feat(models): vendor Nemotron-H Puzzle checkpoint support

[pierre427]

feat(models): vendor Nemotron-H Puzzle checkpoint support

Source: ml-explore/mlx-lm#1536

Review state: validated — real 75B server/tool-call gate passed; MTP unavailable in the tested artifact

Why

Nemotron-H Puzzle matters to Rapid users because it is a practical 75B / ~9B-active hybrid MoE that can run locally on Apple Silicon with a 262k context window and strong tool-calling behavior. The port lets Rapid instantiate the heterogeneous per-layer Puzzle variant while keeping hybrid SSM and speculative-decode safety gates conservative.

Summary

Add native-first nemotron_h compatibility for the Nemotron Labs Puzzle
checkpoint while preserving Rapid-MLX's conservative hybrid-model safety
gates.

AI assistance disclosure

OpenAI Codex Terra implemented the initial adaptation. The coordinator
independently inspected the diff, repaired process-global test cleanup,
verified public checkpoint metadata, corrected the tool parser profile to the
checkpoint's Nemotron XML template, and reran focused tests, Ruff, format, and
whitespace checks.

Implementation

• Vendor the Nemotron-H architecture only when the installed mlx-lm has no
  native module.
• Register the architecture in Rapid's direct loader fallback.
• Add a text-only Nemotron Puzzle profile with the verified Nemotron tool-call
  parser and Qwen3 reasoning parser.
• Preserve the checkpoint's MTP sidecar metadata without claiming safe
  speculative decoding: Rapid still lacks validated hybrid SSM
  rollback/replay for this architecture.

Test plan

• pytest -q tests/test_nemotron_puzzle_vendored.py tests/test_aliases_contract.py — 2545 passed.
• Ruff check on all changed Python files.
• Ruff format check on all changed Python files.
• git diff --check.
• Public checkpoint metadata verified: nemotron_h, 88 block
  descriptors, one configured MTP layer, 13 weight shards, and zero MTP
  tensor keys in the tested community 6-bit artifact.
• Parser profile corrected to match the checkpoint's
  <function> / <parameter> Nemotron template.
• Real 75B GPU server validation on Apple Silicon: Rapid loaded the
  community 6-bit artifact, /v1/models advertised 262144 context with
  text/tools capabilities, ordinary chat returned Paris, and a forced
  Nemotron-format weather tool request produced
  get_weather({"city":"Toronto"}).

Remaining validation

The tested community 6-bit artifact has no MTP tensor keys, so MTP decode was
not a valid validation target for this PR. mlx-lm#1533 remains deferred until a
Puzzle checkpoint with MTP weights is available and Rapid has a safe hybrid SSM
rollback/replay contract.

Attribution

The vendor module preserves Apple copyright and Apache-2.0 attribution and is
adapted from mlx-lm #1536 for Rapid-MLX's native-first vendoring seam.

---

Draft opened from pierre427 fork after duplicate-checking fresh raullenchai/Rapid-MLX@b3ccb921cd2d435b7238d89b079134258f7d3485.