Portrait HDMI rendering checkpoint

This change is experimental and open. No performance winner or default image has been selected. The physical acceptance tasks remain unchecked.

The HDMI trial uses Sway output transform 90 at physical 1920×1080 and logical 1080×1920. Pixman performs software rotation. The renderer candidate does not change the panel driver. The HDMI mapping now compensates absolute touch calibration independently, as described below. The committed DRM dump advertises rotation only on a YUV video overlay, not the RGB desktop plane.

The opt-in card-shell-hdmi-trial package contains two alternatives:

• WLR_PIXMAN_QUARTER_TURN=1 freshly copies eligible textures in tiles before ordinary composition; scaled bilinear input retains the stock sampler.
• WLR_PIXMAN_OUTPUT_TURN=1 composes a logical frame into per-pass scratch and rotates it once directly into the physical target. Normal/unsupported formats, visible software cursors and damage highlighting retain the original path.

Scratch allocation is capped at 64 MiB. Neither path changes normal defaults. Per-draw and final-copy traces are bounded; frame CPU measurement must include the entire copy. Touch diagnostics record routing coordinates and event origins without device names, application text or network settings.

Recovered source proof

After the execution environment discarded /tmp checkouts, the original ten HDMI implementation files were recovered from their intact worktree. In the persistent checkout, these commands passed:

python3 tests/test_pixman_quarter_turn.py
python3 tests/test_card_shell_telemetry.py
python3 tools/hdmi-shell-performance.py --self-test

The pixel test passed 1,152 exact comparisons against stock Pixman, including crop/scale, opacity, rounded clips, damage and source updates. See recovered output. The telemetry test and eleven strict parser/dispatch fixtures also passed. The capture and scene entry points now delegate to their bounded helpers. The board helper has separate mocked capture/cleanup fixtures; these are host proof, not observations from the board.

The reconstructed final-frame patch now compiles as part of the native pinned wlroots build. A socket-free public-render-pass probe confirms rotation direction and byte-exact nearest-filter output for both quarter turns and ARGB8888/RGB565. It also reproduces bilinear differences; see the probe and measured limits. This is small-surface renderer proof, not a running Sway scene or board proof. A prior headless run was blocked by sandbox Unix-socket bind() returning Operation not permitted; it was not passing scene evidence.

Correctness and physical limits

A prior host diagnostic compared normal composition followed by final rotation against stock per-texture affine rotation. All 576 nearest-filter cases matched; 336 of 576 bilinear cases differed, by up to three ARGB8888 channel steps or one RGB565 channel step in the tested patterns. The diagnostic source/build artifact was lost with /tmp and must be reconstructed before this finding becomes a reproducible acceptance result. It is recorded as a limitation, not as a passing pixel gate or permission to relax the requirement.

The first per-texture board trial remained slow. The earlier committed board baseline also shows expensive rotated software composition. Neither establishes the proposal’s controlled 24-drag budgets for the final-frame candidate.

The operator reported that app-edge gestures worked after recognizing the touchscreen’s 90-degree orientation. This does not complete the named physical capture, keyboard/shade checks or latency gate; no recognition threshold change is justified by that earlier observation alone. The later operator correction below proves a separate orientation mismatch.

Next gates are reconstructed scene/pixel proof, exact cross-build identity, reserved one-/two-card physical trials, and normal-panel recovery. Mainline kernel porting proceeds separately. No archive is justified at this checkpoint.

Reproducible capture entry points

The capture helper runs on the board, where the live Sway IPC socket and compositor journal are available. Reserve the board before invoking it. Stage tools/hdmi-shell-performance.py and tools/hdmi-shell-capture-board.py together. Supply the exact source revision used to build the installed compositor; the helper also records hashes of the actual executable and mapped Pixman library. A supplied revision alone does not authenticate binary/source correspondence.

For injected profiling with exactly one or two already-open Foot windows:

python3 tools/hdmi-shell-performance.py --capture --variant candidate \
  --input injected --source-revision DEPLOYED_COMPOSITOR_REVISION \
  --seconds 120 --drags 24 --output NEW_CAPTURE_DIRECTORY

For physical contact, use --input physical --operator-confirmed-contacts. After READY, swipe up from the ordinary app into overview, then make 24 center horizontal drags. Record the operator observation separately. Baseline requires both rotation opt-ins off; candidate requires an opt-in in the actual process environment. Run one- and two-card workloads in separate new directories. The helper never launches/closes apps or changes output settings; it stops the benchmark, returns from overview and checks output/window counts after capture. Failed cleanup and incomplete traces are not accepted measurements. Their allowlisted capture.log is retained for routing diagnosis, while metadata.json has no accepted runs.

