[nanotchec]

Change

Keep an indirect mesh conversion when its eight input bytes are disjoint from every known storage write made by the preceding mesh draw. Overlapping records are discarded. Unknown writes, unsized buffers and sparse ranges retain the existing invalidation path. Also record the Metal indirect mesh selector on the optional encoding thread.

This is a stacked draft based on seedds#1312. The base provides the indirect mesh path.

Measurements

Ocarina of Time 1.1.1, M3 Max, docked, game NDSR, fixed view at the top of Link's ladder in Kokiri Forest. Both versions used the same wide-write optimization and game resources. The direct-encoding A/B produced 24.8 → 27.4 FPS, then 24.4 → 26.4 FPS in the reverse order. In four profiled images, indirect conversion splits dropped from 24 to 8, and the main colour/depth passes averaged 28.4 → 25.3 ms of GPU work. Frame captures and the local test covered this route. The reported foliage animation issue remains open.

Checks

• t_mesh_indirect passed with Metal validation in both encoding modes. It tests GPU-produced arguments, overlap with the second source word, preservation of a disjoint record, and the unknown-size fallback.
• Docs tests and the scoped LLVM 23.1.4 ratchet passed.
• The 1.1.1 gameplay check ran on the local Ocarina integration. Additional games and a full playthrough remain unchecked.