Direct drawer and shade tracking: headless QEMU

Run at 2026-09-24 06:22 UTC. This is synthetic card_shell test-touch input in a 568×1232 headless Pixman Sway compositor under QEMU. The card is a public probe fixture and the drawer catalog contains one temporary public desktop entry. The adjacent PNGs are original compositor captures; result.json records the measured panel edges. No physical finger, LCD, or latency/CPU measurement is represented.

The combined compositor source is 7eff7613733c05aa4bf9f4fe23911fb59e25d03d (including the direct-travel denominator correction c5cf6146 and diagonal projection correction 48f9a8fa). The card package was /nix/store/j4cq3c28dnk6jybi66k32a1ydzx80ffr-k230-card-shell and the exact Sway executable was /nix/store/6jx6nlxxc2hy5viidq7msnlpshidd5w6-sway-unwrapped-riscv64-unknown-linux-gnu-1.12/bin/sway (SHA256 cb94679b97cdab2c1b61743f163f74d97db8e20e459b1c20f1eb94aa53bbad36). The Rust receiver was /nix/store/rbvcqv3sjq87iami6iymwd9cx7nzhhma-k230-shell-rust-riscv64-unknown-linux-gnu-0.1.0/bin/k230-shell-rust (SHA256 7b0b104cffdc9523f999b2a7bd9762e98bd5ff89a2f2a7e7c9861c7f5221275f). The synthetic live client was /nix/store/6nskh3ldk3ya7mwxbb258qszrxhw2i61-card-composition-probe-client-0.1/bin/card-composition-probe-client.

Reproduce from this repository with:

python3 tests/test_direct_reveal_drag_runtime.py \
  --sway /nix/store/6jx6nlxxc2hy5viidq7msnlpshidd5w6-sway-unwrapped-riscv64-unknown-linux-gnu-1.12/bin/sway \
  --rust /nix/store/rbvcqv3sjq87iami6iymwd9cx7nzhhma-k230-shell-rust-riscv64-unknown-linux-gnu-0.1.0/bin/k230-shell-rust \
  --client /nix/store/6nskh3ldk3ya7mwxbb258qszrxhw2i61-card-composition-probe-client-0.1/bin/card-composition-probe-client \
  --output /tmp/k230-direct-drag-qemu-positive-01

The original control run with the prior Sway executable /nix/store/qwja5838kgidxzvii70l0smf72snzsqs-sway-unwrapped-riscv64-unknown-linux-gnu-1.12/bin/sway failed as intended: a 100 px drawer finger move shifted the captured panel edge 135 px (observed y=1097; exact 1:1 expected y=1132). With the combined candidate, the drawer edge after 100 px was 1132, held at 1132 during a 250 ms pause and a 100 px horizontal perturbation at unchanged y, returned to 1172 after a 40 px reverse, and settled at 234 after a separate full swipe. The corresponding shade edges were 100, 100, 60, and 801. Both below-threshold reversed gestures unmapped and restored the deck. The pixel assertions allow 2 px for movement/reversal and 1 px for no-drift hold.

The test exercises composed intermediate pixels and the existing synthetic input route. It does not establish glass finger tracking, panel presentation timing, real touch controller sampling, or the strict device frame/CPU budget. Those remain separate physical gates.