Repository navigation
AGC: buffer descriptors bypass the resident native MSAA guest-access refusal #6480
Description
2 November fixture clarification
#6498 merged the original native-buffer refusal fix; this issue stays closed for that scope. A later intermittent 'recipe dispatch: missing refusal' assertion is now diagnosed as a fixture handling error: DispatchRecipe returned Rebuild before recording, but the fixture ignored it rather than following the real driver's ordinary dispatch path.
#14961 is the independent test-only correction. Exact main/diagnostic/fix evidence: main 926f7dd reproduces the assertion; an unchanged-production print-only diagnostic captures Rebuild; the correction passes 14 runs across RTX 5080/NVIDIA 621.01 and Radeon Graphics/AMD 26.10.1 with loaded validation and no VUIDs/skips/timeouts. The rebuilt path executes and reaches the required refusal. Recorded calls still must refuse.
This captured case does not establish a new production guard bypass or a reason to reopen the original runtime issue. Earlier references to an unresolved recipe bypass in composed-branch checks are superseded by this diagnosis, while their observed test counts remain source-pinned history. No new title result or publication capability is claimed. AI-assisted: yes (Codex).
Original report and earlier source-pinned validation
Maintenance qualified — 28October2026: b2aad7239cb4 refreshes to tested maind4cb89300fe54cce0ad74edf67803086e302b819,including merged #11369. Published history/authorship and all own topic lines preserved;no force-push. The new depth-buffer upload guard and this color-buffer guard are both present.
Fresh main/control/fix qualification:60positive processes(56physicalGPU/4CPU),six intended alias-control failures. Main/candidate each30/30;merged native-depth buffer-write cases3/vendor cover all D16/D32 x2/4/8 rows,depth/HTILE preservation and adjacent buffers. Candidate color aliases refuse6/6 modes while adjacent/deferred/recipe controls pass. Full native/relinker/real-fixture builds,conventions,stable hashes;zero validation warnings/errors/VUIDs/skips/timeouts,max6.450s. RTX5080/621.01,Radeon Graphics1002:13c0/26.10.1,Windows11 Enterprise26200,loaded Khronos1.4.363;required Vulkan/cache off. Advisory notices disclosed;no fresh title/full-suite/Linux or combined heap/fault-arena claim.
#11369 is merged;Depends on remains none. #6498→#8963 agreed;#6498→#9508 proposed. Actual-head maintainer Build/title gates remain. AI-assisted:yes(Codex).
Publication stack status,28 October2026: five submission heads refreshed parent-first onto tested main3a40945e62b7e2543031d5da2b8e351ab603bc1a;history/authorship and own implementation/test deltas preserved. Current seven-source report:132positive processes and10actual host-only-owner sampling failures;#8223-first and the fork-only initialization/publication corrections remain. All native/relinker/real-fixture builds and conventions pass,zero validation diagnostics/skips/timeouts. #7002 still depends on #6498. Latest-main source previews clean;no final passing standalone/full-stack gate or new title claim.
Separate corrected proposal,26 October2026: Current source,complete synthetic logs,retained failure and analyzer scope. Parent-first normal merges preserve original author,published history and CPU-pin ancestry;own changes remain identical. Both-vendor presentation host-memory failure injection executes on all three main5b85 source pins;latestmain1c90's retired-import rewatch control passes in the original mocked fixture. Integration volume warning counts remain explicitly preserved at earlier pins. No production guard relaxation or new title claim. Earlier actual failures/controls keep their exact historical pins.
Reproduction
A public synthetic regression derived from tests/execution/SampleMultisampleTarget.cpp creates a 16x16 RGBA16F, four-sample native color target, draws into it, then applies a recognized uniform color-memory fill. MIMG reads verify all 1024 native samples equal (1,1,1,1). RequireNoMultisampleTarget(address, 128, ...) confirms this same range is guarded with “whose samples are only on the host”.
A real recompiled tbuffer_load_format_x (32_UINT, 32 lanes), bound through a V# over the first 128 bytes of that range and followed by a buffer store to a separate output, nevertheless dispatches successfully and reads 32 zero DWORDs from stale guest memory. Expected: retain the existing explicit unsupported-resident-access refusal until native MSAA buffer publication is implemented.
Production AGC source is unchanged upstream main 0029e62f974c453a5f6c7333f3bdb13e770e3903. Local diagnostic head 692c29a4fb16daa941b5853cdfbb7fac4f3683f2 changes test code only (plus the independent Windows test-build correction in #6452).
