measure(ENG-EXPERT-STREAM-DEVICE): W0e ran, the CPU arm has a live-cache number and the CUDA arm loads without generating (#1124, #1299)

[localai-bot]

W0a and W0e both ran on dgx:gpu0 inside one rc hold, on the head of #1242.
Two of the three things this row wanted arrived and the third did not, so this
change records all three and claims only what was measured.

W0 works where it was aimed. --device cuda on Qwen3.8-2.4T-A95B UD-Q1_0
loads in 255-272 s instead of refusing, the [expert-stream] ON banner prints
on the device arm, and that banner is the first production evidence that W0b's
CUDA leg is reached and answers true on real hardware rather than on the fake
platform its unit gate registers.

The CUDA arm then exhausts the box inside its first forward and emits no token,
seven attempts, every one the same. That is filed as #1299 and it is not the
expert lane: a 64-slot 0.15 GiB arena dies exactly where an 8000-slot 18.55 GiB
one does, and a 1-token prompt, whose protected set fits the arena with no
in-place tower fallback at all, behaves exactly like a 5-token one. The growth
is RssAnon while RssFile stays flat, so the GPU is not pinning the mapping's
pages either, which was the first hypothesis and is refuted. Host anon plus swap
reaches ~65 GB while the system reports ~119 GiB used, and the ~42 GiB
difference is device memory this unified part does not charge to process RSS.
The non-expert weights are resident twice, and about 50 GiB of that is the bf16
expansion the GDN V-head reorder already forced on attn_qkv and ssm_out.
expert-streaming.md predicted this in prose: "Whoever takes this needs BOTH".
W0 delivered the expert half.

The CPU arm is measured, reproduced, and replaces the figure BENCHMARKS.md
carried as VOID from #912 F1: 11.05 s/token steady over 29 samples, two reps,
with steps=32 and a decode-phase exhausted delta of 0 in the same run, which
is the pairing #912 F1 lacked. Four runs across two slot counts produced
byte-identical token ids. The 8000-slot pair came out 3.6x SLOWER than the
4000-slot pair, which is the wrong direction for a bigger cache, and the swap
column says why: an 18.55 GiB arena does not fit beside 62 GiB of dense weights
on this box, so those runs swapped and one consumed all 30 GiB of swap. That is
a memory result wearing a cache result's clothes, and it is recorded as such.

G0-CORRECT has no CUDA side to compare against and says so rather than reporting
the CPU half as a pass. G0-LIVE passes on CPU and is not reached on CUDA.
G0-SPEED reports the CPU number and no ratio.

The harness lands under benchmarks/ beside marlin_moe_standalone.cpp,
unwired from CMake, because one generation had to yield three things no shipped
command gives together: the generated token ids, a per-step arrival timestamp,
and the stream counters at both ends. Its logits processor is a pure observer
that edits nothing, so the argmax is the one the sampler would have taken
anyway, and the prompt is fixed as token ids so both arms are fed identical
bytes with no tokenizer between them.

Stacked on row/ENG-EXPERT-STREAM-DEVICE-W0 (PR #1242), because every claim
here is a claim about that branch's code and none of it should land ahead of it.
agent-ready.py is green on all 28 record gates and every mutation suite; its
two commit-gate SKIPs are the stacking itself (origin/main is not an ancestor
of a branch based on #1242's head), and both gates were run explicitly over
95883dcae..HEAD and reported OK: commit trailer contract and OK: commit writing style.

Issues: #1124 (the row), #1299 (filed by this measurement).

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5-1m [claude-code]

[localai-bot]

Replaced by #1414.

#1377 landed as a squash (5f4eb356e), which destroyed this branch's merge base. Merging main into row/ENG-EXPERT-STREAM-DEVICE-W0E produces thirteen conflicts, including src/vllm/entrypoints/model_loader.cpp, src/vllm/platforms/cuda.cpp and three test files that this measurement does not touch — git sees the branch's pre-squash W0 commits and main's squashed W0 as unrelated content.

#1414 is the same eight-file delta reapplied onto main, which leaves three record conflicts instead of thirteen, each resolved by key: the benchmark log unions two disjoint sections, and the two keyed table rows take this row's newer state while keeping the keep-quant/keep-f16 mechanism #1378 corrected. Every one of main's #1378 repairs was verified to survive in the spec afterwards.

This branch is left in place rather than deleted, since it holds the original review history.