feat(nemotron3): add Nemotron-3 Ultra chat-template variant

[hallerite]

Summary

Adds the Nemotron-3 Ultra chat-template variant to the nemotron-3 renderer. Ultra's template differs from Nano/Super in three ways:

• Reasoning glue: <think>\n{reasoning}</think>{content} — no \n around </think> (Nano/Super use <think>\n{reasoning}\n</think>\n{content}).
• Historical truncation: dropped-reasoning turns collapse to <think></think>{content} with no separating \n.
• Truncation boundary: thinking is dropped on every assistant turn before the last user message (the template's loop.index0 < last_user_idx rule), rather than preserving only the last plain assistant.

Changes

• Nemotron3RendererConfig.ultra: bool | None = None — None auto-detects the variant from the model name. Marked _internal_fields (it selects a template variant, not a Jinja kwarg), so the parity matrix doesn't cross it as a template field.
• Variant auto-selected from tokenizer.name_or_path via _ULTRA_DEFAULTS + _default_ultra, materialized in __init__ — mirrors Qwen3.5's _ENABLE_THINKING_DEFAULTS. Unknown / fine-tuned / local-path checkpoints fall back to the Nano/Super template; pass an explicit ultra= to override.
• Ultra BF16 + FP8 checkpoints mapped to the nemotron-3 renderer in MODEL_RENDERER_MAP.
• Ultra added to the config-parity, shared-barrage (conftest), and roundtrip test matrices (BF16 as the representative), plus an offline test pinning the name-based selection and the model→renderer mapping.

Validation

Token-parity verified against the canonical Ultra apply_chat_template across system / no-system, reasoning, multi-turn truncation, tool-call (including consecutive tool results and varied argument types), and generation-prompt shapes, with enable_thinking and truncate_history_thinking both toggled. Nano/Super behavior is unchanged — the Ultra branches are gated on the resolved flag, which is False for them.

CI uses BF16 as the Ultra representative; BF16 and FP8 share the same tokenizer and template.

Pre-existing renderer content-normalization behaviors (e.g. inline <think>…</think> inside content rather than reasoning_content, and whitespace stripping of message content) are unchanged by this PR and apply identically to Nano/Super — out of scope here.

---

Note

Medium Risk
Changes tokenization and history-thinking rules for Ultra checkpoints; incorrect ultra resolution would break HF template parity and round-trips, though Nano/Super paths stay gated on ultra=False.

Overview
Adds Nemotron-3 Ultra as a second chat-template variant on the existing nemotron-3 renderer, selected by Nemotron3RendererConfig.ultra (auto from tokenizer.name_or_path via _ULTRA_DEFAULTS, overridable explicitly).

Ultra vs Nano/Super behavior when ultra=True: historical thinking is dropped on assistant turns before the last user message (not only before the last plain assistant); reasoning blocks glue as {content} without separating newlines around and on collapsed turns.

Registers Ultra BF16 and FP8 in MODEL_RENDERER_MAP, excludes ultra from template parity fields (_internal_fields), and extends config-parity, conftest, roundtrip (BF16), plus offline tests for name-based selection and FP8 mapping.

^{Reviewed by Cursor Bugbot for commit f63fa30. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Nemotron-3 Ultra chat-template variant to Nemotron3Renderer

• Adds an ultra boolean field to Nemotron3RendererConfig; when None (default), it is auto-resolved at renderer init time by matching tokenizer.name_or_path against a new _ULTRA_DEFAULTS lookup table.
• Maps the nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 and nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-FP8 checkpoints to the nemotron-3 renderer in MODEL_RENDERER_MAP.
• Ultra rendering differs from Nano/Super in two ways: </think> is emitted directly adjacent to content (no surrounding newlines), and thinking is truncated on every assistant turn before the last user message rather than only at the final plain assistant turn.
• ultra is marked as an internal field via _internal_fields so it is not exposed as a template kwarg.
• Behavioral Change: existing callers using Nano/Super Nemotron-3 checkpoints are unaffected; Ultra checkpoints previously fell back to DefaultRenderer and now use the nemotron-3 renderer.

^{Macroscope summarized f63fa30.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR adds support for a new Nemotron-3 Ultra chat-template variant with new configuration, auto-detection logic, and conditional rendering paths that affect how messages are formatted differently from the existing Nano/Super variants. New feature capability warrants human review.

^{You can customize Macroscope's approvability policy. Learn more.}