Hardware and controls
Windows 11 Enterprise 10.0.26200, WinLibs GCC 15.2r7, Release, Vulkan required, shader disk cache disabled:
| Physical GPU (verified in each run) | Driver | Resident alias | Nonresident kernel | Original MSAA fixture |
|---|---|---|---|---|
NVIDIA GeForce RTX 5080 (10de:2c02) |
32.0.16.1062 | 5/5 reproduce | 5/5 pass | 5/5 pass |
AMD Radeon(TM) Graphics (1002:13c0) |
32.0.21036.18 | 5/5 reproduce | 5/5 pass | 5/5 pass |
All are fresh processes and finish below 30 seconds; maximum across these 30 runs: 0.588 seconds. Resident runs exit 1 on the intentional missing-refusal assertion, without a device loss or timeout. The nonresident control executes the same kernel, verifies its 32 zero results and unchanged surrounding output.
An earlier diagnostic harness waited on allocation removal because imported memory was released before device/import teardown; that harness bug was corrected before these measurements. It is not a GPU hang finding.
Scope and proposed correction
ShaderResources::addGuestBuffer checks overlap with the current draw attachment, but compute V# bindings over a previously rendered native target have no equivalent resident-image guard. Validate logical buffer descriptor ranges, including reuse, without rejecting adjacent logical ranges merely because a larger host-import window overlaps the native target. An explicit refusal is the proposed bounded correction; native tiled/sample-preserving buffer access remains unimplemented.
Related #6475 restores the guard for shader captures and explicitly excludes buffer uploads/input snapshots; this report concerns actual GPU buffer descriptor access. #6175 concerns pass hazard ordering, a different behavior.
This was found during the PPSA99902 bring-up investigation, but is a synthetic confirmed gap, not a newly reproduced title stop or proof of the cause of its historical failure. No new title execution, FPS, PS5 console equivalence or Linux/lavapipe claim. AI-assisted investigation/report: Codex.
Focused buffer-guard fix: latest rebase (2026-10-19)
#6498 is rebased onto main 4ffb0cd12b at e15c428b4d. Guard and fixture additions are unchanged; the documentation conflict is resolved while retaining upstream's retirement description. Fresh Windows relinker/library builds and 39 focused NVIDIA/AMD/CPU passes succeed, with no skips/timeouts or logged Vulkan validation warnings/errors. Main9ca regression controls and prior Linux validation retain their original pins in the PR description. This maintenance rebase adds no title-unblock claim; experimental buffer publication remains separate.
Current focused publication stack (2026-10-18)
The four conflicted PRs are updated on the fork and merge cleanly with inspected main 9f25d0bf5b. Test source is pinned to main 9caecfd2d5 plus each declared dependency stack. The intervening polygon-primitive change does not alter these publication functions or triangle fixtures; its new behavior was not rerun here.
| PR | Published head | Positive executions | Depends on |
|---|---|---|---|
| #6498 | 103e467afc |
60 selected +6 restored | none |
| #6989 | 7c92904c1c |
60 | #6987 and prerequisites |
| #6995 | 3ce433052f |
69 | #6989 |
| #7002 | 53282f62c2 |
96 | #6995, #6498 |
Windows/Linux relinker and system-library builds pass for every resulting head. The selected matrices cover RTX5080, Radeon Graphics, Linux/lavapipe and CPU tests, with verified physical device identities, no whole-test skips/timeouts and zero logged Vulkan validation errors/warnings. Existing capability-dependent subcases retain their fixture limits. A fresh test-only #6498 control with unchanged main9ca production code still misses all six refusals in3/3 runs per physical GPU; adjacent/deferred/recipe controls pass. The exact PR descriptions retain group lists, old fault controls, analyzer advisories and source limits. All four current upstream Conventions checks pass; the maintainer's #6498 rebase request has a direct reply for the #6445 game gate.
#7002 now includes the exact updated #6995/#6498 heads. Its combined fixture consolidates duplicate helpers, keeps all six unsupported-layout refusal cases, and tests supported R_X reads through actual publication. Writable aliases remain refused before read-only publication; binding stays a strict check. Upstream's movedOnly argument and BGRA multisample coverage are retained.
