Override 2: Super Mech League (PPSA01907)

[JJenkx]

Manual-play finding — severe graphics corruption (2026-10-17 UTC)

The subsequent user-controlled run on the same f1ba8bfd0 integration shows severe 3D scene smearing and repeated/distorted geometry, with the HUD still readable. This uses a different mech from the first-blue-mech controls check below. Multiple local captures preserve the failure; the root cause and whether it is mech-dependent are not isolated. The prior movement/punch check passed only its limited input/animation scope: graphical correctness fails in manual play, and this is more than excessive bloom. No fix or performance gain is claimed.

---

Latest verified status — Training controls (2026-10-17 UTC)

Training arena, controlled movement and a right-punch animation are confirmed. Exact source f1ba8bfd0734621c55e12d2e01107fa33be8713d is the public title integration rebased linearly onto main fd9e577d9821b42f6bf0443b1b569625be6776d3, excluding the #5609 and #6636 performance trials. This is not unchanged main, and individual title prerequisites have not been isolated.

On the RX 6650 XT / RADV Mesa 26.2.4 / Arch Linux, the full source-built build passes and all 14 focused checks pass with core Vulkan validation, no skips/errors/VUIDs, including the newly merged f32 compare modes, layered mesh/vertex rendering, draw reuse, shader disk cache, AvPlayer, fiber and late profile shutdown checks. The first test launch used the wrong runtime-library search path and failed symbol lookup before nine tests could run; correcting that launcher environment produced the passing results without a source edit. Tests.

The native run used no profiling or validation layers and no forced focus. Screen-verified normal controller input selected Training, the first blue mech and Skyview Center. After the arena rendered, a five-second left-stick command moved the mech from its starting point to the wall; the subsequent R1 command produced the punch animation visible below. Ordinary close at 481.260 seconds exits 0, process gone, temporary kernel limit restored to 64 and runtime command line unchanged. Source, controller timing, screenshots, checks and limits. The earlier main-f326 integration also reached the arena in both X11 and XComposite captures and closed normally; its earlier stage-selection screenshot did not prove a permanent renderer hang. Earlier run.

Remaining scope: very bright bloom/rendering correctness, opponent interaction and damage, audio, a full match and full-title completion remain unverified. Loading logged a zlib decode error and GPU occlusion-query timeouts in the earlier f326 run, yet the arena rendered; their causes are not isolated. The new run compiled shaders at startup. Neither run is a controlled main-versus-change timing comparison, and no FPS gain is claimed.

For the earlier #5609 investigation, the separate counter commit c56fd05d8 distinguishes successful buffer descriptor stage patches from alias hits. The startup/menu sample before the first controller input contains 38 patches across nine published windows. A patch can belong to a partly matched draw; it is not an executed draw or time saving. Rendered moved-buffer checks pass with and without graphics pipeline libraries; retained and unchanged variants add no patch. Both native diagnostic runs closed normally. Counter source · Results and counter semantics.

AI-assisted: yes. The account remains at 20 open PRs; new independent fixes and test work stay public without opening over the contribution limit.

---

The earlier reports below retain their original source revisions and limitations.

Existing thread

• I searched this category for PPSA01907 and the title; no existing title thread was found.

Title

Override 2: Super Mech League (PPSA01907)

Content version tested

01.007.000

AnyPS5 commit

Main b260d8d792c81a21a638340979ae2f98243547d5 (full base hash in the report), with #6913, #862, #6781 and local diagnostic changes. Exact native source: c01f24339a1b42a98b1b09d27ef5ad3b1e9c82ef. This is an integration report; it does not establish which of those PRs the title needs individually.

System tested

Arch Linux 7.2.9-arch1-1, GCC 16.2.1 Release, Ryzen 9 7950X, Radeon RX 6650 XT / RADV Mesa 26.2.4-arch1.1.

Furthest point

Menu.

What is shown

Readable menu UI over a very dark 3D scene. The diagnostic run's window reported roughly 18 FPS near the end, with synchronous target readbacks and validation enabled; this is not normal-play performance or an FPS comparison. Audio and gameplay input were not assessed in this run. The window closed normally: exit 0 after 97.64 seconds, with the temporary kernel setting restored.

