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.
24 KiB
P2 package C — p2/espryt (Espryt 0b, the first Track H slice) — result v5 (rework round 5)
Tree ~/w7/p2-espryt, branch p2/espryt, base tag p2/contract (9c6a8a25), HEAD 0f29fa58.
Build directories: build-linux (pull, the G1 build), build-push, build-verify
(-DMOBILEGL_PIPE_VERIFY=ON -DMOBILEGL_ITEST_REQUIRE_GPU=ON) and build-nolegacy
(-DMOBILEGL_PIPE_PUSH=ON -DMOBILEGL_PIPE_LEGACY_MEMOS=OFF, untracked, not content).
git status --porcelain at the end of the round: only ?? build-nolegacy/.
The round-4 major (nine files outside C.5's ownership) is closed by waiver: ID-1 in
~/w7/notes/p2/INTEGRATOR-DECISIONS.md records the grant and ID-2 moves espryt before
tracker and magma in the integration order. Those edits are untouched in this round and are not
re-argued here. This round closes the six round-4 minors the rework instruction named as real
defects — (a) the two-holder hazard, (b) the caller-less "backstop", (c) the call-site pin,
(d) the armless case's fixed path and leaked env / config, (e) the sentinel's arming site, and
(f) HandleOf's negative memo — and re-runs the package's verification.
1. Commits
Two new on top of the thirteen from rounds 1–4 (15 total from p2/contract).
| sha | subject |
|---|---|
6507cfae |
[Fix, Test] (Espryt): deliver a death notice to every holder of a kind, retire the twin table's last sweep, and pin the bring-up call site under the armless knob pair — (a), (b), (c), (d), (e), (f) |
0f29fa58 |
[Test] (Espryt): pin that the armless cases put the operator's log path and MG_Config::Features back as they found them — (d)'s own gate |
Files touched this round: MG_Backend/DirectGLES/{SlotTables.h, Managers.h, Managers.cpp} and
MG_Test/SanityTest.cpp — all inside C.2's file list. No file outside C.2 + ID-1 was edited.
git diff --summary refs/tags/p2/contract..HEAD prints only the two create mode 100644 lines
(SlotTables.h, StateObjectDeathNotice.h) — no mode change.
git log refs/tags/p2/contract..HEAD --format=%B | grep -ci 'co-authored\|claude' → 0.
2. The six minors, one by one
(a) The two-holder hazard — FIXED (6507cfae)
Restated. OnFrontendStateObjectDestroyed told one registry per kind, and
BackendSlotTable::DestroyByLifetimeId returned false without freeing when this table did not
hold the slot. With CollectGarbageIfNeeded empty and CollectGarbageNow uncalled, any other
holder of the kind kept a live entry — and its twin's driver storage — for the life of the process.
The real second holder today is the ScopedDirectGLESTextureBindings fixture's saved copy of the
Texture registry. It is made worse, not better, by the allocator: MGPipeSlotAllocator::Free
erases the lifetimeId → slot mapping (SlotAllocator.cpp:113-131), so a notice delivered to
one holder and then resolved by the next would find nothing to resolve.
The fix — one notice, every holder, one free, last. BackendSlotTable links every instance
into a per-table-type intrusive doubly-linked holder list (constant-initialised head, two pointer
writes per link/unlink, no allocation; every constructor — default, copy, move — links, the
destructor unlinks, copy/move-assign carry entries and memo but never the links).
OnFrontendObjectDestroyed(lifetimeId) is static: it resolves the handle once through
FindByLifetimeId, calls the private ReleaseTwinAt(handle) on every holder (twin moved out of
the vector before its destructor runs, memo forgotten whenever it names that slot), and then calls
Free once. The slot is returned whether or not any holder still had an entry — the lifetime
id is dead and MG_State never re-issues one, so a slot that a table = {} reset had orphaned no
longer stays allocated for the life of the process (round-4 debt #2's test-only leak is gone too).
StateBackendObjectRegistry::DestroyByLifetimeId is static as well and forwards to it; the
dispatcher in Managers.cpp:184-214 keeps its per-kind spelling and says in its comment that
naming one instance is a spelling, not a choice of holder.
