Sonic Unleashed Recompiled (PPSA99902): conditional bring-up and remaining blockers

[Amoenus]

User-data scope correction (2026-10-15 UTC): the earlier manifest restored tracked legacy /app0 files; the game uses /data/UnleashedRecomp, which was outside that audit. Protected original/runtime/guest-libc hashes remain verified, but full active-user-data restoration and identical active-save state were not established. The first “None” configuration trial edited the legacy file and is invalid as a no-AA test. Maintained correction and verified two-root snapshot/restore scope. Numeric fixture evidence remains valid.

Independent buffer-alias guard correction (2026-10-15 UTC)

The confirmed synthetic gap is #6480, with a focused fix in #6498 / fork f9feba0b2. Main's fresh, edge-overlapping, written, warmed, deferred and recipe buffer bindings can bypass the resident native-MSAA refusal. Both NVIDIA RTX 5080 and AMD Radeon Graphics reproduce all six paths 5/5 on main fa9365ca7; the exact fix passes 5/5 each, with adjacent-range and native-image controls. The PR keeps #6452 as a separate Windows build dependency and records full-build, regression, fault-control and static-analysis evidence.

This is an independent submission branch; the retained integration and historical title frontier below keep their existing pins. It does not implement buffer publication or establish Sonic's first failing writer/plane. No new title run or FPS claim.

Current retained integration

Integration eb3c113a6 includes main 5d9dcc42d2ec5401aebd089028b8d4d7410a77b4 and preserves the earlier validated fe0a86b6c, 2579536f5 and bring-up checkpoints. Canonical codex/integration-6060-floor290 and checkpoint codex/integration-main5d-sampled-publication point here. The color-publication consumer now uses #6601's shared sampled reader with a sampled/storage raw UINT array view. The CPU layout implementation matches #6588 apart from trailing blank lines. The small foundation-only handoff is separate.

On this exact integration pin, full Windows app/libs, conventions and registration pass. 163 selected groups pass per GPU selector, with physical device selection verified in 47 groups per vendor; maximum 5.539s, no failures/skips/timeouts. The connected color-publication fixture passes five fresh processes per physical GPU, all 48 configurations each, with formatless storage-image reads disabled when creating the device. The old consumer fails all ten controls. Original staged/moved buffer controls retain their earlier source pins; the current matrix passes their fixtures without a new staging trace. Four Cppcheck and three clang-tidy findings match the baseline; they remain unresolved. Exact evidence, controls and scope.

Windows 11 Enterprise 26200, WinLibs GCC15.2r7 Release; RTX5080/32.0.16.1062 and Radeon integrated graphics/32.0.21036.18; shader cache disabled. #6498 remains the separate upstream shader-buffer refusal. #6588/#6601 provide focused shared foundations; this broad integration is not one submission. Native depth publication/HTILE remains coordinated in #6570.

No new title run or FPS result: the historical cinematic-to-level-loading frontier and resident-MSAA write refusal remain the last title evidence. Its first failing writer/plane is still unidentified. No Linux, PS5 console-equivalence or new Addrlib run is inferred from this refresh. AI-assisted: Codex.

Sonic Unleashed Recompiled — PPSA99902, 01.003.000

Focused contributions (current, 2026-10-15 UTC)

Nine contributions are merged: #3754, #4318, #4451, #4556, #5496, #5524, #5554, #5648 and #5790. The native-depth prerequisite chain is merged. Seven PRs remain open; 13 slots remain under the 20-open quota. Current review, mergeability and checks are maintained in the linked PRs. Exact tested heads, devices, controls and limits are maintained in each PR:

PR	Focus	Next action
#5939	Uniform initial native MSAA DCC fast clears, head3851b712c	Review the tested narrow path; additional CMASK/FMASK states await the requested scalar evidence. Existing guards retained.
#6225	Recorder fixture GPU selection	Review; independent Linux helper results are attributed in the PR.
#6359	Expected pending-image identity in repeated stores	Review; separate neighbour-copy correction is #6385. Windows and attributed Linux results remain mode-specific.
#6422	Native VCC sampler execution regression for merged #4996	Review.
#6445	Fragment-aware sampled MSAA memory ranges	Review; shared helper with #6280 is coordinated. No new title result.
#6498	Refuse native MSAA buffer descriptor aliases, including reuse	Review; #6452 is the separate Windows test-build dependency.
#6452	Portable 64 KiB alignment in the new indirect fixture	Review; fixes the current Windows build regression.

Current applicable checks are maintained in each linked PR; full cross-platform CI requires maintainer dispatch and is not counted as passed when skipped.

Earlier focused validation snapshots (historical heads and superseded status counts)