Current completed title evidence: integration e4c preserves all45/8d4 bring-up ancestry on main9ca. Ordinary RTX5080 cache replays record60 writes followed by56 hits and a pipeline-cache reload; #7452/#7462 track that separate correction. The fuller earlier8d4 replay verified movement,12 rings and tutorial continuation. #7260 holds the exact pins and limits. These focused PR tests do not establish new whole-title compatibility, FPS or completion.
This stack concerns uncompressed single-layer R_X color publication. Native depth/Z_X guest commit, HTILE publication, BDA and fixed-function consumer publication remain separate; the #6570 handoff makes no Onimusha compatibility claim. Original dependency authorship, historical branches and runtime archives are preserved. AI-assisted: yes, Codex.
Earlier integration notes at their recorded pins
Current focused publication stack, 2026-10-19
All five focused branches now incorporate main's DCC-refill handling at 1e774a3b22. Each has full default and explicit libs/relinker Windows builds and repeated focused checks on the actual RTX 5080 and Radeon Graphics. Exact dependency pins, drivers, diagnostics and controls are maintained in the PR bodies.
| PR, in dependency order | Qualified fork head | Passing checks |
|---|---|---|
| #6973 | cafd4f749f |
72 |
| #6987 | c4bd172d57 |
72 |
| #6989 | 3e728aadc5 |
75 |
| #6995 | 7b090b67d0 |
81 |
| #7002 | 0cdc81ae9f |
99 |
Each head additionally reproduces the same nine AMD 2x/4x/8x DCC assertion failures as literal main 1e774a3b22 (#7936). These are failures, not passes. NVIDIA completes the refill checks; AMD fails earlier in the clear checks, so AMD refill-assertion coverage is not claimed. Selected logs have zero VUIDs/validation errors/warnings, unsupported subcases, whole-test skips or timeouts; the requested Khronos layer is confirmed loaded. Later main merge previews are not tested compositions.
#6987 adapts one private checker call: ReadSamples requests full color/DCC/CMASK validation after synchronization with checkDccCurrent(target, false). Refill handling retains main's narrower check immediately after a verified metadata refill. The other three children's own patches are unchanged. #7002 retains exact credited #6498 e15c428b4d ancestry and validates all writable bindings before any read-only publication. Historical dependency and reversed-order controls retain their original tested pins; they were not rerun or relabeled as current-main results.
Title boundary: this refresh is synthetic-fixture validation, not a new Sonic launch. #7260 maintains the latest ordinary run (menu/movie/level loading, then a Mesh-preparation unreadable-memory stop) and preserved historical daytime gameplay. Integration f68ea8a529, historical best 0e7fba38, original dumps, saves and prior fixes are retained. Native-owner FMASK work is separate in #7948. Compressed DCC/CMASK publication, general image aliases, metadata/depth publication, PM4 selection, fixed-function vertex/index and dynamic BDA remain separate. No new title/FPS, unchanged peer-branch, Linux/Intel, Mirage or Onimusha result is claimed. AI-assisted: Codex; author responsible.
Earlier foundation validation — original 2026-10-17 pins
Submitted stack: initializer #6973 → snapshot #6987 → explicit publisher #6989 → CPU reads #6995 → read-only shader buffers #7002. The last also depends on the exact published #6498 buffer guard. PR bodies maintain their own dependencies, source pins and detailed matrices; automatic buffer wiring is now submitted.
#7002 validation: five own files, +439/-94. 81 successful synthetic processes across three explicit pins: 57 in the original matrix at 066f52769, 12 after the dominator-bitset main update at 2835de774, and 12 after the sampler/runtime-ABI update at submitted head 3d457b9a8, main 29bfe31dd. Both OS production/fixture builds pass. NVIDIA RTX5080, AMD Radeon Graphics and llvmpipe are verified per process. Core-layer runs pass without validation errors; project shader cache is disabled. The earlier controls retain their original source pins.
- 18 pre-feature failures: 1e823c956 restores pre-feature ShaderResources source/header. Compute publication and device-draw modes each refuse the host-only alias three times per device. Restoration passes.
- Nine proposal-ordering failures: 7698eb434 proves the previous experimental single-pass loop could publish a read-only source first and then prepare an overlapping writable binding without refusal. The probe only prepares a non-null resource plan; it executes no overlapping writes. Correction 066f52769 checks every writable binding before publishing any read-only binding. The regression verifies refusal, complete guest-byte preservation and native-owner retention. This corrects our fork proposal, not a newly asserted main defect.
