1@README.md
2
3- `run_frame` loops `tick()` until the core reports `FrameRendered`. One call
4  is one video frame, never one CPU step; callers count frames with it.
5- Player-one input is `inputs.p1.set_field(button, pressed)`. The method is
6  generated by upstream's `define_controller_inputs!` macro, so it does not
7  appear in a grep for `fn set_`.
8- `wram()` and `wram_mut()` depend on two patches this repo carries on the
9  vendored core (`third-party/rust/jgenesis`, commits "jgenesis: expose work
10  RAM from snes-core" and "jgenesis: let a host write work RAM"). Upstream
11  keeps `Memory` private. Re-apply them after any subtree pull.
12- Index work RAM by offset from `$7E0000`. The RAM map in `../../research/`
13  writes addresses as `$7Exxxx`; subtract `0x7E0000`.
14- The core is deterministic: same ROM, same inputs, same frame count gives a
15  byte-identical picture, optimised or not (checked 2026-09-20 at 1500
16  frames). A test that compares frames is therefore sound.
17- **`snapshot` and `snapshot_fixed` are different formats on purpose.** Every
18  state on disk (`resume`, `home`, the MCP tools' named states) is
19  `snapshot`; only `packages/replay`'s keyframes are `snapshot_fixed`.
20  Switching either one breaks every file already written in it.