• feat(agc): import uniform multisampled DCC fast clears #5939: uniform initial native MSAA DCC fast clears, 347d11b93 on main8f51, depends on: none (fix(agc): handle FMASK decompress on MSAA targets #4767/feat(agc): sample a multisampled color target through its 2D MSAA texture #4820 merged). Full Windows app/libs/checkers and16 groups once+5 under NVIDIA/AMD selectors pass, max5.68s; eight runtime groups verify both actual vendors, while two legacy recorder groups remain NVIDIA-only. Five supported Linux groups pass once+5; unsupported2×/8× explicitly skip. Geometry225×5 and unchanged-main state/pixel failure controls are confirmed. No new title replay of this head. #5336 retains detailed scope.
• feat(agc): sample native multisampled depth attachments #5496: native numeric MSAA-depth sampling, 008f91501 on main at3c87cdfda, depends on: none. Full Windows build/checkers and14 groups once+5 per NVIDIA/AMD selector pass, max5.52633s repeated. Actual-draw feedback and four contradictory-clear regressions catch30 expected mutant failures; production restored/rebuilt/retested. The earlier stencil-address classification correction remains.
• fix(agc): retire fully released native depth owners #5524: complete depth/HTILE allocation retirement, focused change, depends on feat(agc): sample native multisampled depth attachments #5496. Full Windows build/checkers and15 groups once+5 per selector pass, max5.7654s.
• fix(agc): apply completion-proven native depth clears #5554: completion-proven native HTILE clears, focused change, depends on fix(agc): retire fully released native depth owners #5524 and transitively feat(agc): sample native multisampled depth attachments #5496. Full Windows build/checkers and17 groups once+5 per selector pass, max5.44355s. Parent regressions and completed-clear/refusal tests are retained.
• feat(agc): emulate default eight-sample dithered alpha coverage #4451 (merged): default eight-sample dithered alpha coverage; tested head246d16a03 merged into main82f112 by adriadam10. Targetless32_AR behavior and canonical24 compatibility are retained; dither uses25. Full build/libs/checkers, three compatibility/state controls ×20 and13 NVIDIA groups ×5 in both GPL modes passed. AMD limits retain their scoped #4768 reports.
• test(tools): split conventions fixtures across CTest groups #5685 (closed duplicate): superseded by earlier test(tools): shard conventions comment checks under CTest timeout #5678. Our Windows validation checks all54 cases over six nine-case shards, five parallel2 rounds, maximum9.295s process/9.129s unittest; direct whole-module main53/candidate54 pass28.729s/28.666s. This is exported Python-runner validation, not full application/CMake/CI. Historical alternative: split the conventions fixture into complete CTest groups; head dbc2ab33c, main107, depends on: none. Original53 cases retained plus three partition checks; five parallel2 repetitions pass, maximum16.4823s, post-rebase maximum19.62s. Intake/conventions and relinker Linux/Windows/macOS CI pass; full application jobs skipped. Originating test-budget issue #5672.
• fix(shader): accept buffer resource tuples ending in VCC #5648: accept MUBUF/MTBUF resource tuples ending in VCC; head943d362d2, main433, depends on: none. Three production lines; all five original patches range-equivalent after rebase, plus fixture whitespace cleanup. Full build/libs/checkers and three controls ×5 under both GPU selections pass without skips (maximum1.40763s NVIDIA/1.21538s AMD); native logs confirm both devices. #5630 retains historical negative/title controls; main4f618 clean merge preview is unbuilt.

Native-depth heads804/9b394/72f are refreshed/tested/pushed on main411 in parent-first order. The14/15/17 matrices pass once+5 under both selectors with runtime physical IDs verified; CPU checks are separate. Child patches remain range-equivalent and include only their own prerequisites. The source-backed stencil alias report is corrected in #5496; combined depth/stencil import and Windows tracking remain unresolved, and DAVE THE DIVER was not replayed here. #3262 carries exact scope. This is separate from the main4808 title replay below.

Buffer head is now tested/pushed943d362d2 on main433, with the fresh checks above. The following original-decoder and matched-title evidence retains its older source/runtime pins. Original decoder plus identical fixtures fails CPU/native controls on each vendor. The original decoder blob is unchanged between tested a5 and a14. CPU160 opcode/base cases; native22 wave32/wave64 configurations ×64 lanes check loads/stores, both VCC words, bounds, formats and guards. Raw gfx10/gfx1013 remains unmeasured; details: #5630.

Own contributions: 5 merged, 5 open, 15 available slots (live paginated inventory; 20-open limit). #4451 is merged; earlier tested heads and scoped AMD controls remain preserved. #5790, merged 2026-10-14T19:06:25Z, is a three-line StoreOverUntracked fixture lifetime correction, independently tested100 times per GPU on main82, with unchanged-main failure/#5789. It is test reliability work, not a new game capability. Generic HTILE store-back #5627 and our guarded native ownership need routing whichever lands second; coordination. Combined heads are not tested.

Progress summary

Historical controlled runs verified PRESS START, the autosave notice, NEW GAME, the opening cinematic and post-movie NOW LOADING. No interactive level/gameplay is verified. The old presentation-counter campaign reported medians6.06 at PRESS START/13.15 in NEW GAME/59.94 in the movie; these are successful VideoOut calls on the older instrumented088 profile, not current-build or gameplay FPS. Pinned historical evidence and limits.

The latest focused title replay uses 8d375dd9d/main324, implementing uniform initial MSAA DCC clears without the retained integration or dependent native-depth patches. Full Windows app/libs/checkers pass; twelve selected groups pass once and ×5 per NVIDIA/AMD selector without skips. Current compiled DCC geometry matches225 pinned PAL source-model cases ×5; the wrong-configuration control fails. Exactly126 modules plus three toolchain DLLs load5/5 and original-input conversion succeeds. RTX5080 execution advances past the DCC decode refusal, then stops29.343s at initial multisampled DCC contents require watched guest memory, exit0xC0000409/no timeout. No input pulse/current FPS/gameplay claim. Originals, baseline runtime/guest libc and user data are restored/hash-verified, cleanup errors empty. #5336 holds this proposed fix and evidence; the first failed coverage range/cause is unestablished. #3036 preserves the different c77/main23be host-import exclusion evidence and qualifies its two stale runtime outputs. These histories are not a performance comparison.