The caller tests cover fresh, warmed, edge, deferred and recipe dispatches; unchanged-input draw-recipe replay; and moved-fragment ordinary Draw retaining its cached template. Actual RGBA32F pixels distinguish original/current/moved words, including 0x3c003800 and 0x34003a00; source samples/padding, clear/covered pixels, output neighbors and the unused old guard are checked. Six unsupported-layout refusal cases remain, including adjacent/deferred/recipe controls. Their neighbor starts after the conservative guarded extent. Binding stays a strict assertion without waits/submits. New buffer fixtures use 4x RGBA16F; they do not establish every format/sample configuration.
Current experimental integration: f52b77371, branch, includes main 4e15bde13 and the writable-first correction. Both OS builds pass; buffer publication, unsupported refusal, direct draw and the retained older manual raster probe pass once per verified adapter: 12 smoke processes. These are separate integration results. All historical checkpoints remain preserved.
The submitted analysis pin retains two Cppcheck and two clang-tidy advisories in source contexts unchanged from the prerequisite baseline; 40.963 s analysis, no parser/configuration failures, suppressions or automatic fixes. Conventions/registration pass; no full-suite or clean whole-project claim. Hardware: RTX5080 driver32.0.16.1062 and Radeon Graphics32.0.21036.18 on Windows11 Enterprise26200/GCC15.2r7; Ubuntu24.04 WSL/GCC13.3, Mesa25.2.8 llvmpipe LLVM20.1.2. Core validation is not synchronization validation.
Earlier evidence: #6987 retains 49 passes/nine order controls; #6989 19 passes/nine stale-data controls; #6995 31 passes/nine prior-guard controls. Older CPU/compute 401835f1b, manual-raster 3da1cf9c3 and direct-draw 45406ab5c matrices retain their original pins. CPU coverage includes hardware 2x/4x/8x and XOR0/0x8100; software excludes unsupported2x/8x subcases. These are not relabeled as latest-main reruns.
PM4 draw-key/variant selection, fixed-function vertex/index, dynamic BDA, combined #6947 heap/space reuse, general metadata/depth publication, console equivalence, unsynchronized concurrent writes and large-title costs remain unverified or unimplemented. #6947 coordination preserves the handoff before its adjacent early-return edit. No fresh game startup/FPS or SONIC SUPERSTARS/Onimusha validation is claimed. AI-assisted: Codex; author responsible.
Earlier integration validation at its original source pins
Current retained integration
Integration e86150f84 includes main 82ddae75b6761757f3d3ba9c0723647469270f76 and preserves prior validated 65d979b03. Canonical codex/integration-6060-floor290 and checkpoint codex/integration-main82-public-validation point here. The merge keeps the audited integration-only journal State size of768; its codec otherwise matches main. Existing color metadata exclusion assertions and #6789's immutable test-data storage correction remain.
Affected-path validation: six selected targets and their affected libraries build on each OS. 32 Windows and21 Linux fresh CTest processes pass, maximum6.80s, no whole-test skips/timeouts. Conventions and test registration pass. This is a selected build/check campaign, not a new full application or whole-suite result.
- CPU, each OS: constant-bit folding3/3, shader-memory3/3, graphics3/3, journal1/1. This covers main's vertex lane-bound folding change, including the preserved ballot-word form; graphics still passes with test(agc): keep the vertex terminate case's code in static storage #6789.
- Runtime, each actual NVIDIA/AMD GPU and llvmpipe: lazy shader preparation3/3, synchronous-lazy3/3, eager preparation1/1, signed image stores3/3, and retained MSAA sampled-publication control1/1. The changed SINT fixture checks D16 stores across eight formats with full/sparse masks and after a32-bit atomic. These are host/emulator fixture results, not new console measurements.
- Actual selected device identity is checked in every runtime process. Windows11 Enterprise26200 / WinLibs GCC15.2r7; NVIDIA RTX5080 driver610.62/32.0.16.1062 and AMD Radeon Graphics32.0.21036.18. Ubuntu24.04/WSL2 / GCC13.3 / Mesa25.2.8 llvmpipe. Project shader cache disabled; driver caches uncontrolled. Linux heap-mirror cases are unavailable and excluded; software MSAA publication checks16 configurations, with32 unavailable and excluded.
