Record of Lodoss War -Deedlit in Wonder Labyrinth- — PPSA05347

[JJenkx]

Linux/Radeon bring-up for PPSA05347 v01.000.001, tested 2026-10-15. The recorded main control renders the title, save/new-game menus and opening story with readable text and sprites. Gameplay, movement, combat, audio quality and a full playthrough remain unverified. The earlier native runs below include a local SPIRV-Tools repair, disclosed below; the latest check uses the official dependency pin with the build-applied repair in #6476.

Reproducible validator repair — #6476

The validator type-table lifetime repair is now proposed upstream in #6476, using AnyPS5's existing guarded build-source patching mechanism instead of a private SPIRV-Tools dependency commit. The official dependency gitlink remains unchanged.

A fresh native check on main 0029e62f9 plus that patch (407de2d78fec6d840ca495be54cc291de07dc52b) rendered readable opening-story text, accepted the same five scheduled Cross inputs, and closed normally at about 98 seconds (exit 0, prior kernel import limit restored). The final PR was rebased onto ec1193dc3, rebuilt, and passed 13 focused SPIR-V checks plus 2 existing checks on each of RADV and lavapipe. The native check predates that final rebase. Three negative controls isolate the repaired validator tables: restoring either dynamic table or the original validator passes warm validation and fails during exit.

This removes the private dependency checkout from that check's reproduction requirements. It does not establish a general title shutdown fix, gameplay stability, or an FPS gain. #6476 is proposed, not merged.

Current-main behavior

Upstream main 0c4816f61 plus that common dependency repair, private native head ed7b36dffda9c628cacd32b1027a6dda844e61a7, reaches the horse/riders opening-story page after five scheduled Cross pulses (requested 12, 25, 40, 55 and 70 seconds). The matched control closes normally after an ordinary window-close request at about 97 seconds. Shutdown remains intermittent across this investigation: other recorded main integrations abort on close, with the cause unisolated.

The checked opening-story scene runs at about 30 presents/second, with GPU busy time around 3 ms per present. CPU work, synchronization and pacing need further investigation before assigning a root cause.

Scattered-import investigation

A diagnostic found a rejected 8 MiB host import combining 121 registered ranges. This led to testing the existing scattered-backing work in #4449. That upstream branch conflicts with main; draft integration JJenkx/AnyPS5#77 provides a port onto 0c4816f61, retaining its author's commit, current import-watch callbacks, chunk/sparse paths, and two fixture follow-ups. It is optional investigation work; the main control already renders the intro.

Public source e31645c3ad8f03163874926887032ec00c07c6e1; private native validation d08cb0072eb051a984e397f5eb48184282763463. Full Linux Release libraries and selected executables build. Eleven focused CTests pass on Radeon and the same eleven pass on lavapipe, zero skips. The stronger kernel fixture reverses physical mappings and requires three backing runs in guest-address order, checks their file-backed bytes, and rejects a short output array. Deliberately reordering those runs fails the unchanged regression executable; restoring the implementation passes. Radeon host-pointer and DMA-buffer probes report zero stale pages after GPU stores over CPU-written bytes.

A matched warm-cache pair uses APS5_TRACE_SECONDS=75:10, the same input schedule, 3840x2160 guest output and 2972x1672 native window:

	Main control	#4449 port + fixture follow-ups
Presents	301	300
Window ms	10015.546	10021.367
Frame p50 ms	33.279	33.446
Frame p99 ms	42.687	40.082
GPU busy ms/present	2.992	3.000

No FPS gain or title unblock is established. Both traces complete without dropped events. Screenshots are outside the timed window; the game is active in every measured-window sample, without forced focus. No other profiling switch or game-log redirection. The candidate's earlier warm-up closes normally, but its matched run exits 139 after the ordinary close request, after its trace completes. This does not establish sustained stability or fix the shutdown problem.

Timing-only traces, summary lines, digests and synthetic-test hash manifest.

Environment and limits

Ryzen 9 7950X, 128 GiB, RX 6650 XT 8 GiB, RADV Mesa 26.2.4-arch1.1, Linux 7.2.9-arch1-1. Software-driver checks use lavapipe Mesa 26.2.4 / LLVM 23.1.1. Native host imports and pipelined draws stay enabled, import budget 16384 MiB, temporary udmabuf limit 8192 MiB; each completed run restores the limit to 64 MiB.

The public and private candidate trees differ only by private SPIRV-Tools teardown repair b326807c0c9d8682b697fdad3153f77fc2642f5f, excluded from the public source. These runs do not independently validate the official dependency pin. Windows/full-suite coverage remains unverified for this port. #6341 separately adds visibility of the draw committer to CPU traces; it is not included in the matched pair above. #6442 fixes a separate failed-window cache invalidation case, with no title gain established.

No proprietary assets, screenshots, binaries or raw title logs are attached. AI-assisted implementation, orchestration and analysis; native tests on my machine.