The DCC fix is pushed to our fork. #5939 is now open after the sample split/software checks and current-base validation above. The post-creation audit found5939/5648/5790 mergeable with required checks passing;5496/5524/5554 have since been rebased, tested and pushed. Their new automatic checks are tracked separately. Full CI remains maintainer-dispatched; no skipped full build is called a pass.

Main3e stagingb8206b91b fully builds and passes57 focused groups once, but repeated NVIDIA StoreOverUntracked crashes on execution4; unchanged-main controls establish that fixture failure separately in #5789. This profile was not installed/replayed on the game. The fixed main23be refresh now fully builds and passes60 focused/24 critical groups ×5 per GPU selection; its frozen128-module profile loads5/5 and the fresh title replay repeats the33.387s color refusal. This preserves source progress but does not regain the older menu/movie milestone.

Matched title comparison (main + local patches)

On main + local patches, control86df4c5ad, based on main1978, stops at MTBUF resource descriptor register range overflow after649.304s.

On main + local patches, candidate2bd6889f5, the same source plus the buffer correction, passes that boundary and stops after797.332s at AGC graphics: guest writes to a resident multisample target or its metadata are not implemented.

Installed profiles differ only in libSceAgcDriver. All other modules, converted executable/guest libc and six300ms input pulses match; consumption is unlogged. Seven PE metadata-only rebuild outputs retain control bytes. Both exit0xC0000409 without harness timeout. Original dump/modules/runtime/user data and clean source pins are verified restored, cleanup errors empty. Windows11 Enterprise26200 / RTX5080,610.62. No new menus/gameplay, unchanged-main startup or performance claim.

The next refusal comes from the retained resident color owner's Use/ImportPublishedClear path. The message alone cannot identify the changed data/metadata range or failed precondition. Current finding in #3036; no specific producer/encoding or capability correction is confirmed yet.

Current retained integration and title result (maind92, replay at5aa;2026-10-15 UTC)

Canonical integration10ea356a1 retains the broad DCC/transfer/snapshot/publication, native-depth, MSAA-array and resident-import work on main d92a8dad390af1807cc91622b924252141816a5a, including #6225. Its own delta from669 changes only the public synthetic test. Full Windows app/libs/checkers, five complete submitted-clear fixture runs per physical NVIDIA/AMD GPU, and66 related groups per selector pass. Two internal compared-import paths remain untested. The submitted-clear findings and ruled-out hypothesis are below. The measured original-title replay remains5aa,706.328s at the loading-stage guard; no new title replay, gameplay or FPS claim.

Submitted native DCC clears (2026-10-15 UTC)

The test-only fork delta submits the existing known constant-fill kernel through the real driver instead of calling the native-clear notifier directly. Production is unchanged from669. Each run covers64x64 RGBA8 and RGBA16F at2x/4x/8x samples, with private resident section mappings and expanded CMASK/FMASK. All four numeric DCC clear codes00/40/80/c0 produce the exact expected bytes of every sample/channel. A subsequent same-value CPU metadata store must still refuse. Plain-buffer cases verify16/1024/1040/65536-byte fills, an extra dispatched group, an untouched suffix, and the17-thread dispatch form through the shader path.

Five complete runs per actual RTX5080/610.62 and AMD Radeon Graphics/32.0.21036.18, Windows11 Enterprise26200:240 successful native clear operations and60 same-value-store refusals across the dedicated runs. Slowest full fixture0.65s NVIDIA/0.57s AMD. Full Windows app/libs/conventions pass;66 related groups once per selector pass, maximum2.03s/2.00s, no CTest failures/skips/timeouts/device loss. The unit-shadow and kept-index compared-import paths retain their internal “not tested” messages; selector counts are not blanket physical-device coverage. No Linux/full-suite claim.

Ruled out: a provisional constant-fill shortcut and metadata-fill change were withdrawn. Restoring the original shader path still passes once the driver is initialized before the native target is created. An earlier late-device-initialization arrangement failed at the guard, but the reordered fixture passes on unchanged production. Those excluded arrangements do not establish the game's first writer or a missing clear implementation. Only the final test changes are pushed; experimental production changes are not on the fork's canonical branches.

This verifies the basic submitted-clear path on the retained composition. It does not identify the title's failing kernel, metadata state, plane or first writer, or prove the same operation is its loading-stage cause. The measured5aa title result and default-MSAA failure remain unchanged; the existing resident-import proposal remains separately scoped in the linked #5851 report.

Observed stage remains title/autosave → opening cutscene → level loading. No gameplay or FPS result.

Original-dump profile	Result
Previous9e4925b7	Loading; resident-MSAA revalidation refusal after703.882s
Import correction6bd1fff07	Loading; same refusal text after704.404s
Refreshed5aa3282c3 / maind92	Loading; same refusal text after706.328s

AGC graphics: guest writes to a resident multisample target or its metadata are not implemented

All three exit0xC0000409 without a timeout. The latest5aa run sends all six scheduled inputs and records48 ordinary BMPs; inspected frame047 at601.211s shows loading. Its wrapper ended during restoration. A separate recovery completed and hash-verified the originals, baseline libraries, guest libc and user snapshot, archived all48 frames and removed the owned temporary executable; no restoration work remains.

