Edge control intent and Home return: host proof

An edge down now holds a weak target until motion distinguishes a gesture from a tap. A tap sends balanced input to the original live target; an actual edge drag replays the original origin/time into existing native motion policy. Extra contacts cancel a forwarded client stream rather than turn cancellation into a tap. Output/seat loss and keyboard chord takeover clear pending ownership. Real wvkbd keys keep their full input surface. Home paints a themed bottom handle: a stationary tap returns to the retained cards, while an upward drag still opens the app drawer.

Exact matching build identities and source hashes accompany touch-first protocol results, legacy routing results and mouse/Home regression output.

flock /tmp/k230-nix-build.lock nix build .#nixosConfigurations.k230-coherent-shell-hdmi-trial.config.system.build.toplevel --no-link --print-out-paths --max-jobs 2 --cores 8
python3 tests/test_card_touch_routing.py --edge-intent --sway <build.json sway> --output <fresh-private-directory>
python3 tests/test_card_touch_routing.py --sway <build.json sway> --output <fresh-private-directory>
CARD_SHELL_SWAY=<build.json sway> CARD_SHELL_CLIENT=<native-probe-client> python3 -m unittest discover -s tests -p test_card_shell_pointer_navigation.py -v

The native probe client was compiled from nix/card-composition-probe-client/card-composition-probe-client.c with the pinned fixture’s xdg-shell generated protocol and host Wayland headers, cc -std=c11 -O2 -Wall -Wextra -Werror. Tests ran the actual cross-built Sway under qemu-riscv64-static. Eight touch-first cases, seven legacy cases and the mouse navigation regression passed. The overlay/keyboard test clients implement their actual layer namespaces but are protocol fixtures; these results do not establish actual Rust search Backspace semantics or physical feel. After keyboard removal the harness waits 200 ms for the test app to commit its restored buffer; the compositor correctly refuses an unpainted resized source rather than displaying it as a card.

An initial build used a nonexistent Sway geo member and was corrected to use actual scene-node coordinates and current surface dimensions before the passing build. The first host mouse invocation lacked a built native probe client; the recorded invocation uses the freshly compiled source fixture. No failed result was relabeled a pass.

Deployment and board-injected header/search-key checks are task 4.2; real-glass controls and navigation remain task 4.3. No kernel or device tree changes are part of this fix.

Home fixture reconciliation, 30 September 2026

The full Rust suite on unchanged master 7515378f failed the Home portrait pixel-identity test: the earlier Home-handle implementation intentionally added a 96×6 bottom pill but left its committed render fixture stale. Rendering unchanged master source through render_responsive_evidence and comparing RGBA pixels shows the sole difference is (236,1218)..(332,1224), the handle’s intended rectangle. The portrait Home fixture is refreshed to that visually inspected render; no renderer source changes in this correction. This is host pixel evidence, distinct from the operator’s new real-glass report that tapping Home’s bottom handle opened overview correctly. Compact search Backspace still closes the drawer according to the same operator: that negative physical result remains open in the app-drawer standard-keyboard follow-up.