Current blocker

The dark 3D rendering is unresolved. A native readback of the first completed draw to the 32×32×32 R10G10B10A2 color-grading target has populated slices 0–7 and zero RGB in slices 8–31. This matches the earlier frame replay observation, now without RenderDoc mapped and with zero Vulkan validation errors/VUID reports in the corrected native capture. It is a first-draw snapshot, not every frame, and does not yet prove the failing operation or the full cause of the darkness.

Replay diagnostics also show a 32-layer attachment and 1×32×1 mesh workgroups for this pass. Individual workgroups 0–7 reach their matching slices; sampled later workgroups (8, 15, 16, 31) have no marked raster output. Lighting-table fetches and one HDR lighting-volume sample match the corresponding captured CPU data for the probed pixels. Next: isolate emitted counts, primitive validity and vertex fetch bounds beyond instance 7, then make a synthetic regression.

A separate synchronous-path bug transitioned mip 0 regardless of the resident attachment's actual mip/layers. The exact source above includes PR #6934; it removes that layout validation error and does not fix the dark scene. No game code/assets or shader semantics are patched in the native run.

Native report, synthetic readback check and limits. Earlier replay/radiance evidence. Raw game captures, shaders, target pixels and game logs remain local.

How it was run

Used the existing converted title with its converted SDK/FMOD/party modules and the rebuilt libraries from the exact source above. The native supervisor verifies the loaded graphics-driver path and source/build revision, captures without focusing the window, and requests ordinary close after 95 seconds.

In the relinked title directory:

env LD_LIBRARY_PATH=<build>/core/libs/libs \
    ANYPS5_SHADER_CACHE_DIR=<cache> \
    APS5_PIPELINED_DRAWS=1 APS5_HOST_IMPORT_MIB=16384 \
    APS5_DRAW_CACHE_ENTRIES=4096 APS5_NO_SAMPLED_FROM_STORAGE=1 \
    APS5_DUMP_TARGETS=1 ./app.elf

The supervisor also selects RADV and the Khronos validation layer, and temporarily changes udmabuf.size_limit_mb from 64 to 8192, restoring 64 afterwards. Diagnostic flags add overhead. The regular no-target-dump repeat also closes normally with zero validation errors, but does not read back the resident color-grading volume.

AI-assisted: yes.

Mesh layer limit isolated (2026-10-16)

The LUT draw is submitted as vkCmdDrawMeshTasksEXT(1, 32, 1) into a 32-layer view and render pass. The RX 6650 XT / RADV Mesa 26.2.4 device reports maxMeshOutputLayers = 8 (ordinary framebuffer limit: 1024). Vulkan defines mesh output to a layer at or above this limit as undefined.

Offline diagnostic triangles placed into separate tiles on layer 0 confirm that all 32 workgroups execute and compute six vertices / two primitives. Moving those diagnostic triangles to their respective destination layers writes only slices 0–7, even with the original shader body bypassed. Every replacement was removed afterwards and the original target hash verified. This changes the investigation from missing mesh counts to exceeding the host mesh layer limit. Guest code/assets were not modified.

A synthetic 96×16×32 RGBA8 fixture reproduces the cutoff independently of the title: all 32 vertex-rendered layers pass; the mesh draw fails on RADV at layer 8 (initial 64 remains instead of expected 255). An eight-layer mesh control passes on RADV. Core validation emits no errors, but does emit maintenance9 forward-compatibility barrier warnings. Lavapipe passes the vertex control but its mesh execution crashes for both 32 and eight layers; that crash is not evidence for this cutoff.

Validation report, synthetic reproducer patch, RADV 32-layer log, eight-layer control log. Source is preserved on the diagnostic branch, based on main 9e9d86f81 plus #6934. It intentionally reproduces unsupported output and is not a merge-ready fix.

Next: implement general layer-window rendering within the host limit, preserve guest layer values for fragment ancillary input, and rerun the native title. No visual fix or FPS gain is claimed yet. AI-assisted: yes.