Main's lazy registration, constant folding, SINT image-store and bounded exit-logging changes determined this selection. Unrelated clear/alias/relocation matrices were not repeated. No new title run, gameplay or FPS claim; the historical title frontier remains unchanged. This broad fork remains experimental. Separate array-descriptor evidence stays in our #6635 review. AI-assisted: Codex.
Previous main65 evidence, retained at its tested pins: 2a476ed69 passed47 Windows processes, including twelve native-clear processes across default/opt-in modes and both physical GPUs; each checks RGBA8/RGBA16F at2x/4x/8x, four known DCC clears, bounds and the metadata-write refusal. Commit-item eligibility for each clear was not asserted. After merged #6784's journal correction, 5e69b9642 passed10 Windows and21 Linux processes with one separately reproduced CPU graphics failure (#6792). Exact #6789 head b88591fd4 then passed Linux graphics3/3, against main3/3 failures; integration 65d979b03 passed graphics3/3 per OS. The underlying compiler/storage explanation remains unconfirmed. Earlier full builds/static findings retain their original pins below.
Earlier validation and retained evidence at their original source pins
Previous main7dc checkpoint: 3435e26dc built affected libraries/six targets, passed 45 Windows and 27 Linux repetitions (six GPU groups/vendor and three CPU groups x3), maximum18.260s. Software publication checked16/48 configurations; heap-mirror and kept-highest-index omissions were excluded. The earlier full-build/static evidence below retains its original pins.
Current retained integration
Integration ccd37c17f includes main 72c5eed1ff43c4e70275561818427714a9122dbd and preserves previous canonical eb3c113a6, the fully built main332 checkpoint, and the historical bring-up checkpoints. Canonical codex/integration-6060-floor290 and checkpoint codex/integration-main72-public-validation point here. The broader MSAA array/transfer work remains, with #6445's linear-descriptor correction, #6744's shared pending-image wait and deterministic regression, #6601's readback-fixture capability correction, and selected-device reporting for every shared MSAA test mode. #6359/#6385's other wait corrections were already represented.
Validation pins: at 622a03eda (main332), full Windows app/libraries and patched libs build; 128 successful repetitions across 20 selected groups. Physical GPU selection is verified in 17/17 groups per vendor, with the remaining groups providing CPU/mock coverage. Connected color publication passes 5/5 per GPU, all 48 configurations with formatless storage reads disabled. Linux selected targets pass 55 repetitions across eleven groups. Maximum selected test 16.130s; no whole-test skips/timeouts.
At current ccd37c17f (main72), upstream changes touch draw relocation, descriptor-use metadata and shader-cache serialization. Affected Windows/Linux targets and patched Windows libs build. Targeted validation passes 39 Windows repetitions (six GPU groups on each physical vendor plus the CPU disk-cache group) and 21 Linux repetitions, maximum 5.080s. Groups: draw_cache_relocation, image_sample_vcc_sampler, store_over_untracked, store_pending_identity, multisample_guest_access, msaa_sampled_publication, and shader_disk_cache with their normal CTest prefixes. The earlier footprint/readback/layout fixture blobs are unchanged and their unrelated matrix is not rerun or relabeled. Local conventions and registration pass.
Static analysis remains at its tested main332 pin: footprint/wait/typed-readback translation units have no selected diagnostics; the logging fixture retains three Cppcheck and three clang-tidy findings identical to the matched baseline. Two possible-throw-in-noexcept errors remain unresolved. No new selected diagnostic or suppression.
Unsupported 2x/8x lavapipe readback subcases are excluded from coverage; connected color publication on lavapipe checks 16 configurations and reports 32 unavailable. CPU graphics heap-mirror omissions and fixture write-watch omissions are also excluded. Linux is selected-target/software-Vulkan evidence, not a full Linux application build or physical AMD/NVIDIA run. Windows11 Enterprise26200, WinLibs GCC15.2r7, RTX5080 driver610.62/32.0.16.1062 and Radeon Graphics driver32.0.21036.18; project shader cache disabled, driver caches uncontrolled.
The broader earlier 163-group matrix, original negative controls, staging traces and title frontier retain their original source pins. This slice does not rerun the paused buffer/fault experiments, inspect a private binary or replay a game. No new startup/FPS, native-depth/HTILE-publication, console-layout-equivalence or fresh Addrlib result is claimed. The historical cinematic-to-level-loading frontier remains unchanged. #6588/#6601 remain the focused shared foundations; the #6570 handoff is separate. This broad integration is not an upstream PR. AI-assisted: Codex.
