[ashhart]

Stacked on #2293, whose three commits show here until it lands. This PR's own change is the last commit.

On CUDA, a request's finished stats now list tensorfold.prefill_widths too, for the three CUDA families we run: Qwen3.8-27B (and Qwen 3.5), GLM-5.3-Flash and Nemotron. Each entry is one completed target prompt pass with the rows it ran. An entry is added only where the code already waits for that pass to finish; no wait, sync, loop or cut is added or moved.

family	where it appends	what it records	what it leaves out
Qwen 3.5 / 27B CUDA	after the chunk's existing .last forward, whose readback synchronizes	every target chunk's real rows, including request cuts and resumed tails	—
GLM stepped	after the final draw, or after an existing intermediate settle_step wait	the rows each step took	intermediate steps with no settle wait
GLM batched	after prefillMany returns, since it draws each stream's token	each request's prompt length, once	—
Nemotron served	after prefillFrom's existing first-token download, only when exactly one target chunk ran	that chunk's rows	multi-chunk calls, which report nothing rather than guess

glm/cuda_lanes.zig and nemotron/cuda_lanes.zig were already at 600 lines, so each drops seven separator blank lines to stay at the limit. Each also exposes an existing step to the host fixture. Telemetry only: no token, default, kernel or launch change.

Checks:

• Host: 339 of 339 steps, 5,013 tests passed and 28 skipped, across test-cuda-host test-runtime-host check-native-cuda-host test-lane-api-host and the widths steps.
  • Chunked, resumed and one-pass prompts report their real cuts in all three families.
  • With the five appends removed, five of the ten width tests fail.
• zig build -Dtarget=aarch64-linux-gnu compile-cuda-tests: 10 of 10 steps.
• zig fmt --check; lean is unchanged.

[ashhart]

CUDA cells for this change, main against this branch:

• Qwen3.8-27B on a GB10: GREEN.

• Nemotron on a GB10: GREEN.

• GLM-5.3-Flash NVFP4 at TP=2 on two GB10s: tokens equal in all 52 checked cells (0 unequal, 0 unverified). Speed, from two runs an arm:

  • prefill is 6.7% faster at 2K and 4.5% faster at 8K and 32K;
  • decode is +1.2% (chat) and -1.6% (code) at one stream, and +0.3% (chat) and +5.4% (code) at two.

  The prefill gain would need more runs before we'd claim it. The -1.6% is inside the spread we usually see with two runs an arm.

The Metal side was checked with #2293. This lands right after #2293, which it's stacked on.