[ashhart]

What this changes

MiniMax-M3 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, as Gemma 4 (#3025), Nemotron (#3054) and LFM2-MoE (#3058) do.

• Same cuts: one chunk a call at exactly today's cuts. MiniMax-M3 declares no chunk invariance, so the host never caps a step and the cuts never move. The whole pass and the stepped one share the same chunk body.

• Each step first waits for other streams' steps still on the GPU. A stepping drafter (DFlash2 or DSpark) gets each chunk's rows through features(), valid until the next forward, and a mark is kept after the drafter has absorbed its chunk, as lfm2_moe metal: prefill runs a chunk a call, so a new prompt no longer stalls other streams' rounds (#3021) #3058 does for LFM2-MoE.

• Cancel between chunks keeps the state at the last chunk's end; a stream that leaves mid-pass is forgotten.

• Sizes: backend.zig 529 → 467 lines with the pass moved to a new prompt_pass.zig (181); new prefill_steps.zig (52) and its test (292). The new host step is in the macOS host CI list.

• On cost: an M5 matrix on gemma4 metal: prefill runs a chunk a call, so a new prompt no longer freezes other streams' decode (#3021) #3025 showed lone decode level with main (earlier gaps were run-to-run noise), and a new prompt's first token arriving later while other streams decode is the interleaved rounds, the point of Gemma 4 on Metal: a new prompt freezes every other stream's decode for its whole prefill (no prefill stepping) #3021.

• 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. Small adjacent-line conflicts with lfm2_moe metal: prefill runs a chunk a call, so a new prompt no longer stalls other streams' rounds (#3021) #3058 in prefill_widths_metal.zig/host_ci.zig; whichever lands second rebases.

Checks on this head

• Lean checker: 0 new; every file under 600 lines. zig fmt --check: clean.
• test-minimax-m3-metal-steps-host: 5/5 (4 named, including the drafter's absorb order); check-metal-prefill-rest-host passes (it compiles the real prefill, prefill_step and cancel_prefill bodies).
• Not run here: the Metal device steps. We'll run them on an M5 Mac before landing.

Receipt

• Main against branch: owed, on an M5 Mac: replies byte-identical to main at T=0 and seeded, including a long prompt arriving while another stream decodes, a prompt-cache resume from a mark and a cancel mid-pass; lone decode and lone TTFT within run-to-run noise.

Checklist

• The title and description match the head commit.
• 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.