evidence · 4.6 KiB

README.md

Committed at docs/evidence/final-shell-image/README.md. This is what grounds the requirements that cite it.

Open raw source

Final installed video and launcher regression

The coordinator built and flashed source f76754c on 2026-09-23. Later opt-in experiment packages do not change this normal system: evaluation after their integration still resolved the same system path.

nix build --no-link --print-out-paths .#toplevel .#sdImage --max-jobs 1 --cores 8 succeeded. tools/ums-session.py --flash IMAGE --expected-sectors 249872384 --out PRIVATE_LOG held the board lock throughout the flash. Its write returned zero: 2,308,960,256 bytes in 181.3 seconds including sync. The Linux login prompt appeared and the console reported the system path above. Full-image readback was skipped by the established policy. Only the protected Wi-Fi file was restored, with matching bytes and root:root mode 0600; no home directory was restored. Raw flash/network transcripts remain private.

Executed checks

Result records an injected-touch physical-board regression. This is not a real-finger test, optical timing or performance benchmark. The Python native-frame assertions deliberately impose load; their elapsed time must not be read as UI latency. The installed status identifies the actual installed wrappers, services and touch device.

The check covers Apps Next/Previous, Help/Back, Terminal and Monitor focus, keyboard show/hide with the expected 400-pixel reservation, overview up/down and Back, System reboot-confirmation cancellation without reboot, and Apps recovery after forced allocation failure. For the latter, prlimit --as=1 constrained only the identified launcher process; Next made it exit. A new Apps invocation rendered, its Next/Previous changed and restored the canvas, and Back closed it. Recovery frame, shown keyboard, and Help are native grim captures of this installed image. They are not camera images.

Each of three Video trials was launched through Apps. The checker required the actual software video surface and media time beyond one second, then injected Windows followed by Stop, Back or Home. It waited for the state file and tracked controller/player processes to disappear, checked that no session socket or video surface remained, and checked Home focused Terminal. A final Apps/Back interaction and active shell/seatd/Wi-Fi services passed. The public BBB 480x270 software profile and disabled audio were the image defaults. Earlier committed comparison, error, EOF and fallback evidence remains linked from docs/evidence/network-video/; this run adds final-image integration and cleanup, not new sustained-presentation or audio claims.

Procedure and rejected checker runs

The source is check.py. Transfer it to /run, copy docs/evidence/launcher-gestures/metadata-budget/check.py as /run/rollback_check.py, and provide the repository’s /run/inject-tap.sh plus a reserved evemu touchscreen at /dev/input/event1. The image’s Python is /nix/store/v189xydz6qkcd4cbkixcmv91w8hbc560-python3-riscv64-unknown-linux-gnu-3.14.7/bin/python3. Run the checker from the normal home bar with its keyboard hidden. The script sets the normal Wayland/Sway environment and records no window titles.

The first run used a fixed 400ms pause between bar actions and missed Apps after System cancel. It was rejected; its result remains. The next run waited for the rendered bar state and passed all shell controls and allocation recovery. It then rejected an immediate process-existence assertion after Stop: the controller had removed its state before fully exiting. That result is also retained. The corrected checker waits up to 12 seconds for actual process exit. It resumed the three video cases with --video-only on the same verified boot ID, retaining the prior successful control checks. The two logs and video continuation log preserve which checks ran in each phase. No failed assertion was changed into a success without rerunning its corresponding interaction.

The forced allocation test complements the separately committed gesture-disabled rollback. These checks close the integrated-image and failure-recovery gates. The existing real-finger gesture clip remains insufficiently sharp for its separate final camera task; that proposal stays open.