For9e42→6bd, only libSceAgcDriver.prx differs between129 runtime files. For6bd→5aa, all126 PRXs were rebuilt and differ; the three toolchain DLLs, freshly converted executable and guest libc are byte-identical. The new replay is a joint integration/main refresh, not a module-isolated causal test. These timings identify the runs; they are not a speed comparison.

At6bd: full Windows app/libs/conventions/registration pass;75 focused groups once (max24.2387s),40 critical groups five times under each NVIDIA/AMD selector (max5.35228/5.72351s),0 CTest failures/skips.22 critical fixtures record both physical GPUs;18 without device evidence are not counted as GPU coverage. Three internal StoreOverUntracked “not run” cases remain. Windows11 Enterprise26200, RTX5080/610.62 and Radeon Graphics/32.0.21036.18. Frozen126PRX/threeDLL profile loads5/5. No Linux/full-suite or console-hardware claim for this head.

The neighboring-import correction is retained because its public synthetic regression and real copied-buffer GPU fill are verified, with ten intended fault-control assertions and successful restoration. It does not unblock this title's first refusal. The game’s changed plane, first writer and failed clear-import predicate are still unidentified; the shared message cannot establish them. Next: continue bring-up from the measured loading stop, using public source and synthetic regressions to narrow a general fix while preserving the existing checks. Maintained reproduction/proposal and comparison.

Earlier retained integration and title result (main23be)

Integrationbd1e4bb2a, main23be plus retained fixes, fully builds and passes60 focused groups and24 critical groups ×5 under each GPU selection without skips (maximum25.8692s focused/5.44708s NVIDIA/5.3604s AMD). Merged #4451, upstream depth-plane sharing and the #5790 fixture cleanup are retained. These are composed-stack checks.

The frozen128-module runtime loads5/5 and original-input conversion succeeds. Normal RTX5080 replay stops33.387s/0xC0000409, no timeout, at multisample target reuse requires watched guest memory; untracked writes are not implemented (color: range outside write-watch coverage). The coverage exclusion's producer remains unknown. No input pulse/private observer/guard bypass/current FPS or gameplay claim; originals/runtime/guest/user data restore and hashes pass, cleanup errors empty. Exact color scope/#3036. This retained composition is not unchanged individual PR-head validation; see the current native-depth maintenance above.

Earlier integration refresh (main0b6)

All retained fixes, including #5648's correction, are preserved in source feb6c224f, now merged with main0b6. The canonical fork branch codex/experimental-ppsa99902-msaa-restored has advanced from86df to this head; earlier source/runtime checkpoints remain preserved.

Three conflicts were resolved: Windows fixture library paths retain local initial-content refusals and main's resident-present/slice-dirty cases; draw completion retains upstream targetSlices and our native-MSAA dirty marking; serializer tests use the reviewed #4451 fixtures regenerated from canonical main. The production serializer matches that reviewed PR. A duplicate covered-centroid guard already covered by merged #4556 was removed; the main guard remains.

Full build/libs/conventions/registration pass. 35 focused groups pass without skips (maximum25.0001s), including serializer, new memory/presentation groups and existing MSAA/alpha/descriptor tests. Nine critical groups ×5 pass on NVIDIA and AMD without skips (maximum1.28835s/1.18637s). This is composed-stack validation, not unchanged individual PR heads. Win11 Enterprise26200, WinLibs GCC15.2.0r7 Release, SPIRV-Tools ON, bundled dependencies; RTX5080/610.62 and AMD1002:13c0/32.0.21036.18.

Fresh128-module loader checks pass5/5 and original-input conversion succeeds. The normal RTX5080 replay stops after28.675s/0xC0000409, no timeout, at native multisample depth import requires watched guest memory, before the first scheduled input pulse. All originals/runtime/guest/user data restore/hash verification passes with no cleanup errors. No new menus/gameplay/FPS or unchanged-main result.

The failing legacy depth path throws during attachment creation. The focused #5496 implementation records unsupported sampling state while preserving existing drawing, then refuses sampling; this integration is not byte-identical to that PR. #3262 and #3036 are updated in place. No specific changed range or upstream cause is established. The focused sampling/retirement/clear alignment is now c7e3a456d, pushed to the canonical integration branch. Full build/libs/checkers,40 focused groups and14 critical groups ×5 on each GPU pass without skips (maximum24.8258s focused,1.27366s NVIDIA/1.22426s AMD critical). Legacy tests assert sampling refusals after permitted drawing and retain complete-HTILE-store refusals. The native retirement fixture enables and initializes the actual integration color-transfer path; its initial setup failure is corrected. Core depth code retains the focused implementation plus legacy layout/overlap helpers. Fresh full-runtime loader checks pass5/5 and original-input conversion succeeds. The aligned RTX5080 replay stops after34.566s/0xC0000409, no timeout, at multisample target reuse requires watched guest memory; untracked writes are not implemented, in the retained color owner's reuse check. No guard was bypassed; no input pulse was sent. The zero-generation-versus-unwatched range and its cause are not identified. Originals, baseline runtime/guest libc and user data restore/hash checks pass with no cleanup errors. No new gameplay/FPS or unchanged-main startup claim. #3262/#3036 retain this result in place; latest main is staged separately.

The earlier a14 full1163 suite remains1154pass/8skip/1timeout: conventions_function_bodies at30.0414s. It passes alone27.6355s/53 cases; the test/checker/registry blobs are unchanged on current main. Dedicated tooling report #5672 proposes splitting complete case groups without raising the30s budget. The later #5685 is closed as a duplicate in favor of earlier #5678; #5672 and our validation comment on #5678 record the completed Windows shard campaign.
AI-assisted: Codex; author responsible.

