AGC: normal rendering to an 8x MSAA DCC color target is refused (PPSA99902) #5336

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 manual launch, 2026-10-17 17:08:52–17:09:18 UTC

The user double-clicked the prepared launcher for the same c784 integration, with the standard retained conversion. The saved log independently confirms the selected NVIDIA GeForce RTX 5080 (10de:2c02) and the identical final exception: writes to a resident multisampled DCC target or its metadata are not implemented (DCC range 0x40819f4000+0x87000, 8 samples). The console reported exit -1073740791 (0xC0000409). This confirms the current automatic-run stop during an ordinary user launch; it is not an unchanged-main run, menu/gameplay result, or new capability fix. No screenshot or FPS measurement was taken for this manual run.

Preserved local log: sonic-nvidia-31791-5827.log, SHA-256 7126c36cfac8d60ee656991d07cad678cac3a5afb6e13015a9bc924a0b573b4c. AI-assisted log analysis; user-operated launch.

Latest current-integration title replay, 2026-10-17 UTC

Integrationc784faf893, main667 plus retained fixes, #7187cf84 and #7196 head28ea, reaches this resident8x DCC refusal on verified RTX5080 after28.656s from a long physical root (metadata filename274 characters). Range remains 0x40819f4000+0x87000;7 captures, last initial NOW LOADING. AMD Radeon Graphics stops after70.750s at vkQueueSubmit recorder: Vulkan result -4,3 captures, last NOW LOADING. Both exit0xC0000409, no timeout/input, protected inputs/configs unchanged; Khronos1.4.363 loaded with zero observed validation errors/warnings.

All127 built libraries match each respective runtime and between GPUs; a515 executable/guest conversion retained. Full Windows/Linux relinker/library builds and51/33 focused filesystem/stream/archive passes. Current pins, hardware and complete scope are maintained there. This is a combined-integration observation, not isolated proof for either path correction or a common AMD/NVIDIA cause. Source audit correction: e040cb7 is already an ancestor of integrationc784 and earlier observers567/3c4. The active checkDccCurrent expression includes UnchangedSince(...) || !StoredOver(...). Our earlier statement that this candidate was outside the current integration was wrong. The current Sonic refusal therefore occurs with that proof enabled; it does not show that enabling it would unblock this stop. Next observe the existing rejection reason without changing its decision. The #7195 metadata capability is now independently validated through #7196.

Earlier c97 resident DCC refusal, 2026-10-17 UTC

#5939 merged initial uniform-DCC import. Fresh Sonic PPSA99902 validation at integration c97d127064, containing main e09dbc2b42 and the retained #7085/#6280/#6934 composition, reproduces the resident refusal on NVIDIA after27.608s:

AGC graphics: writes to a resident multisampled DCC target or its metadata are not implemented (DCC range 0x40819f4000+0x87000, 8 samples)

Verified RTX5080 (10de:2c02), driver32.0.16.1062, Windows11 Enterprise26200;7 captures, last initial NOW LOADING. This matches the earlier confirmed8398 range/sample count. The DCC plane is known; its first writer, value, ordering and required transition are not. No guard was relaxed.

The matching-runtime AMD run (1002:13c0, driver32.0.21036.18) reaches initial NOW LOADING (3 captures), then stops at71.108s with vkQueueSubmit recorder: Vulkan result -4; a common cause is unestablished. Both exit0xC0000409 without timeout/input, have zero observed Khronos1.4.363 errors/warnings and preserve protected original executable/module/configuration hashes. Fresh ordinary conversion, all127 matching PRXs, isolated fresh assets/user data and unchanged configuration. Exact executable, guest-module, library and dependency hashes match across GPUs. These are integration results, not plain-main startup or gameplay. #3036/#4443 maintain the full current table; #7066/#7085 retain focused extra-mip evidence. AI-assisted: yes (Codex); author responsible.

Earlier bounded replay variation, 2026-10-17 UTC

Earlier5045 with existing tracing was deliberately stopped at300.640s, byte-identical5045 without tracing at316.723s, and exact archived8398 reached its300s harness limit (300.782s total). All last captures show initial NOW LOADING with zero observed core validation errors/warnings and unchanged protected original/configuration hashes. Those non-reproducing results remain valid alongside the new c97 reproduction above; no cause for the variation is established.

The trace/plain pair shares executable/module/all127 library/dependency hashes; archived8398 matches the earlier failed runtime hashes. Existing capture buffer notes did not record overlap with the earlier reported DCC range, but its live allocation was not established, raw CPU stores are not covered and forced termination may lose buffered tail records. This does not identify or exclude a writer. Tracing and #6934 are not established causes. The first writer and required transition remain unknown.