Candidate layer-window implementation (2026-10-16)

A candidate fix is now available at b73f874b5, based on main 9e9d86f81 plus #6934. It renders oversized layered mesh color draws in windows within maxMeshOutputLayers, remaps and culls mesh outputs, and restores the guest layer number for fragment ancillary input. Monolithic pipelines, graphics pipeline library keys and journal state distinguish the windows. Shaders with guest memory writes, clocks, GDS access or unproven replay effects are refused until an output-capture path exists; layered depth remains unsupported.

The synthetic 32-layer case passed. The final fixture uses a 99×16×34 RGBA8 volume with 33 visible layers starting at slice 1, so it covers an attachment offset and a one-layer final window. Every RGBA byte is checked for vertex and mesh white bands, fragment reads of the original layer number, and untouched regions. Zero-vertex/zero-instance draws also retain their no-op behavior even without replay certification.

On RX 6650 XT / RADV Mesa 26.2.4, all four requested combinations of graphics pipeline libraries enabled/disabled and synchronous/normal recording configuration pass with no core-validation errors. Maintenance9 barrier warnings remain; the legacy render-pass path also warns about a Layer-declaring shader on the single-layer final framebuffer. Seven selected regression tests pass with no skips, including graphics, mesh, replay-safety, disk cache and pipeline journals. Full libs builds.

A production-only negative control disables the window dispatch while retaining the 33-layer fixture and shader changes: it fails at physical slice 9 / guest layer 8. Restoring the dispatch passes. The test binary is unchanged across that negative; the subsequent empty-draw follow-up was validated separately.

Candidate validation report, candidate patch, final regression log.

Historical candidate status: no native retest had been completed at that point. The corrected native results are recorded below. Array-target/input-alias review and native LUT/readback plus visual testing remain before upstream submission. No FPS claim. AI-assisted: yes.

Native LUT correction verified (2026-10-16)

The corrected integration now fills all 32 LUT slices. Full-volume snapshots after the four host layer windows show 8, 16, 24 and 32 populated slices, and the bytes from earlier windows remain identical in the final volume. Source: d26ec8200, based on main 2fab88d10, with the same separately declared API fixes (#6934, #6913, #862 and #6781) and diagnostic readbacks. Peer authorship is preserved.

The first candidate stopped at 108.88 seconds with exit 134 because its replay check treated usesDma as a write flag. ResourceTracker also sets it for read-only address access. The corrected classification accepts read-only loads and continues to reject actual address/buffer/image writes, atomics, clocks, GDS and unproven effects. A regression distinguishes those cases. All 181 historical captured shader requests recompile with the corrected classification; that is shader evidence for this title, not 181 title tests.

Two corrected native runs completed:

• With full-volume target readbacks: normal close at 180.08 seconds, exit 0 at 182.08 seconds; all 32 LUT slices verified.
• Normal rendering, target dumps and forced synchronous draws disabled: normal close at 91.58 seconds, exit 0 at 93.58 seconds. The menu displays the robot’s blue/white materials, illuminated signs and floor reflections.

Both corrected runs report zero core-validation errors and no VUIDs, restore the temporary udmabuf limit to 64 MiB, and leave no native process. Khronos validation is enabled and RenderDoc is absent in the verified readback process. Maintenance9 compatibility and unconsumed-output warnings remain. No gameplay, audio-quality or FPS-gain claim. The normal repeat uses a warm cache, and the earlier baseline screenshot uses a different integration base and animation phase.

Native validation report, earlier dark-menu screenshot.

The production fix is submitted as PR #6959, with #6934 declared as its included prerequisite. It excludes the diagnostic and unrelated API changes in this native integration. The published branch passes nine selected RADV checks, all four explicit mesh execution configurations, and an unchanged-binary negative control. Local lavapipe passes vertex/compile controls and explicitly skips mesh execution on its narrower host subgroup layout. Layered 2D arrays are still rejected by the existing decoder, and layered depth/non-replayable shaders remain unsupported. AI-assisted: yes.

Current-main refresh and menu input check (2026-10-16)

Production #6959 is rebased cleanly onto main e9a3343a2, head 5f4891387. Full build passes; nine focused RADV checks and 14 memory/loader/voice checks pass. The broader set remains 626 pass / three skip / the reported Mesa tessellation SIGSEGV (#5444). Hosted full Linux/Windows builds still need maintainer dispatch.

