Actual scene trial: Pixman fallback and its rejection reasons

Two bounded board sessions on 2026-09-23 used the experimental wlroots renderer inside Sway’s existing DRM backend. Neither submitted a GPU frame. These are negative acceleration findings, not hardware promotion or normal-service GPU acceptance. The normal shell and seatd recovered after both sessions.

Each run’s result.json records exact immutable compositor/client paths, the installed system closure, boot ID, scanout format, sole observed Sway PID, exit status and service restoration. The trial used RGB565 at 568x1232, scale one. No image flash, device permission change, second DRM owner, or persistent service override was performed. The root diagnostic and its separately launched unprivileged synthetic client do not prove the normal shell service’s access boundary. Its prior credential-isolation prerequisite is recorded in ../vglite-isolation/.

Commands and retained evidence

The coordinator reserved /tmp/k230-board.lock, imported only missing Nix closure paths over the private network with matching SHA-256 and successful nix-store --import, then uploaded the committed per-run check.py and tools/vglite-root-scene-trial.py. The actual board invocation for the second run was:

/nix/store/v189xydz6qkcd4cbkixcmv91w8hbc560-python3-riscv64-unknown-linux-gnu-3.14.7/bin/python3 \
  /run/gpu-diagnostic-check.py > /run/gpu-diagnostic-check.out 2>&1

The copied checker contains the exact bounded harness/client arguments and native grim capture commands. The harness arms recovery before replacing the normal shell and tears down both diagnostic cgroups before restoration. The diagnostic decisions.log retains only the fixed numeric renderer records; the unrestricted root-private journal and transfer addresses are not published.

An initial attempt used an unsupported synthetic app ID and the client exited 64 before mapping. initial/rejected-app-id.json retains that failure. The corrected runs use the fixture’s supported k230.card.one ID. It is a test surface with independently changing parent and child buffers, not a product application or a card-shell acceptance run.

What the diagnostics established

Of 224 replay decisions, 220 rejected texture_transfer: the ordinary surface has transfer value 8 (GAMMA22), primaries present, luminance one, opacity one, and no YCbCr encoding/range. Four clear-only passes rejected target_attributes. Every logged target has one linear RGB565 plane, offset zero, and stride 1136 bytes. The renderer currently requires 64-byte stride alignment, so the actual width-times-two target fails even without textures. No record reaches result=attempt or result=gpu.

The full-size parent texture also has four clip rectangles due to child occlusion. Since the color rejection stops later preflight, these logs do not prove whether clip/coverage would pass after a color fix. Target allocation, clipping, and partial damage must be investigated independently; this finding does not justify removing their checks.

The subsequent source/host-tested exact default-color exception is documented in ../vglite-default-scene-color.md. It was not the package in these board runs, and does not resolve the observed target stride.

Visual inspection and limits

The coordinator inspected the native PNGs. The initial scene and diagnostic later capture show the synthetic striped parent and contrasting child with frame markers. The diagnostic first capture is entirely black: the checker waited for IPC mapping, which did not guarantee visible presentation before capture. It is retained rather than treated as successful content evidence. Future visual comparison must wait for a presented/recognizable frame.

These captures and logs establish actual scene fallback and its first rejection conditions. They do not establish exact GPU pixels, cache ownership, GPU frame timing, touch usability, optical behavior, or product card performance. Those proposal gates remain open. Pixman remains the normal default.

After the diagnostic, the coordinator ran restoration-check.py on the board using the same board Python path above. restoration-result.json records PASS for normal shell/seatd, injected Apps Next/Previous, keyboard show/hide while preserving Apps, Terminal focus, and Apps Back. The checker imports the retained pixel helper from docs/evidence/launcher-gestures/metadata-budget/check.py (deployed as /run/rollback_check.py) and uses the existing injected device. This is an installed-image recovery regression, not physical-finger acceptance or a measured UI latency trial. The initial helper’s CPU-heavy PNG processing time is not a response-time measurement.

Follow-up: actual GPU submission with padded pitch

The later padded scene trial reached 196 GPU frames and zero CPU replays. Its forced-Pixman comparison uses the same allocation. Source colors differ by a quantization step, and no speedup is accepted. The original zero-GPU records above remain unchanged.