Magma and VertexElementsCso — a correction to the finding, stated rather than hidden. The
rework instruction says the kind "is shared with magma's m3 inside P2". I checked package D's
tree: grep -rn 'SlotAllocator\|MGPipeSlots' ~/w7/p2-magma/MobileGL/MG_Backend/DirectVulkan/ hits
only a comment (MagmaPipeArms.h:170, "nothing in P2 can call its Free"). Magma's
MagmaPipeIdentityTable mints out of its own per-renderer allocator with age-based
reclamation and never touches MGPipeSlots(). So the two backends share the name of the kind
and nothing else; there is no cross-backend holder to notify and no double-free ordering between
two consumers to design for. The holder list is per table type; each of the six kinds has exactly
one table type in this backend, and SlotTables.h's header says so. The hazard the review named
is real for the fixture's copy (and for any future second table of a kind) and that is what is
fixed and tested.
Tests. OneDeathNoticeDropsTheTwinInEveryHolderOfTheKind (kind Fence): two independent
tables plus a by-value copy of one hold the object, a fourth table holds a different object;
one notice while the object is still alive → all three FindByHandle null, both twins'
weak_ptrs expired, the kind's LiveCount down by exactly one, a second notice answers false,
the bystander untouched. ASavedCopyOfARealRegistryDropsTheTwinOnTheSameNotice does it through
the real TextureImpl::g_backendTextureObjects, a copy taken exactly as the fixture takes it,
and the real ~TextureObjectBase notice. AWholeTableSavesResetsAndRestores now also pins
HolderCount() across copy / reset / restore and that the saved copy drops the twin on the
object's death. Negative control NC8 (§3.4): release only the first holder → exactly those
three cases fail, naming the holder.
(b) The caller-less "backstop" — the claim and the code are DELETED (6507cfae)
ReclaimDeadSlots(), CollectGarbageNow(), CollectGarbageIfNeeded() and m_isCollecting leave
BackendSlotTable. grep -rn 'ReclaimDeadSlots' MobileGL/ → no hits. The header no longer claims
a backstop; it says what is true: a notice dropped by InProcessTeardown() leaves a twin that is a
deliberate leak (the process is exiting, the driver reclaims the object, and a twin destructor
must not call into a driver that may already be unloaded), not garbage awaiting a collection. The
per-entry weak_ptr stays for exactly one reason — ForEachLive()'s strong reference to the
callee — and the comment says so. StateBackendObjectRegistry::CollectGarbageNow is pre-P2 API
(p2/contract Managers.h:364, no caller then either) and is kept for the legacy arm; on the
handle arm it returns. CollectGarbageIfNeeded on the handle arm is the predicted branch plus a
return with nothing to forward to.
(c) The call-site pin — ADDED (6507cfae)
EglBringUpUnderTheArmlessKnobPairReturnsInsteadOfStopping runs the real bring-up entry point
under the pair. InitDisplayAndContext is file-static, but its two public callers are not;
InitPbufferSurface(1, 1) reaches it, and its twin-arm call is the first statement after
DestroyEGLContext(), ahead of eglGetDisplay. So in a forked death-test child the case sets
the knob pair, installs an EGL table whose eglGetDisplay answers EGL_NO_DISPLAY
(SetEGLFuncsTable), calls InitPbufferSurface, and std::exits with a code that says "returned
false". No display, no GPU, no skip: the predicate is ExitedWithCode(0x51). If the site is
edited back to ResolveEsprytSlotTablesArm() the child dies of SIGABRT and the case fails
with a message naming MOBILEGL_PIPE_PUSH / kMGPipeSubsystemEsprytSlots and
MOBILEGL_PIPE_LEGACY_MEMOS=0 — and it additionally reads the child's log and asserts the
diagnosis line is there and no Fatal{ is. It runs in all three push builds (in build-nolegacy
the verdict is Handles and bring-up must still return — it does not skip there) and skips only
in the pull build, where the twin table does not exist (G2 parity).