Rejection reason confirmed, 2026-10-17 UTC

Public-source observer7b5b036a9f preserves the existing decision and reports which StoredOver predicate rejects. NVIDIA RTX5080 stops after28.644s at the same DCC range, with this intact record:

plane=DCC address=0x40819f4000 bytes=0x87000 generation=18598 samples=8
unchanged=0 stored=1 reason=driver-overlap

This identifies a retained driver-store piece intersecting the range after its saved generation. It rules out a coarse-only rejection, CPU-page branch, unknown/evicted-history branch and Windows host-region uncertainty as the branch taken in this query. It does not yet identify the producing AnyPS5 call or establish the exact underlying GPU store. The e040 adjacent-store proof is already active in this integration.

AMD Radeon Graphics still stops at Vulkan device loss-4 after73.150s, without a stored=1 rejection record. Both runs exit0xC0000409 without timeout/input;7/3 captures, last initial NOW LOADING, original inputs/configs unchanged. Windows11 Enterprise26200; drivers32.0.16.1062 /32.0.21036.18; Khronos1.4.363 loaded, zero observed validation errors/warnings. Compared with integrationc784, runtime roots have equal lengths, executable/guest module/dependencies and123 of127 PRXs match;4 changed modules are Agc, AgcDriver, SystemService and VideoOut. These are diagnostic integration observations with the a515 conversion retained, not unchanged-main or isolated AGC-DLL results.

The observer's full Windows/Linux relinker/library builds pass. Existing2x/4x/8x DCC controls pass3 times per physical GPU (18 passes); Linux4x plus write-watch controls pass3 each (6). Adjacent writes still pass and same-value overlapping stores still refuse. Clang-tidy completes without selected diagnostics on both changed TUs; Cppcheck retains the pre-existing conditional-unique_ptr advisory, without parser failures. Next trace the source call that records the overlapping driver store. No new capability fix or common AMD/NVIDIA cause is claimed.

The recording call is identified

Observer78db9b00c9 adds source-location records for writes overlapping registered resident metadata ranges. NVIDIA reproduces the same refusal after28.624s. The rejected piece is 0x40819f4000+0xc000, generation19101, newer than saved18564. It exactly matches a record from BdaResources::markWrittenPages, line322.

That call records five ranges whose union is the DCC plane0x40819f4000+0x87000, plus CMASK0x40819e0000+0x14000. This identifies the AnyPS5 recording path, not the exact GPU instruction or written value. BdaAbi::ForEachWrittenRange emits4KiB page ranges, so full page-range coverage does not prove every byte was overwritten or establish a uniform clear. The overlap refusal remains unchanged.

Full Windows/Linux relinker/library builds pass at78db; the existing positive/refusal fixture gives18 physical-GPU passes and6 Linux controls again. Static analysis retains the same unchanged-code Cppcheck advisory and no selected clang-tidy diagnostics; no parser failures. The NVIDIA run retains the matched root length, executable/guest module/dependencies and123 libraries; the same4 modules change as in7b. No timeout/input or original-file changes,7 captures with last NOW LOADING, zero observed Vulkan validation errors/warnings. No AMD title rerun at78db is claimed; its7b result remains separate.

The e040 candidate is already carried upstream by MarcZX100 as #7132, preserving our three original commits unchanged. It is active in these integrations; no duplicate PR is needed. Next establish the metadata-write semantics with public-source/generated fixtures before proposing a capability change. In particular, page coverage alone is insufficient justification for accepting a full metadata reset.

Confirmed neighboring-store gap; title causality under validation

Main c9e6c45de9 has a separately reproduced false refusal: after importing/drawing a DCC target, a driver store outside its4KiB metadata range but inside the same64KiB tracker block makes the next resolve throw the resident-DCC exception. The supplemental regression fails3/3 on verified RTX5080 and AMD Radeon Graphics with unchanged production code (test-only975c3ee6b3). This does not identify Sonic's writer.

Retained focused candidatee040cb7f75 preserves the existing unchanged-generation proof and additionally accepts the existing finer StoredOver proof after the same uncached collect. Actual overlapping writes still refuse; unknown/evicted fine history stays conservative. Keeping the original proof matters because StoredOver deliberately treats Windows host regions as uncertain. This supersedes the query-only6eeb draft; the draft's title observation was deliberately stopped at63.195s after source review, not an application failure or completed test.