A separate clean native integration, 9a294a6bc, adds #6913 and peer #862 (lvmino97) / #6781 (XairRrR), preserving authorship. Eleven selected integration checks pass. Normal pipelined rendering, core validation, new empty AnyPS5 shader cache, existing Mesa cache, APS5_NO_SAMPLED_FROM_STORAGE=1, no target dumps or forced synchronous draws. The helper did not request window raising or focus.

The corrected lighting remains visible and controller inputs reach Select Mech. This 303.55 s run closes normally with exit 0 after the planned close at 301.54 s; zero core-validation errors/VUIDs, temporary kernel limit restored to 64 MiB, native PID gone. The run does not establish Training or gameplay: later snapshots remain at Select Mech. Rendering is slow (exploratory title-bar samples 2.29–4.30 FPS from 150–300 s); no matched performance trace or cause is established. Next work is to profile the warm scene and resolve the character-selection flow. No FPS gain, audio-quality or new LUT-readback claim.

Current native report · Integration tests

Warm native menu trace and f32 correctness work (2026-10-16)

A completed, normally closed run at 46458ceb458465ef35e07ab4af25e580a85b4e2a records the warm animated main menu, silver/gray mech, on RX 6650 XT / RADV 26.2.4 / Arch Linux. Over 45.004 seconds: 1,996 presents (44.35/s), frame p50 22.361 ms, p90 25.318 ms, p99 28.577 ms; GPU busy 18.181 ms/present. Core validation is disabled, output goes to a PTY, only APS5_TRACE_SECONDS=100:45 is enabled, and no input is sent. The source adds #6341 tracing to the preceding native integration. Earlier low-FPS runs used different validation, cache, output and mech conditions; this is a baseline, not a measured speedup. Draw-committer and queue0 timelines overflowed, so their timeline-derived averages are incomplete. Training/gameplay remain unverified in these runs.

Exact baseline source · Trace · Digest · Summary · Scene

A subsequent testing integration rebases #6503 and fixes its request-version collision with main's spanned-wave field. Additional checks exposed the existing mode-3 f32 preservation gap with hardware flushing disabled. At ea0b9c77bf8ef4b561f39cb58633126ef9d3f167, preserving shaders are excluded from hardware flush, and supported devices explicitly request DenormPreserve 32. Four numerical add/multiply checks pass. The full build passes; focused core-validation checks pass 12/12 with hardware flush and 12/12 forced software, with no validation errors or numerical skips. The broader suite has 628 pass, three skip and the previously recorded Mesa tessellation SIGSEGV. Devices lacking independent f32 preservation retain the documented gap; modes 1/2 remain refused. This commit has not had a native title run or paired performance capture, so no title gain is claimed.

Testing source · Correctness report and scope · Hardware validation · Forced-software validation

Standalone f32 preservation fix on upstream main (2026-10-16)

The preservation fix above is now isolated as one commit, ec019f3b646f4edb02a6690589dfdb425ccf9f6b, on upstream af0fff762. It has no dependency on #6503 or the game integration. An attached five-line numerical fixture fails on main 40255d774 with production sources unchanged: max-subnormal plus max-subnormal returns zero instead of 0x00fffffe.

The standalone full build passes. All 13 focused RADV/core-validation checks pass with no validation errors or numerical skips; the broader suite has 629 pass, three skip and the same recorded Mesa tessellation SIGSEGV. Lavapipe passes compatibility checks, with its numerical preservation subcase skipped because independent f32 preservation is unavailable. Modes 1/2 and unsupported-device limitations remain documented. No native title run or performance gain is claimed for this standalone source.

Standalone source · Exact control, results and limits

Hardware/software f32 timing follow-up and diagnostic shutdown fix (2026-10-17 UTC)

The source-built integration on main 29a546d37, exact head 871fdc7cd5b27ac993c29998a4abe63e740fa68e, completed two warmups and four native timing captures. All six ordinary closes exited 0, with native PID disappearance and kernel limit restoration verified. Hardware/software f32 flushing was toggled in the same binary; this is an exploratory integration comparison, not an isolated main-versus-PR speed claim. The captured GPU reports hardware flush with 0/5,472 probe mismatches.

