2026-10-03 / cloud Linux / agent-operated functional QA
This is actual native input, not human enjoyment evidence or a Windows run.

Original first native session: media/native_20261003T024604_056645/ (79 events).
Its source, event records and PNGs are preserved. Two later corrections mean its app/model hashes differ from current sources; it is not relabelled as a final-source capture.
Observed: held Space starts only; held Return opens movement forecast only; H→R costs 2; Z opens tools; initial offset2 trial fails without resource consumption.
During the first trial, P stopped trial at frame8, Return was ignored, and after 1089 render frames paused, P resumed from the same frame8. It completed at trial frame120.
Down twice sets offset0 for free. New trial passes. Repair preview and repair charge exactly materials2/time1 and reveal R–K cost1.
Picked up the bottle, used the repaired R–K edge, saw fish crates at K, picked the bell, received one tool.
S at K recorded time16/materials3; X undid the tool (materials2); L restored time16/materials3 with supplied=True. No extra tool was claimed.
Travelled to F, toggled gear edges 1/2/3 from mask3 to mask4, rehearsed and repaired. Used new F–H cost1 and finished with time12/materials1, both repairs, all3 towns visited, bottle+bell, fish/seedlings/bell life events.
Retry, old hand-method tools → menu → comparison B → start reset correctly to H/map/gear. Left+Right chord, comparison-specific help, screenshot, exit were recorded. Window absence confirmed.

Final-source native session: media/native_20261003T025455_922758/ (16 events).
Current source hashes match this folder exactly. Started, moved H→R, adjusted offset0, ran free trial.
At trial frame120, both cam marks are3, overall trial success is displayed, and corrected current-time label reads『発車待ち』rather than falsely saying phases differ. PNG captured with F12.
S saved, Up changed phase, L restored the saved state and normalized to the map, another PNG captured. Esc+Return exited; window absence confirmed.
601-decision long-history persistence is covered by an automated regression, not claimed as 600 physical native keypresses.

Native launcher exit codes are0. Audio initialization reported unavailable; no audio is implemented or claimed audible. Every recorded model state replays exactly under current code; original records and source snapshots remain untouched.
