[peachtaozhutou]

The bug

The shared layer's dispatch() skips an empty grid before anything else:

if (!dp->x || !dp->y || !dp->z) return;

The guard is right for direct dispatches (Metal 4 rejects zero dimensions, and BotW's ordinary loading path submits them). But an indirect dispatch is recorded by nvnCommandBufferDispatchComputeIndirect (hle_nvn.c) with only its caller and the argument-buffer address — x/y/z are zero at record time by design; the real grid lives in the 12 bytes the buffer holds. The guard therefore returned before dp->indirect was ever looked at, and every indirect compute dispatch was silently dropped — the exact class of failure hle_nvn.c's own comment warns about ("entries that would silently drop GPU work (indirect draws/dispatch …) were removed").

The regression

This is #2152: Ocarina of Time's grass is generated through indirect compute dispatches, so Hyrule Field and Lon Lon Ranch lost their vegetation on the shared layer. The reported bisect (good b0d2f353, bad f9f7aa8d) is exactly the window this guard came in with. Ocarina of Time's own graphics layer had no such guard, and its indirect dispatches read their arguments and ran. Trees, characters and everything drawn through ordinary draws were unaffected, which is why the fields broke alone. The v28 mesh-output fix (#2166) did not touch this path and did not restore the grass.

The fix

One condition: if (!dp->indirect && (!dp->x || !dp->y || !dp->z)) return;. Direct zero grids keep being skipped; an indirect dispatch whose argument buffer holds zeros still runs and does nothing, as before.

Verification

MacBook Pro (Mac16,8) M4 Pro 24 GB, macOS 27.2.1, Ocarina of Time v131072, same save and settings, update_runtime.py rebuild:

• before: Hyrule Field / Lon Lon Ranch grass gone (as Ocarina of Time 131072 (Mac): grass fields draw as flat streaks on the shared graphics layer. Good at b0d2f353, bad at f9f7aa8d. #2152's screenshots, streaks on the reporter's M3 Pro)
• after this one line: the grass is back, everything else unchanged, no new log lines

Separately, this checkout also carries a local workaround for the pinned .sha256 format change of PR #2166 (#2184) — it is not part of this PR, which is the single nvn_mtl.m hunk only.