Negative control NC10 (§3.4) applied exactly that edit at DirectGLES.cpp:10238: the case fails
("Terminated by signal 6", both knobs named, the Fatal{PipeLegacyMemosDisabled …} line quoted)
while, in the same build, MOBILEGL_PIPE_PUSH=0 MOBILEGL_PIPE_LEGACY_MEMOS=0 ctest -L integration-gpu -R CrossFrameBufferScenario reports 100% tests passed, 0 tests failed out of 13, all
***Skipped — i.e. the lane hides precisely what the pin catches.
(d) The armless case's fixed path and leaked state — FIXED (6507cfae, gate in 0f29fa58)
Two RAII guards in SanityTest.cpp's anonymous namespace: ScopedLogFileRedirect (closes the log,
saves the operator's MOBILEGL_LOG_FILE_PATH — value or absence —, sets the redirect; on
destruction closes the log again, puts the previous value back or unsets, removes the file) and
ScopedArmlessKnobPair (saves and restores Features.PipePush / PipeLegacyMemos).
UniqueScratchLogPath() builds <tmp>/<stem>-<pid>-<counter>-<steady ticks>.log, so two
SanityTest processes on one host and two cases in one process cannot collide.
AnArmlessKnobCombinationStopsInsteadOfSkippingTheLane uses both guards; its ASSERT_EQ(…, NoArm)
now sits after the knob guard is constructed, so a failure there restores too.
TheArmlessCasesLeaveTheLogPathAndTheConfigAsTheyFoundThem (0f29fa58) pins the restore byte for
byte with a pre-set operator path and with none. ls /tmp | grep -c mobilegl-espryt after the
whole binary → 0.
(e) EnsureProcessTeardownSentinel — MOVED to the slot table's first insertion (6507cfae)
BackendSlotTable::GetOrCreate arms it on the non-null path before Acquire; on the handle arm
StateBackendObjectRegistry::GetOrCreate dispatches to the table before its own arming call, so
the registry arms only on the legacy arm. D13's sentence is now literally true. Round-4 §6.9's
objection — "a BackendSlotTable used outside the registry would arm it" — was a reason not to
bother, not a reason it is wrong: the atexit handler only latches g_processTeardown, which no
unit binary reads to any effect. Pull-build codegen is unchanged (the dispatch is under
#if MOBILEGL_PIPE_PUSH; in the pull build the sequence is still assert → sentinel → map), and
G1 confirms it: contract → HEAD 0 added, 0 removed, 0 resized, 0 renamed, .text +0.
SlotTables.h re-declares the function because it is included from Managers.h above line 61.
(f) HandleOf's negative memo — REMOVED (6507cfae)
if (!MGPipeHandleIsNull(handle)) RememberHandle(…). The reason is the one the review gave: the
memo is per table, the allocator per kind, so once another holder can be the one that acquires, a
cached "no handle" here would outlive the twin's creation over there and nothing on this table's
acquire path would ever refresh it. ReleaseTwinAt also forgets the memo whenever it names the
freed slot, before checking whether this table has an entry there, so a positive memo learned
from the allocator for another holder's twin cannot survive the slot's next handout either.
RepeatedLookupsOfALiveObjectKeepOneHandle lost the sentence that pinned the old behaviour;
ANegativeLookupIsNotCachedAcrossAnotherHoldersAcquire is the new gate (first table asks and
misses, second table acquires, first table must now answer the same non-null handle).
Negative control NC9 (§3.4): memoise unconditionally again → exactly that case fails.
3. Verification — every command and what it printed
All from ~/w7/p2-espryt at 0f29fa58 (the test-only second commit; the library gates were run at
6507cfae and the pull library is byte-identical between the two — §3.1 last row),
CCACHE_BASEDIR=/home/swung/w7, with MOBILEGL_BACKEND_TYPE=DirectGLES and
__EGL_VENDOR_LIBRARY_FILENAMES=/usr/share/glvnd/egl_vendor.d/50_mesa.json for the GPU lanes.
Summary of the long lane: ~/w7/p2r5-espryt-summary.txt.
