docs: Qwen3.8-2.4T-A95B UD-Q1_0 has no reproduction recipe and no checkpoint pin in USAGE.md

[localai-bot]

Qwen3.8-2.4T-A95B UD-Q1_0 serves on one GB10 today: the borrowed-mmap tier makes it
fit (ENG-EXPERT-STREAM, #912), --offload-config's vllm_cpp key configures it
(#1119), and --device cuda is refused at load with the footprint arithmetic instead of
dying mid-stream (#1123, #1132). None of that is reachable from the docs as a recipe.

The checkpoint is not pinned anywhere a user reads. AGENTS.md requires every ported
model to document the checkpoints it was built and gated against in docs/USAGE.md —
file name, size, exact repo and revision, sha256 for a quantized artifact. Revision
567d3e6ac26c5474b18311e619c04350fb9a5556 appears only inside .agents/issue-index.md
row #1123 and .agents/specs/expert-streaming.md. docs/USAGE.md names the checkpoint
three times and pins it zero times, so nobody outside the agent record can reproduce a
single number this project publishes about it.

Nor is there an end-to-end recipe. docs/USAGE.md's "Streaming routed experts from
disk (capacity mode)" is a good reference for the mechanism — schema, precedence, where
the config form does not reach, the statistics line, the CUDA refusal — and it is only
that. It does not say which file to point --model at, what the load costs, what decode
costs, or where the ceiling is. The four sibling per-model recipes in the same file
(Muse Glimmer 30B, Nemotron-3.5-Lightning-30B, MiniMax-H3, MiniMax-Music3) all do.

Two defects in that section, found while reading it.

1. Both example invocations name -00001-of-00008.gguf. The checkpoint has ten
   shards: UD-Q1_0/Qwen3.8-2.4T-A95B-UD-Q1_0-00001-of-00010.gguf through
   -00010-of-00010.gguf, 397,256,393,248 bytes total, confirmed against the HF API at
   the pinned revision and consistent with the 1702-record census in Qwen3.8-2.4T on --device cuda: the model loads, then the first forward dies on cudaMalloc: out of memory #1123. Copying the
   published line gets a file-not-found.
2. Line 3848 writes the server as ./build/vllm-server. examples/ is added with
   add_subdirectory and nothing overrides RUNTIME_OUTPUT_DIRECTORY, so the binary is
   at build/examples/vllm-server — which is what the same file writes at lines 1401 and
   2634, and what .agents/specs/expert-streaming.md records as the working invocation.
   Line 2354 carries the same defect for Muse Glimmer.

Not a gap in the mechanism reference. The fix is a recipe section beside its four
siblings that links to the streaming section rather than restating it, so each fact keeps
one home.