Caesar III personal review policy and startup input readiness
Evidence: scratch/runs/20261117-caesar3-reviewed-fifteen-second-city-consequence/result.json.
The broker and Caesar controller now accept gameplay captures younger than 15 seconds. Menu review remains 30 seconds. Original capture clocks, retained complete PNG bytes, personal review timestamps, focus/actor/window generation/context/target checks and one nonce per finite intent remain binding. Confirmation, independent physical UP and aggregate observation remain 500/1750/3000 ms. Stale gameplay images close the controller before intent/DOWN; UNKNOWN endpoints remain permanently closed.
Real native tests delivered a 10-second-old gameplay capture and refused a capture older than 15 seconds with zero DOWN and no later HTTP. Native delayed confirmation, delayed independent release, late aggregate observation and late immutable encoding all closed correctly. Personally viewed complete published fixture images produced an intent at a real capture age of 8.432 seconds and native DOWN at 8.509 seconds, followed by independent release and native fixture Quit. These are authored fixture results, not Caesar city progress.
One original ran on the exact published e27e6c155 graph and module 5a4178950559d4f677d86aa75403908abee343493805f01a0b8aae8f1df2429a. Its 114 recursive media hashes were recorded before the fixed 2400-second native / 2420-second independent watchdog clock, with a 240-second Quit reserve. Fresh sealed main/CPUWorker/OS-child inheritance matched the completed native and own-parent-supervisor proof. A separate preparation on an unrelated Neverhood publication refused its mismatched graph before starting any original. The retained small launcher selected e27 explicitly through the same published supervisor API; it did not create another original allowance.
The actual title artwork still had a blank loading rectangle. Its first finite title click reached the physical handler but no native USER input return arrived by 500 ms. The retained public API trace shows ongoing WAV loading and no GetMessage/PeekMessage before cancellation. Independent physical UP acknowledged at 1437.5 ms. The endpoint closed permanently. Native, parent and watchdog waits completed with EOFs and an empty process scope. Host cleanup is distinct from native Quit. No housing, road, population or city-entry consequence was reached.
The resulting repair requires a public input-readiness preflight before an original capture. It observes the first completed unfiltered GetMessage/PeekMessage poll of the focused owner thread and HWND. A bounded first-poll record survives the diagnostic trace limit. This is readiness evidence, never a delivery receipt, inferred image meaning, guest API invocation or an automatic input. control ready reports unavailable without capturing, binding or pressing. Capturing also checks readiness; personal review and strict native confirmation are still required afterward.
An authored native fixture delays its first pump for four real seconds. The controller refuses the early capture with zero PNG, binding and DOWN, then allows a fresh reviewed finite click after the ordinary owner poll returns. Wrong actor, target, filter, stack, yield and timestamp records refuse readiness. The supervisor also accepts an explicitly selected published ancestor so unrelated HEAD movement does not silently choose another source graph. Reuse still compares the complete source map, module, seals, native waits, EOFs and completed parent scope.
The single original allowance is spent. A future separately authorized run should wait for public owner input readiness, personally inspect a new complete image, and continue through normal UI to housing/road/population consequence. The closed endpoint must not be queried, rebound or replayed.