3.1 Build and static gates
| gate | command | result |
|---|---|---|
| build | cmake --build {build-linux,build-verify,build-nolegacy,build-push} -j 12 |
all rc=0, 0 warnings in the push build log |
| G1 | symbol_report.py --before ~/w7/p2-before-libMobileGL.so --after build-linux/libMobileGL.so --threshold 0 |
27799 -> 27799 defined symbols: 0 added, 0 removed, 4 resized, 0 renamed, .text +160 — the four are ID-3's admitted set |
| G1 attribution | same, --before ~/w7/p2-contract/build-linux/libMobileGL.so |
0 added, 0 removed, 0 resized, 0 renamed, .text 10792739 -> 10792739 (+0) — re-run after 0f29fa58: identical |
| G5 | git show {48268068, refs/tags/p2/contract, HEAD}:…/DirectGLES.cpp | awk '/namespace RenderStateImpl {/,/} \/\/ namespace RenderStateImpl/' | sha256sum |
all three d8fd1c48716056c5…, equal to ~/w7/p2-before-syncrenderstate.sha |
| G13 | grep -rc 'pGLContext' MobileGL/MG_Backend | grep -v ':0$' |
empty, rc=1 |
| G13 | python3 scripts/check_include_closure.py |
4 probes, 0 skipped, 0 problem(s), rc=0 |
| G13 | gen_pipe.py --check / --self-test |
rc=0 / 7 negative-control trip(s), positive control OK |
| G13 stdio | the pipe-gates alternation (test.yml:1515-1522) over MG_Backend + MG_State |
0 hits |
| C.2 grep | grep -n 'OwnerEquals|TwinLookupMemo|g_fbSlotCache|CollectGarbageIfNeeded' DirectGLES.cpp |
37 hits, unchanged from v4 (DirectGLES.cpp is not touched this round); all inside MOBILEGL_PIPE_LEGACY_MEMOS arms as verified in rounds 3–4 |
| G2 | ID-4's ctest -N extraction × 4, LC_ALL=C sort, diff |
build-linux 2385, build-push 2385, build-nolegacy 2385 — both diffs IDENTICAL. build-verify 3203 (its verify-only lanes) |
| G14 | LC_ALL=C comm -23 <sorted ~/w7/p2-before-ctest-names.txt> <pull names> |
0 removed. Added 22: the contract's 4 placeholders + 18 DirectGLESSlotTable.* |
3.2 Tests
| lane | command | result |
|---|---|---|
| unit × 4 | ctest --test-dir <d> -L unit --no-tests=error -j 6 |
100% tests passed, 0 failed out of 1507 on build-linux, build-push, build-verify, build-nolegacy |
| slot × 4 | ctest --test-dir <d> -R DirectGLESSlotTable --no-tests=error |
18/18 in all four; 0 skipped in build-push / build-verify; 2 skipped in build-nolegacy (AnArmlessKnobCombination… only — the bring-up pin runs there); all 18 present and skipping in build-linux |
| sanity, handle arm | ./build-push/MobileGL/MG_Test/SanityTest |
[ PASSED ] 100 tests |
| sanity, legacy arm | MOBILEGL_PIPE_PUSH=0 …/SanityTest |
[ PASSED ] 96, 4 skipped (the four that are about the handle arm, incl. the new ASavedCopyOfARealRegistry…) |
| sanity, order independence | --gtest_shuffle --gtest_random_seed={1,2,3} |
100/100 × 3 |
| armless hygiene | whole binary, then ls /tmp | grep -c mobilegl-espryt |
0 |
| integration, handle arm | ctest --test-dir build-push -L integration-gpu -j 4 -R DirectGLES |
446/446, 100 (Skipped) |
| integration, legacy arm | same with MOBILEGL_PIPE_PUSH=0 |
446/446, the same 100 skips |
| integration, pull build | ctest --test-dir build-linux -L integration-gpu -j 4 -R DirectGLES |
first run 444/446 (see §5.1), rerun of the two alone 2/2, full lane rerun 446/446 |
| integration, no legacy compiled | ctest --test-dir build-nolegacy … |
446/446 |
| integration, verify build | ctest --test-dir build-verify -L integration-gpu -j 4 -R DirectGLES |
855/855 |
| G4 (a) | ctest --test-dir build-verify -L integration-verify --no-tests=error -j 4 |
818/818; grep -c 'Fatal{' over the 10 lane logs = 0 each |
| armless red proof | MOBILEGL_PIPE_PUSH=0 MOBILEGL_PIPE_LEGACY_MEMOS=0 ctest --test-dir build-push -L integration-gpu -R '^DirectGLES\.CrossFrameBufferScenario' |
0% tests passed, 13 tests failed out of 13, 13 × Subprocess aborted (log ~/w7/p2r5-espryt-armless.log) |
| G3 (DirectGLES half) | retrace_gate.py --tree ~/w7/p2-espryt --lib $PWD/build-push/libMobileGL.so --out ~/w7/retrace-out/v5-espryt -j 4 --only 'DirectGLES$' |
rc=0, passed 39 / 39; failed: [] |
| G4 (b) | MOBILEGL_PIPE_VERIFY=1 retrace_gate.py … --lib $PWD/build-verify/libMobileGL.so --out ~/w7/retrace-out/v5-espryt-verify … |
rc=0, 39/39; of the 39 */output/mobilegl.log: 0 Fatal{, 0 PipeVerifyDiffer|UnmigratedPipeInput|PipeResidualDiverged, 0 unarmed (grep -L 'MGPipe: verify armed') |
The two retrace output trees (1.4 GB each) were deleted after the counts above were taken.
