Box 10.1.0.201, GPU0, registry.arbi.work/arbi-serve:test-e54f6a4, single-GPU
(ARBI_TP_SIZE=1, no --prefill-pipeline), --prefill-capture=full, GDN prefill
capture opt-in default-on, --chunk-prefill=2048, --max-batch=4,
--gpu-memory-utilization=1.0, --max-context=auto, canonical
Qwen3.8-27B-exl3-4.0bpw + DFlash2 K=7 + tkv k4v4 v42 config.

Measured through the admin console, GET /measure/ab, axis=flags, target=local,
a dummy two-arm run (arm A: no overrides; arm B: mtp_accept_dump set to a
scratch path — a flag not expected to move TTFT, used only to satisfy the
axis=flags contract that both arms carry a flag_overrides field) so the
canonical measurement path (canonical prompt, TPOT-by-tokens, bisected
prefill target) is exercised rather than a hand-rolled request. prefill=131072
tokens, concurrency=1, reps=2, cache=auto (bypassed — no flag under test
needs it), schedule=baseline_once.

Raw rung_done events (ttft_ms), in run order:

  rep 1, arm A: ttft_ms = 29161.377570126206
  rep 2, arm A: ttft_ms = 29306.02747411467
  rep 1, arm B: ttft_ms = 29344.41608004272
  rep 2, arm B: ttft_ms = 29351.383399916813

For comparison, the same box's single-GPU EAGER (--prefill-capture=eager,
the default for this model pre-existing this task) baseline from the sibling
receipt (pp-prefill-dispatch-overhead-2026-09-15.md): 29.35s-29.50s, 6 reps.

The four capture-mode numbers (29.16-29.35s) sit inside the eager baseline's
own spread. No measurable TTFT delta from enabling whole-forward prefill
capture on this workload.