Historical evidence (pinned reports retained below)

Windows 11 Enterprise 26200, RTX 5080 (driver 32.0.16.1062), WinLibs GCC 15.2.0 r7, Release. This report tracks a conditional fork integration. Interactive gameplay, audio and rendering accuracy remain unverified; there is no new clean-main game launch in this update. Private binaries, shaders and captures remain local.

Earlier captured graphics stop (preserved integration)

A bounded replay of integration 0f6edc64d, with scalar descriptor logging, stops after 771.826 seconds, exit 0xC0000409, without timeout:

rendering into a packed mip tail needs the resident image of its surface

The captured target is D64_X (tile mode 26), RGBA32_UINT, 16-byte texels, surface 128×64 with 10 mips, selected mip 2 at 32×16, pipeBankXor 0x4000, one sample, no DCC/CMASK. gpuTiling=0 despite an available detiler. Current draw routing enables GPU tiling for R_X only; the packed-tail fallback requires that route. A native captured-state replay on unchanged later main 5f01c8082 now passes the normal resident path 5/5 on each GPU, including draws into mips 2, 3 and 1 and full-surface canonical readback checks. Only the forced no-resident diagnostic stops at the R_X-only fallback guard. The old integration stop therefore does not by itself establish a new current-main title blocker; the routing candidate remains unsubmitted pending a refreshed title run. Merged #4328's R_X tests remain valid within their scope. The descriptor and proposed follow-up are maintained in #4328.

This run stopped before the MSAA clear/stamp comparison and supplies no new write-origin or coherence verdict. Original dump/modules, baseline runtime and user data were restored. Logging/timing can change which stop is reached on the older conditional integration.

Own contributions and current validation

Our two merged contributions are #3754 and #4318 D_X layout, merged at 6f438a1c2 on 2026-10-12. Two upstream PRs remain open. #4451 is rebased and fully rebuilt/tested on pinned main 075169f7b; #4556 retains its validated base 43e3a70f8 and remains cleanly mergeable. Later-main merge previews are distinct from those pinned tests. The merged #4318 row retains its older test base 2f4cc8983.

Contribution	Validated fork head	Latest native results
#4318, now merged	b6f1a4a69	Full NVIDIA 989 passed / 0 failed / 7 skipped of 996; 13 selected NVIDIA groups ×5 pass. Corrected AMD GPL-on: 11 groups ×5 pass; host-exclusion fixture and extended array draw fail. Residuals remain reported.
#4451 alpha8	0bcf45842	Main 075: NVIDIA 1054 passed / 0 failed / 8 skipped of 1062, maximum 24.0818 s; 15 groups x5 both modes. Matched-main AMD undithered-on/wave32-off mismatches remain (#4768). Format 22 retains canonical 21/20; twelve original goldens per version, three gate fault controls. Title prerequisites: #4556/#4443.
#4556 inactive covered centroid	5d0f57d7d	Full NVIDIA 1012/0/8 of 1020; 12 relevant groups ×5 pass on each GPU; recorder-helper device caveat retained. Depends on: none.

Harness correction: APS5_NO_GPL checks presence; 0 also disables GPL. The corrected on path leaves it unset; off sets 1. Earlier 0/1 matrices exercised off twice. Recorder-based helper controls independently select NVIDIA, so they are not AMD coverage. Driver-backed GPU selection was checked in logs. The newly confirmed executable-placement/GPL sensitivity, reproduced by matched main controls, is recorded in #4768; it is not established as an alpha regression. Earlier AMD artifact/control findings are consolidated in #4435, without claiming recurrence of its resolved lifetime bug. Full Linux/Windows CI jobs were skipped by automatic eligibility; local validation is not a full upstream CI pass. These native runs are not title startup validation on current main.

Retained MSAA findings and limits

The integration reaches PRESS START, the menu, opening movie and post-movie NOW LOADING/tutorial screen. Earlier bounded runs (785–794 seconds, restored assets/runtime/user data) reach a local MSAA guest-write/metadata reset guard, before XOR retile/reinterpretation. The decoded draw has eight samples, DB_ALPHA_TO_MASK=0x18701, PA_SC_AA_CONFIG=0x2030e003, export format 4. At the same color allocation the retained/new uses differ in XOR and FMASK pointer. Synthetic XOR controls prove that ignoring the view difference changes layout in that model; they do not establish the console transition.

Tracked-generation changes cover the 510 MiB color surface and FMASK/CMASK/DCC regions. Collected CPU flags do not prove actual guest CPU stores: repeated reports without stores occur in pinned shared-memory controls, and ImportWindow stamps also set the flag. Live snapshots describe shared, clean, unpinned views, without overlapping host-write windows, pin requests or captured handled write faults. There is no live evidence that pinning explains this title stop. Exact live stamp creation/age and completed writes remain unresolved; no guard was relaxed.

Recognized fill requests include expanded FMASK 0x76543210, CMASK 0xffffffff/0xcccccccc and DCC zero. None of the 3,074 completely parsed records targets the color allocation; one split record's continuation is retained. Requests are not completion or uniform overwrite proof, so #2132 is not validated as the solution to this reset. The ELF revision matches the pinned Mesa FMASK initialization source; exact executed build flags/call paths remain unverified. Detailed controls remain in #2132.

