILEGL_LOG_ACTIVE_LEVEL STREQUAL …_INFO/…_DEBUG) message would make the dependency a build fact rather than documentation. Deliberate or oversight?


3. Gate table — one row per gate t1 added

# gate the mutation that must make it red observed red?
1 ctest -L integration-split --no-tests=error matches something in build-split run it in a build without -DMOBILEGL_BUILD_DISAGGREGATED YES — rc=8 in both build-push and build-linux
2 nm --defined-only build-split/libMobileGL.so | grep -i MG_Remote non-empty (build-linux-split, gate part 1, retrace-split swap) point it at a pull library YES — pull=0, push=0, verify=0, split=188; the guard's [ "$remote" -lt 1 ] branch fires
3 same nm on the pull library must be 0 (G1's split half) a leak of MG_Remote into the pull build not constructed; current reading 0
4 run_trace_case.cmake transport-resolution assertion on the SPLIT retrace replay with a pull library under MOBILEGL_TRANSPORT=inproc YES — ssim 1.000000 and then CMake Error … never reported resolving it (~/w7/p5-t1rev-pullneg.log)
4b …the same assertion, against a DEBUG-level pull log craft the Accepted env variable: line NO — it goes GREEN (M-5)
5 integration-split step "The split lane really resolved the transport" delete/skip DirectGLES.Split.PersistentMapArm., or a monolith library not run in CI here; by inspection it reds on a missing log and on a missing line, and shares M-5's substring weakness
6 the arming conjunction keeps the 11 Split. entries skipping while c0's stubs remain remove the stubs' text without implementing anything NO — 8 of 11 went GREEN on monolith (M-1); comment-only disarm also possible (M-2)
7 negative control E1 MOBILEGL_IPC_VERB_BARRIER=0 must red the split entries arm the lanes with nothing implemented YES, the control worked — it reported the 7 entries stayed green, i.e. ::error in CI
8 negative control E3(a) MOBILEGL_IPC_PERSISTENT_BLOCK_KB=0 same YES, the control worked — 3 entries stayed green → ::error
8b …the same two controls in wsl_p5_gate.sh give them a filter that matches nothing NO — they print "RED, as it must be" (M-4)
9 Split.PersistentMapArm.…TheMapLandsInTheArmItsLaneDeclares arm assertion (emulated) run the lane against the monolith path on a driver that adopts YES — declared Which is: "emulated" / std::string(arm) Which is: "adopted". The single strongest guard in the package, and driver-dependent by D-3's own argument
10 add_trace_replay_test(… TRANSPORT x) without VARIANT → configure FATAL_ERROR pass TRANSPORT alone by inspection only
11 trace_cases.py manifest validation of split ci:false / no DirectGLES / non-bool YES (3 of 3 raise ValueError); NO for an unknown-key typo → empty matrix, silent (N-4)
12 G14 (no ctest name ever disappears) / G2 (pull≡push) / G1 (symbols) — reproduced clean, not mutated
13 ARCHITECTURE.md:521 three-arm name+status identity make the inproc arm skip entries the monolith arm ran not checkable in CI — the names file is written and never read (N-3); only wsl_p5_gate.sh compares

5 of 13 rows were observed going red for the reason they exist (1, 2, 4, 9, 11-partial), plus two negative controls (7, 8) observed doing their job. Two rows were observed going green when they should have been red (4b, 6) and two more are structurally unable to fire (8b, 13).


4. What I verified and it held

Everything re-runnable in t1's own ~/w7/p5-t1/build-* trees, at 8d0850bd:

  • G1 symbol_report.py --threshold 0 vs ~/w7/p5-before-libMobileGL.so: .text 10806323 -> 10806323 (+0, +0.000%), 27811 -> 27811 defined symbols: 0 added, 0 removed, 0 resized, 0 renamed. Identical to the report.
  • G1's split half: nm --defined-only … | grep -ic MG_Remote = 0 / 0 / 0 / 188 for linux / push / verify / split; MGPipeApply* = 0 / 39 / 42 / 39.
  • ctest -L unit: 1785 / 1785 / 1785 / 1842. Exactly the preamble's baselines.
  • ctest -L integration-gpu in build-split: MOBILEGL_TRANSPORT=monolith 1139/1139, =inproc 1139/1139, JUnit name+status diff = 0 lines, 204 skipped in each arm.
  • ctest -N: linux 2913, push 2913, split 2981. G2 pull-vs-push diff 0 lines. G14 vs p5-before-ctest-names.txt (2902): 0 removed, 11 added — and I checked all 11 by hand: none is a rename of an existing name, all 11 are genuinely new ({DirectGLES,DirectVulkan}.{Triangle,PersistentCoherentMap}Scenario.* plus the single DirectGLES.PersistentMapArm. lane; there is no DirectVulkan.PersistentMapArm., deliberately asymmetric but harmless). Split-only additions are 11 DirectGLES.Split.* plus the 57 MG_Test/Wire cases, all inside if (MOBILEGL_BUILD_DISAGGREGATED), so G2 is untouched.
  • ctest -L integration-split in build-split: 11/11, all skipping, 22 Skipped markers.
  • Flake hunt: the full integration-gpu label at -j 8, five runs (two A/B arms + three dedicated): 1139/1139 every time, zero failures. No RESOURCE_LOCK-class flakes reproduced.
  • E2, the OpenRA inproc retrace, through the registered SPLIT variant in a -DMOBILEGL_BUILD_TRACE_REPLAY=ON split build: MobileGLTraceReplay.OpenRA.DirectGLES.SPLIT Passed 1.30 s, ssim=1.000000, and MGPipe split: OpenRA DirectGLES transport=inproc, Fatal{ lines in …/OpenRA/DirectGLES-SPLIT/output/mobilegl.log: 0. The census was read from the right file for the right arm, and the -SPLIT suffix really is carried by both directories (measured from the generated CTestTestfile: OpenRA/DirectGLES, OpenRA/DirectVulkan, OpenRA/DirectGLES-SPLIT; artifacts actual-images vs actual-images-SPLIT). D-5 lands as claimed.
  • The variant entry's labels really are two: generated as LABELS "retrace\;retrace-split", and ctest -N -L retrace-split selects exactly the one variant.
  • ctest ENVIRONMENT discipline. All three new env lists go through mgl_itest_join_environment and all three append ${MGL_ITEST_COMMON_ENV}; the generated property is one [==[…]==] blob with ;-joined entries, and the EGL vendor pin survives into the Split entries (measured in MobileGLIntegrationTest[4x]_tests.cmake). MGITEST_SPLIT_LANE=1 appears in exactly 11 generated entries and nowhere else. MOBILEGL_ITEST_REQUIRE_GPU is correctly not named in any new property, so CI's job-level export reaches the lanes.
  • Workflow. .github/workflows/test.yml parses (js-yaml 4.1.1, 20 jobs). Nothing t1 added is behind github.ref == 'refs/heads/feat/disaggregated' — the only such conditions in the file are the pre-existing monolith-symbol-report steps (:2202,2206,2235,2239). The three new jobs have no if: except retrace-split's needs.*.result guard, and their failures are visible as their own red jobs (there is no aggregate gate job to add them to). build-linux-split's configure flags are a faithful clone of build-linux-verify's (only -DMOBILEGL_PIPE_VERIFY ↔ the three disaggregated flags differ), and the nm step is present in the build job — the thing whose absence lets a split lane pass having run monolith.
  • Scenario shapes. glFlush() really is an empty body (MG_Impl/GLImpl/Exporting/Definitions.cpp:112 is one MGLOG_D and nothing else) and TriangleScenario correctly has none — glReadPixels is the ordering point. The draw is VBO-backed glDrawArrays with no client-array index pointer, so no MGHostSpan in the first frame. PersistentCoherentMapScenario's WriteThroughTheMapThenDraw really is std::memcpy + glDrawArrays with nothing between, the sequence really is map/write/draw/write/draw/readback, the second write is announced by nothing, and the second case repeats it across a Present. The static_assert that FLUSH_EXPLICIT is absent is real. The arm is read from IsBackendPersistentMapped() through Harness/PersistentMapPeek (b1's spelling) and re-asserted after the draws, not inferred from pmap — and CounterAsDoubleOrAbsent is the right fix: pmap genuinely lives in FormatWindowLine's bytes/f[…] bracket as FormatFixed2 (PipeStats.cpp:205), which strtoll would read as 0.
  • Commit hygiene. Four single-line commits, no bodies, Co-Authored-By count 0. Working tree clean. The p5-before-* snapshots exist and are consistent (2902 = 2913 − 11).

5. Verdict

Admit it, with M-1/M-2/M-3 and M-6 tracked as blocking the arming commit rather than this merge. Nothing here breaks a gate that is green today; every number in t1-v1.md that I could re-run came back identical, and the package's two headline discoveries (the symbol-only probe, and SSIM being useless for the split retrace) are correct and well evidenced. What it has not achieved is the thing it claims in §3(a): the replacement arming condition is not layout-independent, it is a string grep, and I made 8 lanes go green against monolith with one sed. The right fix is already written down as t1's own debt §6.3 — assert the resolved transport per case — and it should land before c1/s1/v1 can arm the lanes, not after. M-4 (the gate script's zero-selection control) and N-3 (the dead names file) are ten-minute fixes and should not wait.

Logs kept: ~/w7/p5-t1rev-mutA-isplit.log, ~/w7/p5-t1rev-pullneg.log, ~/w7/p5-t1rev-retrace-split.log, ~/w7/p5-t1rev-ab-{mono,inproc}.txt. Review worktree removed; ~/w7/p5-t1 left clean with its submodule symlinks off, as found.