3.3 Negative controls — each new gate goes red for the reason it exists
Each perturbation applied alone, built, run, restored from a /home/swung/w7 copy (no git stash),
then a clean rebuild and 18/18; git status --porcelain afterwards only ?? build-nolegacy/.
Log: ~/w7/p2r5-espryt-nc.log.
| # | perturbation | result |
|---|---|---|
| NC8 | OnFrontendObjectDestroyed releases only the first holder (holder = nullptr after one) |
only OneDeathNoticeDropsTheTwinInEveryHolderOfTheKind ("holder a kept the twin", "holder b kept the twin"), AWholeTableSavesResetsAndRestores and ASavedCopyOfARealRegistryDropsTheTwinOnTheSameNotice fail; 14 pass |
| NC9 | HandleOf memoises unconditionally again |
only ANegativeLookupIsNotCachedAcrossAnotherHoldersAcquire fails ("the first holder kept answering the null handle it cached before the second holder acquired"); 16 pass |
| NC10 | DirectGLES.cpp:10238 → (void)ResolveEsprytSlotTablesArm(); |
only EglBringUpUnderTheArmlessKnobPairReturnsInsteadOfStopping fails: Result: died but not with expected exit code: Terminated by signal 6, the message names both knobs, and the log quote is the Fatal{PipeLegacyMemosDisabled, …} line. In the same build the integration lane under the pair prints 100% tests passed, 0 tests failed out of 13, all ***Skipped |
(NC5–NC7 from v4 were not re-run; the classifier, the per-kind walk and the destructor notices are unchanged this round except that the walk's death step now goes through the static notice.)
3.4 Gates this package cannot reach from its own tree
Unchanged from v4 §3.5: G6/G7 (package A's spans, E's g7_negative_control.sh), G8
(HandleRecycleScenario, E), G9 (gen_pipe_dirty_surface.py --check), G10, G11, G12, G14's
CI-matrix half. DriverBench and the two-device runs (D.4.2/D.4.3) remain owed; this round's
per-draw delta is a removal (the registry's handle-arm CollectGarbageIfNeeded no longer makes
a call at all) plus one MGPipeHandleIsNull test on HandleOf's miss path.
4. Deviations from the brief and from the rework instruction
- ID-1 / ID-2 — the six destructor sites stay in the B/D-owned files under the recorded waiver; not re-argued.
- The "shared with magma's m3" premise of (a) is not what D's tree does (§2(a)): Magma
never calls
MGPipeSlots(). The fix is made and tested for the holders that do exist (the fixture's copy, any future second table of a kind); no cross-backend consumer ordering was built because there is nothing to order. - (c) is a unit-binary case, not a ctest lane entry. C.2 owns no CMake and E owns
MG_IntegrationTest/**, so the "lane-level" check is aSanityTestcase that runs the real bring-up entry point under the pair in a forked child. It never skips in a push build and it cannot skip on the pair by construction (no GPU is needed); it skips only in the pull build, where the table does not exist, for G2 parity. - (e) moved as D13 says; the objection recorded in v4 §6.9 is withdrawn (§2(e)).
- Carried unchanged from v4 §4: 4.2 (first-use Fatal instead of startup Fatal), 4.3 (GC deleted on the handle arm only), 4.4 (e2-before-e3 satisfied in content, ordering deviation on squash-merge), 4.5.