Capture	Mech at 80 s snapshot	Presents/s	Frame p50 / p90 / p99 ms	GPU busy ms/present
HW B1	silver	45.52	21.571 / 24.146 / 29.389	15.262
SW A1	silver	43.75	22.497 / 25.808 / 30.172	18.229
SW A2	blue; excluded from silver comparison	43.63	22.658 / 25.822 / 29.421	17.461
HW B2	silver	39.39	25.489 / 29.437 / 38.196	16.285

No repeatable FPS gain is established. Hardware GPU busy time is lower in the silver samples, but the second hardware FPS result regresses. Snapshots do not prove scene identity throughout every captured frame. All four timelines dropped events; aggregate summary totals remained complete (totals_full=false), while timeline-derived thread averages and slowest-frame CPU attribution are incomplete. Inclusive draw/resource scopes suggest further CPU investigation, but do not establish a cause. Resources.Complete builds descriptors/images/resources; it is not GPU writeback. Shader compile/link scope includes cache and draw-stage work and does not establish fresh shader compilation time.

Exact timing source · All four summaries, traces, digests and limits

A separate APS5_PROFILE_DRAW=1 diagnostic aborted on ordinary close with profile output is stopping. An unchanged minimal finalizer reproducer fails with the same exception on main's header. The independent fix 107d18ba8 on main 85643f75e drains/joins the diagnostic worker, then writes late static reports synchronously through a sink retained until process exit. Owned logger Stop() remains strict. The full runtime build, four focused tests and ThreadSanitizer pass; the unchanged reproducer exits 0 and emits its late report.

Native verification adds only that fix to the failing integration, head 4f706559400e690926cddce8bb7c883afaf75ad9. Ordinary close at 201.927 s exits 0; the process is gone, final resource counters appear after the pipeline-journal report, and the kernel limit plus temporary command-line override are restored. Four storage-image destruction warnings about pending GPU results remain. This verifies one profiled shutdown, not GPU writeback correctness, gameplay or a performance gain. -FixedSeed did not stabilize the menu mech across the two diagnostic launches, so it cannot be used as a same-scene guarantee.

Standalone shutdown fix · Exact native integration · Red/green and native validation · Sanitized final diagnostics

The standalone shutdown and f32-preservation fixes are public and queued for upstream PRs: this account is currently at the contribution limit, 20 open / 15 merged. No additional PR has been opened over that cap. Training/gameplay and repeatable performance improvements remain outstanding.

Address-based draw reuse trial (2026-10-17 UTC)

Rebased the title integration onto main e751dd205 and tested #6636 at its current head 89ab0ea75a987e130068334b462d2373558ec2dc, with original authorship retained. Exact integration ec39b59a5ba7e7d7dc3274cf51d6e01e983bc95c builds in full. RX 6650 XT / RADV Mesa 26.2.4 / Arch Linux: all 14 focused core-validation checks pass, including address-based reuse, with no skips, errors or VUIDs. The broader 640-test sweep has 636 pass, three skip and the previously observed tessellation SIGSEGV.

The profiled native menu run closes normally at 201.913 seconds, exit 0; native PID gone, kernel limit restored, runtime command line unchanged. Across 18 published counter windows, the address-based path accepts zero reuses: 45,632 proofs refuse changed snapshots, and 346,450 lookups find an earlier draw still in flight. These window sums omit any last unreported partial interval. This establishes that the path is reached, but no benefit for this title scene is shown. No same-scene unprofiled main/change comparison or FPS gain/regression is claimed, and #6636 is not being added as a demonstrated performance requirement.

Exact trial source · Hardware tests, native result and limits · Counter windows

The next candidate is #5609's stage relocation when buffer addresses move; its #4943 prerequisite is now merged. No result for that candidate is claimed yet.

Stage-relocation trial and stronger rendered regression (2026-10-17 UTC)

Tested #5609 head 09238da8206e0aa70d8966c69f822c297da36308 on the title integration based on main e751dd205, independently of #6636. Original peer authorship is retained. Integration adaptations preserve main's optimized-pipeline scheduling and current result layout; the combined disk-cache schema uses format 40.

