docs: BUILD.md and serving.md name the nemotron seat, and the README's limit line points to serving.md

[midagedev]

Follow-up to #278 (its Follow-up item). docs/BUILD.md and docs/serving.md on main 75dc51a6 do not name the nemotron seat that v0.3.3 serves.

What

• docs/BUILD.md, the bloomery-serve row: --model lists nemotron, and the Models cell names Nemotron 3 Nano 4B. The row's feature set already carries the seat (it sits behind deepseek41, which glm5next enables).
• docs/serving.md, "The seats": one paragraph on the nemotron seat, held to serve_seats/nemotron.rs's module doc and ACTS_ON: one card, one slot; --parallel above 1, a multi-card or tier-card placement and a draft lever refused by name; one decode step per prompt token; the step-back rule; no prompt cache, a snapshot or resume refused by name.
• docs/serving.md, "Concurrent requests" said every generative seat takes --parallel N, which the nemotron seat refuses; it now names the exception. The Limits line on one-request seats names it too.
• README's Known issues line on Nemotron is cut to one sentence and a link to serving.md's seats section, so the seat's rules are written once in the docs (beside the seat's module doc). The README's Use table stays --hf rows only, as docs: README and models.md list Nemotron 3 Nano 4B, served by -m until its Hugging Face repo is named #278 left it.

공통화한 것 (Commonized)

• The nemotron seat's rules have one place in the docs, serving.md's seats section; README points to it.

Evidence

• crates/gpu-gates/src/bin/shared/serve_seats/nemotron.rs at 75dc51a6: lines 22–33 (one sequence, --parallel above 1 refused, no prompt batch, the step-back rule, no prompt cache), 36–40 (placement), 96 (ACTS_ON), 204–209 (snapshot and resume refused).
• crates/gpu-gates/Cargo.toml:117: glm5next = ["deepseek41", …], so BUILD.md's glm5next,clef,vision,mimo2 build carries the seat (the serve gate builds bloomery-serve with the same set).
• just contrib-check --tier t0 on 41ab6b4e: every step rc 0 but check-recipes, whose one red was tools/review/local.py --self-test's "SIGTERM to local.py ends it and the role's process group: ended True, child alive True". Run alone three times on the same tree: rc 0, 0 FAIL each. An earlier run on 55cbc8b8, beside a pre-review, had its one red in gpu-gate.sh --self-test (a 4 s hold bound waited 282 s); alone it gave 105 of 105 ok. This diff touches neither tool: README.md, docs/BUILD.md and docs/serving.md only.

Follow-up (not in this PR)

• Two check-recipes self-tests went red once under load and green alone: local.py's SIGTERM process-group case and gpu-gate.sh's fresh-hold wait. Both read a timing the load moves; a bound or a wait on the child's exit, not a sleep, is a tool change of its own.
• serving.md's "Tool calls on every generative model" lists five models and not Nemotron. Which ToolFormat (crates/serve/src/dsml.rs:294) the Nemotron file's template resolves to is not checked here; a serve request with a tool would settle it.

Pre-review CLEAN at 41ab6b4ee8635be5cb6d30c6df103d0b4244067a; 2389edf is the same patch on main d213b59 (patch-id 36f92ec4).

🤖 Generated with Claude Code

[github-actions]

Verdict: CLEAN: The documented Nemotron build, serving limits, and prefix behavior match the implementation.

Checked

Duplicates: 0 new functions, types, or constants; the README links the detailed seat description.
Complexity: no functions changed; no crapper candidates.
Silent failure: 0 changed executable sites.
Tests: no tests changed, so mutation checks do not apply. Documentation link target and build feature declarations checked statically.
Behavior: checked CLI dispatch, single-slot and placement guards, draft refusal, prefix reset, and snapshot/resume refusal against the implementation.
Gates: none changed; no pins moved.
Escape catalog: no C1–C4 patterns matched.

Notes

Dryer and crapper did not run because no Rust files changed.

_{🔨 Shingler, bloomery's review bot: approves a clean pull request and requests changes on a must-fix (how it works)}