The host scene entry point currently exercises the per-texture candidate, stock 90-degree sampling, and 180-degree fallback with a changing native client:

python3 tools/hdmi-shell-performance.py --check-scene --variant candidate \
  --sway UNWRAPPED_RISCV_SWAY --client NATIVE_ANIMATED_CLIENT \
  --output NEW_SCENE_DIRECTORY

It requires QEMU user emulation, Grim and permission to bind Wayland sockets. It is not proof for the final-frame candidate. The current runner cannot bind those sockets or reach the Nix daemon, and has no serial device, so runtime, cross-build and physical tasks remain open.

HDMI touch axes correction

On 2026-09-29, the operator clarified that touch was still rotated: a physical left-to-right swipe triggered the drawer’s bottom-to-top gesture. This supersedes the earlier inference that recognizing the screen orientation had resolved touch mapping. No successful physical acceptance is claimed here.

Pinned wlroots types/wlr_cursor.c applies the mapped output’s transform to both touch-down and touch-motion coordinates. Sway sway/commands/output/transform.c inverts clockwise CLI degrees into Wayland anti-clockwise enums. At CLI transform 90, the resulting enum270 maps raw (x,y) to (y,1-x): left-to-right motion becomes upward motion.

The board’s glass retains its native portrait axes while controlling the external monitor. nix/hdmi-touch-calibration.nix supplies the compensating affine matrix for each supported HDMI transform. For the current 90-degree profile it is 0 -1 1 1 0 0; wlroots subsequently maps the calibrated point back to (x,y). nix/shell.nix applies it only when HDMI is active, after mapping touch to it, and explicitly restores identity calibration in the panel configuration on reload. This uses Sway/libinput configuration and does not change the touch driver or gesture thresholds.

A socket-free native test evaluates the Nix matrices and feeds synthetic calibrated contacts through the compiled pinned wlroots cursor/touch APIs:

python3 tests/test_hdmi_touch_mapping.py \
  --wlroots-source PATH_TO_BUILT_PINNED_WLROOTS \
  --output docs/evidence/hdmi-shell-performance/touch-mapping-native.json

All four rotations passed nine points each for both down and motion, including corners, horizontal and vertical swipes, and output-layout coordinates. The uncalibrated control failed the expected coordinate assertion for every rotated output. Results record the actual library/source hashes and limits. The affine libinput step is simulated; installation, device calibration support and real glass remain unverified.

For a reversible live trial on the current transform-90 board session:

SWAYSOCK=/run/shell/sway-ipc.sock swaymsg 'input type:touch calibration_matrix 0 -1 1 1 0 0'

The operator must confirm that upward glass swipes open overview and horizontal glass swipes remain horizontal. The prepared Nix source has not been built or installed by this constrained runner. The normal panel remains a separate identity-calibration case requiring its named regression check.

Access-restored board and cross-build checkpoint

Tool permissions were subsequently restored. Both the trial package and ordinary system cross-build succeeded; build identities record the exact source revision, command and store outputs. The runtime calibration was applied over the reserved physical serial console and get_inputs reports the expected six values; see live board configuration. This is configured state observed on the board, not real-finger acceptance. Coherent-profile deployment and the operator’s direction check remain separate gates.

Activation regression and correction

The first coherent-profile activation returned touch calibration to identity. The generated follow-output script passed the matrix as separate swaymsg arguments; its negative coefficient was parsed as an option (invalid option -- '1'). The interactive command had quoted the complete IPC message and therefore worked. Both generated input commands now quote the whole message. The operator reported the pre-activation direction fix working, then reported rotation and performance regressing after the profile switch. The faster per-texture trial was restored and the quoted calibration reapplied.

k230-coherent-shell-hdmi-trial explicitly retains the opt-in per-texture renderer in a reproducible system profile (WLR_PIXMAN_QUARTER_TURN=1, WLR_PIXMAN_OUTPUT_TURN=0), rather than depending on a runtime service override. The ordinary coherent profile remains the performance comparison baseline. Neither this profile nor the operator’s qualitative speed report meets the unchecked frame budgets, full scene correctness, or normal-panel regression requirements. Physical trial results and deployment identity are recorded separately when observed.

The corrected profile’s real activation is recorded in profile-activation.json. Its startup script applies the compensating matrix automatically, and the running compositor is the explicit per-texture trial with no renderer service override. The operator subsequently confirmed top-down notification gestures but reported bottom-up Home/overview failing. This is an open routing issue, not shipped gesture acceptance; group 3 remains unchecked while bounded real-contact trace collection is enabled.

Exact bottom endpoint recovered

The physical trace showed a bottom contact at y=1920, exactly the output height, rejected by the half-open route bounds. The adapter now moves only exact boundary coordinates just inside the output, preserving genuine off-output rejection and keyboard reservation. The real-policy host regression passes. After trace and activation/operator record preserve the installed trial identity and the operator’s “oh hey bottom gesture fixed” confirmation. That supersedes the preceding bottom-route failure; the complete physical gesture sequence and budgets remain unchecked.

The real compositor headless scene now passes the named dispatcher and grounds task 2.3 only. This change stays open for the remaining physical, pixel-coverage, performance and panel-regression gates.