MVC1 Cross Fever V3 experimental preview: bounded public evidence

Start with V3_SIMULATED_INPUT_RESULTS.json and REPLAY_STATIC_REVIEW.json.
Explicit path references are relative to the archive root. Included references
bind sanitized public files and retain the original file SHA-256 separately.
Runtime hashes inside receipts remain the original tested-source hashes.
PUBLIC_MANIFEST.json maps every included original file to its public copy.

Inputs are ordinary simulated libretro controller callbacks. Physical controllers,
a four-machine internet match/rematch, and full roster/form acceptance are not proven.
Mac lifecycle replays compare game work RAM only. The retained Linux/Wine replay
limitation includes one differing post-load video frame and QSound state differences;
the exact V2.2 Linux control reproduces those counts in the measured window.

ROM members, keys, core binaries, RAM/serialized-state dumps and verbose raw
trace JSONL are excluded. Their original hashes remain in the evidence.
Raw traces are retained locally; this is a bounded receipt/evaluation package.
The frozen matrix evaluator matches the original report hash. The separate
source package adds an import-path-only portability fix after the frozen tests.