5. Where the tree contradicts the brief, or is flaky
5.1 NEW — two E-owned integration lanes are flaky under -j 4 in the pull build
First full pull-lane run: DirectGLES.UnlocatedIoBlocks.UnlocatedIoBlockScenario.TheEmulationIsActuallyArmedWhenTheEnvironmentPinsItOn
and DirectGLES.PointSizeDemotion.PointSizeDemotionScenario.TheDemotionIsActuallyArmedWhenTheEnvironmentPinsItOn
***Failed (~/w7/p2r5-espryt-it-linux.log); both pass alone and the full rerun is 446/446
(~/w7/p2r5-espryt-it-linux2.log). Both passed in all four other lane runs this round. The pull
library is byte-identical to the contract's (G1), so this is not package C's code. Likely cause,
for E: MG_IntegrationTest/CMakeLists.txt:368-387 gives each of those lanes one fixed
MOBILEGL_LOG_FILE_PATH (unlocated-io-blocks.log, point-size-demotion-gles.log), the
arming assertion reads that file, and under -j 4 the lane's other cases open the same file with
"w" in parallel — the same fixed-path collision shape as this package's review minor 5.
5.2 Carried from v4 §5
retrace_gate.py has no --backend / --ssim (ID-4 records it); G4's MGPipe verify: grep
and <out>/*/mobilegl.log glob are vacuous as written (ID-4 records the arming line; the path is
<out>/<case>/<backend>/output/mobilegl.log); _GLOBAL__sub_I_DirectGLES.cpp is in ID-3's
admitted set; section A's ^\s+Test # drops four-digit ids (ID-4); git stash -u destroys the
copied fixtures; p2/contract is an ambiguous refname here (the tag wins).
6. Unfinished, and debts carried
- The death-notice ops table is still a single global pointer
(
StateObjectDeathNotice.h). No second consumer exists in P2 (§2(a): Magma has its own allocator and no notice), so this is a P3 precondition, not a P2 defect; the fix is a small fixed-size consumer list. Still the top cross-package item. - Slot reuse between a table's reset and restore (v4 debt #2) is now moot in its leaking form — a dead object's slot is returned whether or not a holder has an entry — but the semantic hazard (a restore reinstating an entry for a slot the allocator re-handed while it was reset) is still only avoided because nothing acquires between the fixture's reset and restore. A refcounted ownership token is still the right shape if that ever changes.
- Debt #3 (two holders) is closed by the holder list; what remains is that the list is per table type, so a second type of the same kind in this backend would need to join it.
- The backend still mints client handles (
SlotTables.hDEBT block; P3+, belongs inMEASUREMENTS.md). - The one-entry resolution memo is the wrong shape for
BindCurrentFBOandResolveUnitSamplerBackend(per-unit / per-target memo is the fix; G11 prices it). - Device measurement (D.4.2 / D.4.3, G11) and DriverBench are owed.
- G6–G12 not exercisable from this tree.
- No
PipeStatscounter for this subsystem (package A's file). - Round-4 minors left open on purpose: minor 3 (
build-nolegacystill instantiates the legacyCollectGarbage()— nothing ticks it; the sentence in v4 was wrong, the mechanism is not), minor 6 (the D13 "must not break" pins make no acquisitions; coverage lives in the per-kind walk), minor 7 (the populated fixture save/reset/restore is covered byAWholeTableSavesResetsAndRestoreson a fake table, now including the death path, not by the fixture itself), minor 10 (the registry still asserts non-null ahead of the arm dispatch), minor 11 (the out-of-line destructors are push-only), minor 13 (kMGPipeSubsystemsMigratedAtP2duplicated in the test environment), minor 14 (device measurement).
7. Logs kept under ~/w7 (everything else of rounds 1–5 deleted)
p2r5-espryt-summary.txt (the long lane, verbatim), p2r5-espryt-nc.log (NC8–NC10 builds),
p2r5-espryt-armless.log (the red proof), p2r5-espryt-it-linux.log / p2r5-espryt-it-linux2.log
(§5.1). Deleted this round: armless.log, all e2-* and e3-*, p2-espryt-*.log,
p2-espryt-verify/, rev-espryt/, both retrace-out/v5-espryt* trees, and every other
p2r5-* file this package wrote. rt-restore2-build.log (487 bytes, 2026-09-06, a test-target
build log of uncertain origin) was left in place. Other packages' p2r5-* files were not touched.