I strengthened its moved-buffer regression in 5477e1cf4: an existing full-screen triangle fixture now verifies actual rendered bytes after moving to buffers containing 1.0, 0.25 and 0.5, alongside the cache-hit counters. Suppressing only the relocated descriptor-word update fails at 255 rendered, 64 expected. Restoring it passes with the identical test executable, both with and without graphics pipeline libraries and no core-validation errors. This is an injected negative control, not a claim that the original PR has that defect. The new check makes stale-binding regressions observable.

Rendered test commit · Rendered test patch for the #5609 branch · Red/green evidence and limits

The full build passes. Final focused hardware checks pass 23/23, with no skips, errors or VUIDs. The broader sweep on the same rendering code before the test-only addition has 636 pass, three skip and the previously observed tessellation SIGSEGV. Native menu runs at 5477e1cf4 and 53da10dbb both close normally after the planned 200-second diagnostic, exit 0, PID gone, kernel limit restored and runtime command line unchanged. Four storage-image pending-results shutdown warnings remain in each run. These are profiled diagnostics, not performance captures; no gameplay or FPS gain/regression is established.

The first run exposed a reporting gap: relocation counters printed only when user-word patches had been produced and a 100,000-lookup threshold was reached. 53da10dbb emits them with the regular draw profile report; three relevant hardware checks pass with no validation errors. The second run also enables the existing detailed draw-cache trace. Its published windows sum to 784,064 lookups, 781,273 without an entry and 2,248 complete hits. Of the missing-entry cases, 712,897 are classified as user-word/pointer changes. The alias report shows 33,084 lookups and no accepted moved-user-word donors. Those alias counters do not count every buffer-only patch or existing data-only stage reuse, so they cannot establish zero buffer-stage reuse or attribute a performance result. Window sums omit the last unreported partial interval. The runtime reports the same cache stamp in both runs and reuses cache entries on the second run.

Exact integration · Build/test/native scope and complete counter data · Detailed diagnostic output · Reporting fix patch

The reporting fix is also isolated directly on main as one commit, 03dac4ba4e51c0ed74ea6c360fcd89fbbc6ecae3, with no #5609 dependency. Its full runtime libraries build, and both draw-cache hardware checks pass with core validation and no errors/VUIDs. Standalone source · Standalone validation.

Next: separate buffer-only patch counts from alias hits and use the changing-user-word/register evidence to target CPU work, then validate any proposed gain with a reproducible scene and unprofiled traces. The account remains at 20 open PRs, so the new test and reporting commits are public without opening another PR over the contribution limit.

Image-history candidate tested on AMD (2026-10-17 UTC)

Investigated #6868 (aa8f734907686de67c0f16924e84ef5b4762f606, original authorship retained) because it fixes partial driver-store tracking that can replace GPU-written image data with old memory contents. Exact title integration 7ed307ec3fe908892fee7ad5856dcff26a55afec is linear on main f643e61d9cb60325a38cb727d3c6715d57d60a9d. The source-built runtime libraries and both test targets build; eight focused RADV/core-validation checks pass without skips/errors/VUIDs.

On the RX 6650 XT, restoring only pre-PR GuestMemory.cpp makes both new tests fail: the bookkeeping check reports false overlap, and the watched-host-import fixture loses 2,048 of 524,288 GPU-written texels. Restoring the peer production code passes both checks three times with the identical recorder executable SHA-256 c854e9c995f65478109b2554f5775c59be5e0a17cf1386d519cae2f487e43e24. Temporary control changes were removed and the source is clean. This independently reproduces the actual image-data loss and its fix on AMD; it does not identify this title's failing resource.

The native attempt selected the gold AURA mech, whose portrait matches the failed manual-play capture, but entered River, not the failed Skyview Center scene. Selected snapshots show camera rotation/movement without the earlier severe smearing; this difference in scene prevents attributing that result to #6868. Ordinary close at 361.562 seconds exits 0, PID gone, temporary kernel setting restored and runtime command line unchanged. The title's graphics failure remains unresolved. Next: the same AURA/Skyview Center scene and controls with and without only this peer change. No performance claim.

