006 native GUI check — 2026-10-02 17:01–17:04 UTC
Agent-operated physical-key input through cloud desktop CUA. Not human playtest.
Native Pyxel 2.9.9, Linux. Windows native execution not performed.

Observed passes:
- Title Japanese text legible; A/D selects authored service, Space starts.
- A: Space picks material once; D moves to oven; Space starts cooking.
- First readiness lesson visibly freezes time until input; no instruction-time burn.
- Space retrieves; two separate D presses reach delivery; Space shows warm1/4.
- P overlays help and freezes time. Z restores undelivered plate and prior order state.
- B: S chooses the second ingredient (toast), then Space picks it; stove cooks it.
- Took other raw plate while oven occupied. Used shelf to free hand, retrieved
  cooked toast, exchanged it with raw thick dish. Shelf cooked/raw hand visibly
  matched expected ownership. Screenshot media/manual_runtime.png is from F12.
- Raw thick dish at delivery is rejected with "まだ生です。焼き台へ戻せます".
- Toast cooled on shelf; swapping back and delivering increased delivery to1/4
  while warm count stayed0. No failure or hidden deadline.
- Thick dish returned to oven resumed heating; left on stove to burn. Black plate,
  "焦げ！" and actionable message visible. Space then X held1s removed it;
  waste became1 and same order could be made again.
- R from pause instantly reset B to0/4, warm0, waste0. P help worked again.
- Escape returned to title; Q closed window. Second duplicate launch also closed.

Not directly exercised in native input: full four-order ending, N next-result,
V reverse and every randomized abuse sequence. These are covered by model tests
or actual-runtime scripted demo where applicable, not claimed native tested.
No sound essential. Runtime reports failed audio device init; audibility unverified.

Design result remains limited: shelf enables preparation overlap; all tested
multi-order reversals favor the same fast pipeline. A new order-optimum reversal
or meaningful warmth-vs-speed tradeoff has not been proven. Human enjoyment,
voluntary replay and perceived busywork remain unvalidated.
