docs: README and models.md list Nemotron 3 Nano 4B, served by -m until its Hugging Face repo is named

[midagedev]

Follow-up to #268 and #259. #268 merged without the Nemotron lines, so README and docs/models.md on main 3898696d do not name Nemotron 3 Nano 4B, which v0.3.3 (5626a096) serves.

What

• README Status (0.3.3) names Nemotron 3 Nano 4B on its own seat (gpu, gpu-gates, model, refset: Nemotron 3 Nano 4B on the card and bloomery-serve --model nemotron #259).
• The seat table counts seven seats. The Use table stays the --hf rows that tools/hf-arch-check.py checks against Hugging Face, so the nemotron seat is a sentence below it with its -m PATH/NVIDIA-Nemotron3-Nano-4B-Q4_K_M.gguf command: the Hugging Face repo is not named for the docs yet, so no --hf line is written. A later docs change moves it into the table as an --hf row.
• Known issues (0.3.3) gain one line, held to serve_seats/nemotron.rs at 5626a096 (its module doc and ACTS_ON): a prompt runs one decode step per token; one slot on one card; --parallel above 1, a multi-card or tier-card placement and a draft lever are refused by name; a kept prefix only when the prompt extends the whole held sequence; no prompt cache.
• docs/models.md gains a Nemotron row before lev.

The same lines are in the v0.3.3 release body.

공통화한 것 (Commonized)

• Nothing: a docs change. The known-issue line states the seat's own module doc rather than a second wording of it.

Evidence

• git show 5626a096:crates/gpu-gates/src/bin/shared/serve_seats/nemotron.rs: lines 7 and 90 (usage with -m PATH | --hf <repo>[:<quant>]), 25 (--parallel above 1 refused), 26 (no prompt batch), 28–33 (the step-back rule, no prompt cache), 96 (ACTS_ON = card budget and screen, so BLOOMERY_DRAFT is refused at main).

• The tested file is tools/ref/models/nemotron_h.sh:24's NVIDIA-Nemotron3-Nano-4B-Q4_K_M.gguf.

• just contrib-check --tier t0 (on 09bbb77's fold, before the rebase onto 2b43cebf): rc 0, no step red; the CI red on 09bbb77 was hf-arch-check.py's "the README's own Use table parses into rows of known seats" on the -m row. python3 tools/hf-arch-check.py --self-test: rc 0.

Follow-up (not in this PR)

• docs/BUILD.md:72 lists --model ds41|qwen38|glm|qwen3|mimo2|decide, and docs/serving.md's "The seats" (line 82) has no Nemotron text. Both gain the seat in a docs change of their own.
• bloomery-serve's own help already prints --hf nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF:Q4_K_M (bloomery_serve.rs:145). That repo is public and not gated, and its one GGUF is byte-identical to the tested file: 2837072864 B, sha256 be5d9a656a51922f24f1f09a759cebb694e1f5d9728bf0ef9f8c972c5a0b5ef2 on both (Hugging Face tree API and the box's sha256sum, 2026-10-16). The row takes --hf once the maintainer names the repo for the docs.

Pre-review CLEAN at 4d04001.

🤖 Generated with Claude Code

[github-actions]

Verdict: CLEAN: The Nemotron commands and limits match the existing serving implementation.

Checked

• Duplicates: 0 new items checked; no code definitions added; no dryer pairs available.
• Complexity: no changed functions.
• Silent failure: 0 added code sites.
• Tests: no added or changed tests; mutation checks do not apply. Runtime tests were not run for this documentation change.
• Gates: no pins, thresholds, or coverage changed.
• Escape catalog: no C1–C4 patterns matched.
• Behavior: checked seven-seat dispatch, local-file parsing, default single-card placement, parallel/placement/draft refusals, and prefix reuse/cache rules.

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)}

[github-actions]

Verdict: CLEAN: The Nemotron command, seat count, and documented limits match the implementation.

Checked

Duplicates: 0 new functions, types, or constants; no dryer pairs reported.
Complexity: no functions changed.
Silent failure: 0 added sites; documentation only.
Tests: no tests changed; mutation check not applicable. Checked command routing, single-slot and placement refusals, draft rejection, prefix retention, prompt stepping, and cache configuration against the code.
Gates: no gates or pins changed.
Escape catalog: no matching patterns.

Notes

Dryer and crapper did not run because no Rust files changed. No runtime tests were run for this documentation change.

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