Earlier integration validation at its original pins
Experimental integration: color sampled readback and buffer publication
Current integration eb3c113a6 includes main 5d9dcc42d, preserves the previous validated fe0a86b6c, and connects #6601's shared sampled reader to the retained color-publication path. The raw UINT array view declares sampled/storage usage, while CPU R_X retile, metadata expansion and ownership checks remain. The layout implementation matches #6588 apart from trailing blank lines. The minimal foundation-only branch for #6570 remains separate.
Full Windows app/libs, conventions and registration pass; all 163 selected groups pass per GPU selector, with physical device selection verified in 47/47 groups (NVIDIA/AMD). Maximum matrix test 5.538s. The new connected color-publication fixture passes 5/5 fresh processes per physical GPU: 48 configurations each, with formatless storage-image reads disabled when creating the Vulkan device. Its regression-only control fails 5/5 per GPU on the old consumer's feature requirement. The four Cppcheck and three clang-tidy findings match the baseline in message and source context, with no suppression. These include two existing Cppcheck errors about potentially throwing Vulkan function lookup during fixture teardown; those remain unresolved. No new selected diagnostic. Windows 11 Enterprise 26200, RTX 5080 (10de:2c02, 32.0.16.1062) and Radeon integrated graphics (1002:13c0, 32.0.21036.18), WinLibs GCC 15.2r7 Release; shader cache disabled.
Retained buffer-read validation: staged fresh, deferred, warmed and explicit-recipe V# reads span a 64KiB native image and neighboring output. Source+0/+4/+8 staging ranges were measured at the earlier pins. The refreshed matrix runs those fixtures without a new staging trace. The moved graphics-snapshot case still validates Revalidate, MovedReadOnlyBuffers and PrepareDrawBindings against explicit publication. Original direct reads, native-write refusal, neighboring input/output sentinels and all 1024 native samples remain checked. It does not execute a raster draw consuming the moved snapshot.
Original intentional controls remain at their tested pins, 5/5 fresh processes per GPU:
- Remove compute-reuse synchronization: ordinary/staged warmed and recipe reads become stale; fresh/deferred reads and write refusal stay correct.
- Remove completion-stage synchronization: exactly the two deferred reads become stale.
- Remove moved-snapshot synchronization: eight compute reads pass; only the moved snapshot is stale, with old-range refusal and current-image controls intact.
#6498 remains the separate upstream refusal fix. This is synthetic integration evidence: no new title run, startup/FPS improvement, Linux, PS5 console-equivalence or native depth/HTILE-publication claim. Dynamic BDA heap access is now separately reproduced in #6625 on both GPUs after #6498; it remains outside these publication controls. Vertex/index fetches and end-to-end moved-draw output also remain outside these controls. AI-assisted: Codex.
Activity
Amoenus commentedon Oct 22, 2026
A composition boundary found while refreshing #6973 -> #6987 -> #6989 -> #6995 -> #7002 after merged #8847:
8b3e39832a:agc_driver_multisample_srgb_twinpasses on RTX5080 and Radeon Graphics with required Vulkan, the Khronos layer and application shader caching disabled.a877ad4cf4: the NVIDIA run reaches the fixture's final unrelated-format case, then refuses the initial CPU read because it overlaps a multisampled colour target whose newest samples exist only on the host. No VUID or validation warning. The earlier twin-sharing/fill checks have already run before that point; this is not a failed sRGB twin lookup.The refresh keeps that strict refusal. Its regression now requires the specific native-owner overlap refusal, then checks that the original image/view/owner and its zero samples survive. The sRGB twin-sharing and fill checks are retained. Validation completed: fixture-only main control
3248ef763bfails the missing-refusal assertion 3/3 per physical GPU. Refreshed #69739290696549and all four dependent heads pass their selected checks on both GPUs; the final tip is #700247613d5d23. The shared matrix has 96 passes plus six expected control failures, no VUIDs, validation errors/warnings, whole-test skips or timeouts. All runtime libraries/relinker and selected targets build. The root description names the changed final-case expectation and full qualification scope. The broader live unrelated-format handoff remains unsupported. This is a changed final-case expectation, not a claim that the original fixture passes unchanged with #6973.No new issue is needed: the live unrelated-format alias boundary belongs in this existing alias tracker. No title replay or performance result. AI-assisted investigation and report: yes.