At exact e040, the full fixture passes2x/4x/8x three times each on both physical GPUs (18 passes) and4x three times on llvmpipe (2x/8x unsupported/skipped). Adjacent driver/CPU stores preserve rendered samples; same-value writes inside DCC and the existing CPU color/DCC/CMASK, unwatch and unmap controls still refuse. Windows Khronos1.4.363 runs have zero observed errors/warnings. Both OS relinker/library builds and CPU graphics groups pass3/3. Clang-tidy is clean on both changed TUs; Cppcheck has only the unchanged conditional-unique_ptr warning at line48, manually triaged. No globally-clean-analysis claim.

Latest title check: observer3c4 on refreshed integration15f/main21 completes300s on each physical GPU at initial NOW LOADING. The Pad layer consumes all six scheduled command records in each process, while neither log emits a DCC proof-entry or coarse-change decision. This does not exercise or validate the finer proof for Sonic. The no-candidate a515 control and earlier5b5 observations also lacked the refusal; c97 remains a confirmed earlier reproduction. #3036 maintains exact durations, hashes, captured screens and limitations. The focused candidate remains synthetic-tested; no title-fix or common-cause claim.

[videoout] presentation failed: AGC graphics: DCC over a multisampled color target is unsupported (slot 0, samples 8, CB_COLOR_INFO 0x18066730, CB_COLOR_CONTROL 0x00cc0010, CB_COLOR_ATTRIB2 0x03bfc86f, CB_COLOR_ATTRIB3 0x4dc6c000, CB_COLOR_VIEW 0x00000000, CB_DCC_CONTROL 0x00100038)

[videoout] presentation failed: AGC graphics: DCC over a multisampled color target is unsupported (slot 0, samples 8, CB_COLOR_INFO 0x18066730, CB_COLOR_CONTROL 0x00cc0010)

Separate tracker-proof finding

The written-page source identified above led to #7220: on unchanged main production code, sparse BDA stores can falsely satisfy complete 4 KiB overwrite coverage. Independent wave32/wave64 tests reproduce 3/3 on both physical NVIDIA and AMD GPUs. The issue links the failing test branch and an experimental proposal. Its 54 earlier controls passed, but a new real-GPU producer control shows that proposal rejects legitimate full BDA-written native HTILE clears on both GPUs; it is not ready for promotion. The unchanged-main positive control and proposal regression are preserved on the fork and documented in #7220. This does not identify the actual bytes/values written in this title's metadata, prove that its resident-DCC refusal is wrong, or provide a DCC capability fix. Follow #7220 for the tracker semantics; this issue remains the title's current NVIDIA stop.

BDA producer narrowed, 2026-10-17 UTC

Observer a221350909, on the same mainf219 composition, connects each BDA page report with its resource's BDA shader-stage/variant candidates and the compute driver program mapping. It retains multiple program candidates rather than overwriting a shared variant with a guessed last shader; it does not change write classification or any clear/refusal decision.

The completed RTX5080 run reports six overlapping ranges, all compute (stage=0x20), variant32, with the single recorded program candidate 0x20b8b6900. One covers the whole CMASK range; five page-report intervals together cover the DCC range. Generations19217–19222 follow saved generation18683. This narrows the producer from generic BDA reporting to that compute resource/variant. These remain page-granular reports, not proof that every byte was stored, nor an instruction-level program attribution. The program candidate differs from the accepted HTILE kernel at 0x20b8b6700; its exact operation/required native-image transition is not yet established.

The guard still refuses with driver-overlap, DCC and CMASK snapshots remain identical to the table below, and the final inspected frame is NOW LOADING. Exit0xC0000409 after28.610s, seven frames, no timeout/input, verified 10de:2c02, zero observed Vulkan validation errors/warnings. The a515 conversion is retained. This replay reused the completed013 runtime's assets/user state rather than making a fresh-state A/B claim. Prior code, frames, active data and config were archived and 142 prior files were hash-verified restored afterwards; new code and captures are preserved separately, and protected originals/configs remain unchanged.

Full Windows/Linux relinker/library builds pass. Five relevant groups repeated3 times give 15 confirmed-device passes each on NVIDIA and AMD; the standalone BDA-device fixture also passes3 times in each selector run, but does not report its selected device and is not counted as dual-vendor evidence. Linux gives9 passes across three groups and3 explicit skips; its standalone fixture also states that CPU-Vulkan BDA execution was not tested. Static analysis parses all four changed TUs, with advisories in unchanged timer/capture-trace code and an unchanged ShaderResources null-reference path; these are untriaged, not a clean-analysis claim. No AMD title replay or new capability fix is claimed.