The conditional MSAA/native-readback foundations and older evidence remain in the fork integration and #419's validation report. #419 is closed; this combined stack is not a standalone current-main submission. Related reports: #3036, #3262, #3219. The guards above belong to this local integration and are not claimed as unchanged-main regressions. Movie presentation near 59.94 successful display submissions/s is not unique rendered-frame or gameplay FPS.

Refreshed integration and latest title stop (2026-10-13)

The refreshed fork head 6cc5beb7f, branch codex/experimental-ppsa99902-main5f01, merges main 5f01c8082 with the preserved older integration. All 29 textual conflicts are reconciled. Merged entropy/TLS and pthread scope implementations/tests are retained; request version 19 retains original version 18 compatibility, and debt is migrated to per-library files. Full build/all libraries, conventions and registration pass. Full NVIDIA: 1027 passed / 0 failed / 8 skipped of 1035; all 15 relevant groups pass five times. The original five-failure campaign and cancelled invalid retest remain separate artifacts. Repairs corrected merge discrepancies and stale fixture assumptions; they are not reported as new main defects.

The guarded replay of its frozen 124-library patched runtime, keeping the fixed f609 conversion, stops before rendering after 9.998s, exit 0xC0000409, no timeout/frames:

fopen_nid_postfix: open failed: ".../ressources/game/movie/Inspire": Permission denied

A verified libc-only logging replay reproduces the same stop after 9.626s and confirms /app0/ressources/game/movie/Inspire, mode rb. The path is an existing directory. FileAndHeap.cpp is unmodified by this integration and unchanged through later main 43e3a70f8. The observed abort, primary contracts, duplicate audit and proposed focused investigation are now in #4814. The independent temporary-directory regression now aborts on unchanged main 43e3a70f8; a focused failed-open NULL/errno candidate passes the initial native fixture 20 consecutive times. An upstream solution is now open as #4822. Independent Windows validation of its exact 77ede919c head passes five filesystem/stream groups ×20, with a failing-main production control; the PR comment records commands and scope. Our stronger independent fixture also passes five groups ×20 and is preserved in the fork, alongside the experimental adaptation. Full Windows build/libraries of #4822 head77ede919c and NVIDIA suite now pass (1011/0/8 of1019, maximum22.809s). A fresh frozen-6cc control reproduces the fopen abort after10.126s; replacing only libc changes the observed abort to non-socket poll after9.981s, without frames. Originals/runtime/guest/user data are restored/hash-verified; this is conditional validation and does not establish console directory-stream equivalence or gameplay. #4814 is updated in place with the native evidence and limits. This is a conditional integration run, not unchanged whole-main startup.

Original executable/bundled libc, baseline runtime, converted guest module and saved user data are restored/hash-verified after both replays. The old 0f6edc64d branch/runtime remains preserved. This combined experimental branch is not an upstream PR candidate. Private captures/assets remain local.

Current poll stop and existing fixes

The libc-only #4822 profile now stops at poll: descriptor 3 is not a socket / descriptor5 on another thread. A temporary scalar observer independently identifies read pipes requesting POLLIN with an infinite wait, count1; its descriptor numbers are4/5. The pre-fix integration and pinned main share Udp.cpp. A clean current-main carry of #3644 passes six related groups ×20 and full NVIDIA1011/0/8 of1019, maximum22.172s; its kernel-only conditional replay renders the SEGA startup logo and reaches the existing private MSAA residency guard after789.240s without timeout. All baseline data is restored/hash-verified. The detailed validation distinguishes input delivery, rendering and remaining unknowns. The experimental headfac106b2b is preserved in the fork. Existing #3521 and overlapping #3644/#3649 cover this guard. Our existing issue comment holds the scalar evidence and scope. No duplicate issue or own poll PR is being created. The exact-#4822 experimental head088e3512b is preserved in the fork; the diagnostic observer remains local and its source is restored.

Input/capture and MSAA certificate follow-up

The fresh diagnostic replay on the validated two-API-fix stack ends at the same private MSAA guard after 801.118s, no timeout, with all baseline data restored/hash-verified. All six pulses are confirmed consumed/released; inspected sparse captures show the title prompts and opening movie. Gameplay/FPS remain unverified. The refusal has clearGeneration1951852 and matching current/certificate write serial44350, alongside newer DCC/CMASK stamps recorded after the existing clear-import attempt. The existing #2132 observation contains exact values and the remaining provenance/first-predicate uncertainty. This is an unchanged private refusal; no new main-defect issue or relaxed guard.

Our #4451 uses format 22, preserving canonical 21/20 and versions 16–19; twelve goldens per version came from exact main-4bb/main-075 original serializers. The existing #3763 coordination comment is updated; merged #3763 establishes canonical21, now preserved by our tested format22. Compatibility with other unmerged proposals is untested. #3644 has now merged, including a closed-pipe follow-up; our native/title results above remain scoped to the original725 feature and its45 current-main carry, not that later follow-up.

First-predicate replay (795.601s, restored baseline): the first refusal is color UnchangedSince against initialGeneration878765, with newestStamp1945436. Clear serials44298 match; DCC and CMASK certificate checks passed before the color rejection. Uniform-metadata validation/clear import were not reached. The newer metadata stamps from the prior801.118s trace do not identify its first rejection. The existing #2132 report now records the exact short-circuit result, scope and next color-stamp provenance check; six inputs were confirmed consumed/released. No actual CPU-store/full-color-fill or gameplay claim, and no unchanged-main defect established.