Public source · Scope, red/green results and limits · Eight focused checks · Negative control · Restored three-repeat checks.

Reproduction baseline on refreshed main (2026-10-17 UTC)

Rebased the title prerequisites onto main 9bf01db663445820687e85fdbff08959d5193a5a, exact baseline 77cb07e494b572dbb68f40185b88c6ba21f01eba, without #6868. The source-built runtime libraries and selected graphics targets build; all 16 focused core-validation checks pass, including main's new ldexp denormal modes. The original graphics failure is still unresolved.

AURA and Skyview Center were verified visually before loading; the engine loaded Maracanan/Maraca_Map. The recorded camera-right, forward movement, right-punch and camera-left sequence completed, including a 15-second settled capture, without reproducing the severe manual-play smear in the selected images. Bright bloom and other rendering artifacts remain. Ordinary close at 601.941 seconds exits 0; process gone, kernel limit restored, command line unchanged. The window geometry changed from 1536×864 to 3072×1728 after the scripted exercise; its cause was not established. This is not a performance capture.

A second run explored kicks, both punches/kicks, dash, jump, grab, combined camera/movement and the visible L3 Ultimate action. It entered River after a confirmation retry and does not count as a Skyview comparison. Its checked captures also did not reproduce the original severe scene duplication; it closes normally at 541.294 seconds with the same cleanup checks. The menu path needs fresh-screen waiting before retries; remembered mech selection and delayed transitions make fixed-time input assumptions unreliable. No graphics settings, guest code/assets or shaders were patched in either run.

No #6868 title fix is inferred: the baseline did not reproduce the failure with these inputs, so a peer-enabled run cannot discriminate the suspected cause yet. The independent AMD red/green image-data-loss evidence above remains valid. Next establish a failing baseline using the original scene/input conditions before attributing a rendering fix.

Exact baseline source · Run scopes, inputs, captures and limits · 16 focused checks · Settled Skyview baseline capture.

[pushkarreddyy]

Taking this up: investigating the 3D scene rendering corruption on Override 2: Super Mech League (PPSA01907). Examining descriptor attachment transitions and draw pipeline resource bindings to isolate the corrupted draw state.

[pushkarreddyy]

Diagnostic Analysis — 3D Smearing & Dynamic Resolution Extent Transitions

Examining the manual-play smearing capture and the reproduction baseline logs:

1. Dynamic Resolution Scaling & Extent Shift: The transition from 1536x864 to 3072x1728 indicates the engine engages dynamic resolution scaling upon gameplay initialization. When the render target extent shifts, resident storage images in CachedStorageSurface retain the earlier allocation dimensions unless explicitly invalidated by layout recreation.
2. Subresource View Coherence Across Draw Passes: In Draw.cpp (recordDraw), when the camera rotates or performs fast movement, intermediate velocity/temporal post-processing buffers sample from the preceding frame's scene color target. If the attachment view transitions only slice 0 / mip 0 while the mesh draw outputs across layered attachments (or vice versa, as noted in fix(agc): transition resident attachment subresources #6934 / fix(agc): render layered mesh targets within device limits #6959), the prior frame's unevicted texels smear into the motion accumulation buffer.
3. Descriptor Invalidation Verification: To isolate whether the geometry smearing is driven by temporal history accumulation or unevicted attachment subresources, running with APS5_SYNC_DRAWS=1 or forcing fixed window geometry across the sequence can isolate whether the artifact persists independent of dynamic scaling.

[EagleBearerYS]

@pushkarreddyy, a question about the 1536x864 ↔ 3072x1728 switch in Override 2: Super Mech League (PPSA01907): is that the size of the display buffers the game flips, or only of its own render targets? A run's log answers it. Each time the flipped buffer's size or the window's size changes, AnyPS5 prints a line like [mouse] video driver <name>: guest 3840x2160 presented at 1500x844, delta scale …. The tag is the mouse's, but guest is the size of the display buffer being flipped. Could you post those lines from a run that goes through a switch? If guest drops to 1536x864 while the window stays larger, the game is presenting below its output size.