Earlier accepted-kernel observation, 2026-10-17 UTC

Diagnostic 013d65ff34 is based on refreshed integration5f7a/mainf219. It preserves the existing refusal and metadata matching/notification decisions. It adds bounded records at the existing accepted HTILE-mask dispatch recognizer and snapshots both metadata planes at the refusal.

A fresh isolated NVIDIA RTX5080 (10de:2c02) run exited 0xC0000409 after 29.102s, without timeout or input. Seven captures were saved; the inspected final frame remains NOW LOADING. The same DCC range is rejected with unchanged=0 stored=1 reason=driver-overlap.

The recognizer logged six dispatches, all at program address 0x20b8b6700: twice each to 0x402ea00000+0x400000, 0x404f600000+0x400000, and 0x4091b00000+0xa0000. None overlaps either plane above; the512-record cap was not reached. This excludes those six accepted cases as attribution for this range. It does not exclude other kernels, rejected recognizer variants, or other draw/dispatch consumers. The actual producer is still unknown.

Snapshot equality does not prove that no store occurred, a complete clear happened, or host-image reuse is safe. Same-value/partial stores and conservative page overlap can leave identical bytes. No guard was relaxed and no capability fix is claimed.

Validation of013d: full Windows/Linux relinker/library builds; native-depth BDA positive control plus2/4/8-sample DCC controls repeated3 times on each verified physical GPU (12 NVIDIA +12 AMD passes, no skips). Linux/lavapipe4-sample DCC passes3 times; native-depth and2/8-sample cases explicitly skip, accepted by the CI skip classifier. Both changed TUs parse in static analysis. Cppcheck reports the unchanged conditional-command-batch advisory; clang-tidy also reports a possible null call in unchanged PerformanceTimer.hpp:74, not yet classified against a baseline. Static analysis is therefore not claimed clean.

All127 built PRXs were copied from013d; the standard a515 executable/guest conversion was retained. Fresh original assets/user data/configs were copied into an isolated runtime; protected original inputs/configs remain unchanged. Khronos validation1.4.363 loaded, with zero observed validation errors/warnings. This is a diagnostic integration run, not an unchanged-main result. No new AMD title run, menu/gameplay or FPS result is claimed.

Previous snapshot46d7fb06ee, on main667/integrationc784, already found unchanged all-zero DCC bytes after28.644s. Its18 physical-GPU DCC passes and3 Linux passes retain that original pin; the current replay extends observation to CMASK and the accepted dispatch recognizer.

Next: attribute the retained driver-store overlap to its actual producer before implementing a metadata/coherence transition. #7220 tracks the separate coverage-proof mismatch; test-only #7228 protects valid full native-depth GPU clears and does not unblock this color-DCC stop. AI-assisted diagnostics, validation and reporting: Codex.

Historical handoff recovered with #7250, 2026-10-17 UTC

The old08891d938 menu/cinematic branch preserved constant-fill recognition plus a resident DCC clear transition. #6543 restored recognition to6146, but its no-observer Sonic replay still stopped at the same8x DCC guard after28.101s. The new existing #7250 implements the corresponding transition through today's NoteColorMetadataFill/resident-image backend.

Canonical fork6d76fb2098 now tracks main922 and composes #7250 head257a4707, preserving original authorship and all previous6146 fixes. Detailed validation: full Windows/Linux builds,21 repeated passes per verified NVIDIA/AMD GPU without skips; supported Linux/lavapipe groups pass,2/8-sample cases explicitly skip. The new refill tests fail before the change on both physical GPUs.

Sonic on RTX5080 now reaches PRESS START (120s no-input observation), then the CONTINUE menu and opening cinematic in a second240s observation (240.899s recorded,61 frames,four ordinary controller-file pulses). Both runs were deliberately terminated at their bounds with no fatal line; no gameplay/FPS or natural-exit claim. All127 libraries came from6d76; standarda515 conversion and prior isolated assets/user state were reused. This is not a fresh-state or isolated-main comparison. Protected originals/configs remain unchanged;142 prior files hash-restored after each run, new code/captures archived.

The historical insight has now recovered a concrete milestone on the current backend. Continue beyond the movie and use preserved historical calls/behavior to investigate the next refusal. No separate duplicate PR is needed for this handoff; validation is on #7250. Earlier producer/page-granularity limitations and the held fe08 proposal retain their original scope. AI-assisted: Codex.

Extended frontier: the same6d76 build now reaches the post-movie level-loading/tutorial screen, then exits after758.222s (no timeout) on a different refusal, mipmapped 3D color targets are unsupported.96 frames/four normal input pulses;142 prior files restored and originals unchanged. The earlier resident-DCC stop did not recur. Issue7260 owns this distinct missing path; gameplay remains unverified.