Color-stamp provenance replay (801.006s, baseline restored): all8160 color blocks have newer latest Cpu stamps created by Windows dirty-page collection; no latest Driver/ImportWindow/Fresh stamp. First/newest block0x405fc00000 has generation1943351, prior37566, against initial869030. The color check rejects after preliminary clear-certificate checks pass. This identifies the tracker path, not the backend dirty event or an actual guest CPU store. Rebuilt instrumented native controls pass three groups×20; the earlier libraries-only control scope is corrected. The existing #2132 comment consolidates current/older evidence without repeating every run. Next trace shared-page invalidation/collection events and aliases; #3874 is not confirmed as the solution. No guard relaxation or new-main-defect claim.

Shared-backend lifetime follow-up (787.785s, restored): sampled first color views are fresh shared-map records at the same virtual addresses (seen0/generation2, map-shared, one alias, later event serials), before all8160 color blocks receive new Cpu stamps. This establishes new mapping records, not completed stores, physical-backing identity or successful registry-unmap notification coverage. The consolidated #2132 report holds exact values and limits. Existing #4474 is the next retirement candidate to validate; no duplicate PR/new main-defect claim. The refreshed fork integration0c74657e7, branch codex/experimental-ppsa99902-main495, is now fully built/tested on main 495bf91a04b2564772c4f0fdc19012bed9b8af01. Sixteen textual conflicts are resolved; canonical request19 is preserved with alpha in20, including eight original19 fixtures, and the merged pipe-poll follow-up is retained. Main's sampled depth-plane/array changes and read-guard changes are reconciled with the private native-MSAA paths. Full Windows Release build/all libraries, conventions, registration and diff checks pass. Full NVIDIA1034 passed /0 failed /8 skipped of1042, maximum22.738s, parallel2/30s bound. All14 correctly registered native groups pass×5 without skips: shader-memory, pixel-ancillary, write-tracking, guest-memory/poll, MSAA initial-content refusal, graphics/alpha state, dithered coverage, MSAA color/depth execution, depth-plane sampling/guest layers, and pixel ancillary execution. A124-library patched profile is frozen and hash-verified. No game launch on this new profile yet, no AMD/Linux/full-upstream-CI claim. All title observations above remain on frozen fac106b. This combined branch preserves development progress; it is not an upstream PR candidate.

Existing #4474 Windows validation: unchanged026c198e on parent main1acfa7f78 passes full build/libraries/conventions/registration, eight groups×20, and full NVIDIA988/0/7 of995, maximum23.883s. Reverting all three changed production files to parent main makes the unmap/removal regression fail; exact restoration/full rebuild passes. Our existing comment records commands, original-head scope, skips and the current merge conflict. The conditional adaptation b553a8f76 is now built/tested and pushed to the fork: full NVIDIA1035/0/8 of1043 (max23.514s),16 relevant groups×5, and an expected failing pre-adaptation control followed by exact restoration/full rebuild. Its grouped color-plus-metadata retirement retains partial/failed/protection refusals. The existing #4474 comment records precise scope and the no-op OS-unmap callback used by the native fixture. The controlled title comparison uses122 byte-identical baseline modules and only two changed libraries: before0c stops at resident-color writes after784.791s; adaptedb553 passes that guard and stops at guest writes over native multisample depth are not implemented after791.335s. Both show the opening movie, exit0xC0000409 without timeout, send the same six pulses and restore/hash-check protected originals/runtime/guest/user data. Existing #3262 records the newly reached private depth frontier and its two possible predicates. Notification counts/ranges and completed CPU stores are unverified; no gameplay/FPS/unchanged-main claim, duplicate PR or new issue from this private guard. The scalar depth observer then ends after802.420s at the indirect-context-register CPU-read guard, with no coherence-predicate label emitted. It records the sole prepared native depth/HTILE pair fully covered by a successful0x4081c00000+0xffa4000 unmap with the observer present; the stale private depth guards have no retirement. Existing #3262 records exact ranges, duplicate piece/whole-line interpretation, instrumentation/smaller-release limits and restoration. The tested private proposal42aebfe20 is pushed to the fork: shared color/depth release registry, complete original-byte interval accounting across split releases, deferred old Vulkan objects and inactive depth exclusion. Full NVIDIA1036/0/8 of1044 (max23.940s),17 groups×5 and24 new depth-retirement configurations pass; b553 production with the new fixture fails, exact restoration/full rebuild passes. #3262 now links proposed code and precise tests/limits. The guarded title comparison is complete: production b553 stops at native-depth writes791.335s/61 captures; tested42a stops at MIMG descriptor register range overflow807.203s/59 captures, after GPU submission37032/37033 stalled at37029 and fence error0xFFFFFFFC. Only AGC changes;123 other runtime modules are byte-identical. Restoring the exact124-module b553 profile repeats the original native-depth write refusal after786.544s/60 captures. All three abort0xC0000409 without timeout, six pulses sent/consumption unlogged, originals/executable/runtime/guest/user data and archived capture hashes independently restored/rechecked. This strengthens the profile comparison but does not diagnose the later GPU/MIMG failure. Native tests prove retirement behavior; the changed frontier plus restored old refusal does not establish the first depth predicate, gameplay or a missing MIMG capability. The refusal-only observer now captures image_sample at program0x20ba53400/PC0x260/index152 with S#base104 (four words), T#base72 (eight);787.253s/59 captures,0xC0000409/no timeout, restoration independently verified. No submission/fence error is logged in this observed run. New #4991 records exact-current-main c2565f0 GPU-free decoder reproduction and LLVM's s104/s105/VCC tuple source; raw gfx10 tuple execution and the earlier GPU failure remain unverified; existing VCC component reads are now checked by the supplemental tests below. #3262 is edited with the exact comparison/limits; no new issue/duplicate PR or FPS claim. The proposal depends on private native-MSAA/#4474; our #4451 rebase is now pushed and clean. Current sampler follow-up: @MarcZX100's #4996 covers this boundary; our unchanged-head and supplemental native validation is now posted, with test code preserved on the fork. No duplicate fix PR. The conditional integration now passes full build/libraries/checkers,26 relevant groups x5 on both GPUs, and full NVIDIA1037/0/8 of1045 (max23.804s). Its exact124-module frozen profile loads x5; only AGC differs from42a,123 other modules are byte-identical, and AGC named imports match. The guarded comparison is complete: prior42a stopped at MIMG descriptor register range overflow after807.203s/59readbacks; the credited-fix profile stops at MTBUF resource descriptor register range overflow after792.920s/59readbacks, with no MIMG range exception logged. Both exit0xC0000409 without the1080s timeout and send the same six300ms pulses (consumption unlogged). The new run also logs GPU submissions36322/36323 not reaching36321 in10s and fence error0xFFFFFFFC before the MTBUF exception; their cause remains untraced, as in the earlier uninstrumented42a run. Protected originals/executable, exact baseline runtime/guest, saved user data, all59 archived readback hashes and source cleanliness are independently rechecked after restoration. This is a changed decoder frontier on a conditional integration, not unchanged-main startup, raw hardware equivalence, gameplay, rendering accuracy, FPS or a speed improvement. MTBUF instruction fields are not captured yet, so no new missing-capability claim or issue follows solely from this refusal. The private scalar observer preserves the original errors/reads and passes full build/libs,26 groups x5 before and after exact source restoration, full NVIDIA1036/0/8 of1044 (max22.770s), and accepted/rejected scalar controls. Its frozen profile changes only AGC from42a with123 other modules identical. Observer source/binaries/captures remain private; the production worktree and protected title data are restored. No raw-hardware equivalence, post-fix title progression, FPS or gameplay claim.

