[ashhart]

What this changes

LFM2-MoE on Metal ran a new stream's whole prompt in one call on the lane thread, so with shared rounds every other stream's reply waited for the whole prefill. The backend now declares prefill_step and cancel_prefill beside reports_fill, so the lane host runs one chunk a call and decode rounds run between chunks. This is the third family after Gemma 4 (#3025) and Nemotron (#3054).

• Same cuts. prefill_steps.zig's chunkEnd is main's loop cut moved out line for line: at most options.chunk rows, cut at the first planned chunk start past at, then at the next kept mark. LFM2-MoE declares no chunk invariance, so the host never caps a step and the cuts never move. The whole pass (prefill) and the stepped one share the same chunk body.

• Drafter taps. A drafting stream's sink gets each chunk's tapped rows, served from the latest chunk while no forward has run since; the last row's taps are kept on the last chunk as before.

• Cancel between chunks keeps the state at the last chunk's end; no prompt work runs.

• A stream that leaves mid-pass has its pass forgotten in release.

• Issue: Refs Gemma 4 on Metal: a new prompt freezes every other stream's decode for its whole prefill (no prefill stepping) #3021.

• Stacked on: main.

• atc overlap --json: not run.

Checks on this head

• Head commit: 6496bef7f16e44ea3b3665c3238d53891004f272
• Lean checker on the base and on this head: 345 and 345, 0 new.
• Tests run:
  • zig fmt --check on the six files: passes
  • zig build test-lfm2-moe-metal-steps-host check-metal-prefill-rest-host: 10/10 steps succeeded, 5/5 tests passed. The tests check that the stepped calls' chunks equal main's whole loop's, that a decode runs between chunks, that each chunk's taps reach the sink, and that a pass resumed from a kept state cuts as main does.
• CI: runs on this push.
• Not run, and why: zig build test-lfm2 and every Metal device step. We'll run them on an M5 Mac before landing.

Receipt

• Main against branch: owed. On an M5, one long prompt beside a decoding stream, greedy, with and without a drafter: both streams' replies must equal main's. The decoding stream's inter-token latency while a 32k-token prompt fills should stop stalling for the whole fill.
• Paths and platforms: Metal only; CUDA's LFM2-MoE path is untouched.
• Exactness: the cuts are main's (host test above); device token equality owed with the run.
• Licence of any new weights or drafter: n/a.

Checklist

• The title and description match the head commit.
• New tests pass after the change. They use the new step driver, so they can't compile on main.
• Comments and docstrings are one line.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.