[ashhart]

A request's finished stats now say how its prompt was cut on Metal: tensorfold.prefill_widths lists one entry per completed target prompt pass, with its real row count, for Qwen3.8-27B, Nemotron, GLM-5.3-Flash and every other Metal text family: Gemma 4, LFM2-MoE, Step5, Spark2.5, GLM-DSA, Kimi K3, Llama, MiniMax-M3, Qwen 3.5 and Qwen4. It answers "how was my prompt chunked?" without a profile build.

• What it counts: only completed target prompt passes. A drafted verify round isn't a prompt pass, and neither is the drafter absorbing the prompt. A pass that fails is dropped. If an allocation fails, the stats come back without widths and the reply still succeeds.
• Where it's counted:
  • The 27B appends in the shared Metal fill path.
  • GLM appends after its synchronous prompt chunk.
  • Nemotron appends when each pipelined prompt command buffer lands (moved into prompt_flight.zig, with the same chunk cuts, wait, error and timing path).
  • The other ten families each get one added line at the point their existing prompt loop finishes a target pass. No loop is moved, and no original line changes.
• Telemetry only: no token, default or kernel changes. The only additions to the hot path are a list append, and a field on the lane stream that is freed with it.

Checks (the third commit adds check-metal-prefill-rest-host and test-qwen4-prefill-widths-host: 31 of 31 steps and 16 of 16 tests, with the earlier steps):

• zig build test-qwen27-prefill-stats-host test-qwen27-prefill-widths-http-host test-glm-prefill-widths-host test-glm-scheduling-host test-nemotron-prefill-widths-host check-nemotron-prefill-widths-host: 19 of 19 steps, 78 of 78 tests. Chunked prompts report their real cuts, a one-pass prompt reports one entry, a resumed prompt reports only what it ran, and a failed chunk reports nothing.
• zig fmt --check on every touched file; lean unchanged at 426.
• A plain zig build passes: 255 of 255 steps.

[ashhart]

The LFM2 rerun on an Apple M5 Ultra is level (LFM2.5-24B-A2B 4-bit, main against this branch, warm base new new base base new, three runs an arm):

• Prefill is within 0.4% at 16K, 32K and 64K, at every start offset.
• Decode medians are within 1.6% at every cell but one.
• The exception is -3.3% at 16K from offset 5,000. There the branch's three runs are 277.9, 264.3 and 263.9 tok/s against main's 273.3, 268.2 and 274.6. The runs overlap, with the branch's fastest run faster than any of main's, so it isn't a steady change.

With the first LFM2 run's REVIEW now settled, this lands from the queue, and #2304 follows once its CUDA cells are in.