Caller/source backlog refresh (main377): Twelve recorded APIs are rechecked against pinned main377959777;21/21 original reference bytes/GOT operands (17 calls/four tail jumps) still match the original ELF, and two corrupt-reference controls are rejected. Main now logs the heap diagnostics, implements the moved module-segment query and relative symlinks, and no longer has the old combined pthread attribute refusal. LoadExec exit and observed UMTX2/3/11 remain supported; process replacement/fork/CPU temperature still refuse. Symlink target strings are runtime-selected, so absolute-target reachability is unknown. The old58c implementation baseline was our experimental fork; its SoC stub is not evidence of a regression on main. Existing #4588 now has the rechecked sensor0 and1–15 caller evidence, with an output-sentinel test suggestion and explicit static-only scope. No new independent title blocker, game execution of these APIs, new Ghidra analysis or new issue is claimed. Historical inventory/project snapshots remain preserved.

#4451 current rebase: 0bcf45842 on pinned main 075 preserves reviewed undithered/debt corrections, canonical 21/20, and alpha in format 22. Twelve original main-4bb (20) and independently emitted main-075 (21) goldens establish the compact compatibility loop. Build/libraries/checkers, all three incorrect-gate controls and restored compatibility/undithered x20 pass. NVIDIA 15 groups x5 in both modes; full 1054 passed / 0 failed / 8 skipped of 1062, maximum 24.0818 s. Matched main/restored original fixtures reproduce existing AMD alpha-on/wave32-off mismatches (#4768). BDA device/span/interlock each pass five serial repetitions with matched main and restored candidate (#3291). Sparse CPU/native controls pass x5. Existing reports are edited in place; #4556 and remaining MSAA/startup prerequisites remain required. Older main-77/4bb results retain their own head/schema scopes. No new title or full upstream-CI claim; later merge preview is clean and unbuilt.

#4474 rebase aid: Upstream-only fork828e6e461 rebases the original retirement commit onto maine514, preserving tiled guard sizing/tests and current fill handling. Full build/libs/conventions/registration, failing-main retirement control and eight groups×20 pass. Full NVIDIA1017/3/8 of1028 (max22.678s); isolated BDA device passes×5, while span/interlock fail on matched main and restored rebase. The existing PR comment now addresses JJenkx with the branch link/scope; our #3291 comment records the residual comparison. No private integration/new duplicate PR/full-green claim.

Next steps

Nine contributions are merged; seven remain open (#5939, #6225, #6359, #6422, #6445, #6452, #6498), leaving 13 slots. The retained integration tracks tested main03c4cb5e7 with previous progress preserved; the buffer-publication dispatch gap is maintained in #6480. Continue extracting small, nonduplicate general fixes from the retained work; maintain peer validation when another PR covers a finding. Investigate #5939's extra CMASK/FMASK states only from the requested source/scalar evidence. Keep source pins fixed during tests and update main between completed slices. No new title-replay or FPS claim.

AI-assisted investigation and code: Codex; author responsible. This discussion is edited in place as evidence changes.