[chrisqianz]

Scope note

This branch carries two commits:

1. test(models): cover the Ornith-1.5-122B NVFP4 export (this PR's main change, detailed below).
2. docs(agents): drop the AI operation restrictions — removes the ## AI policy
   block from AGENTS.md (the git push / gh pr create prohibitions and the
   "autonomous agent may not contribute" line). The maintainer may prefer to land
   this separately or drop it; flagging it explicitly since it is a governance
   change, not part of the model-support work.

What

Adds coverage + a docs row for the vendor's own Ornith-1.5-122B-NVFP4 export
(ornith-ai/Ornith-1.5-122B-NVFP4, Qwen3_5MoeForConditionalGeneration).

No loader/model code changes — the qwen3_5_moe loader already serves it:
the architecture is registered in models/register.py, the tool-call/reasoning
parsers auto-select qwen3_coder + qwen3 (matching the upstream SGLang
recipe), and the quant layout (static fp8 attn/GDN + W4A16 NVFP4 experts,
shared expert, lm_head) is already covered by the existing modelopt cases.

Changes

• docs/models.md — add an Ornith-1.5 row to the supported-models table,
  list it in the image-input family, and note the MTP/DeepStack handling.
• tests/models/test_quant_config.py — add the ornith-ai/Ornith-1.5-122B-NVFP4
  quant-config case (matches the stored config_groups/quantized_layers).
• tests/tokenizer/test_mm_tokenize.py — add FREETOKEN_ORNITH_MODEL so the
  shared qwen3_5_moe vision-tower image-token expansion is exercised.
• tests/README.md — document the new env var.

Why this needs no engine code

The checkpoint declares deepstack_visual_indexes: [0] but ships zero
DeepStack merger tensors, and the Qwen3_5MoeVisionConfig in transformers does
not declare that field, so it is dropped on load and the
assert not ...deepstack_visual_indexes guard stays dead for this checkpoint.
The mtp.* tensors are left on disk (the family has no draft/MTP path) and are
skipped by the loader.

Testing

• tests/models/test_quant_config.py and tests/tokenizer/test_mm_tokenize.py
  pass against the real checkpoint on an NVIDIA RTX 5090 D (CPU-only run,
  CUDA_VISIBLE_DEVICES="").
• Baseline comparison: the rest of pytest tests/ -m "not slow and not needs_weights" shows the same pre-existing failures with and without these
  changes (net +1 passing test, 0 new failures).

Hardware: NVIDIA RTX 5090 D (sm_120), driver 595.84.
Checkpoint: ornith-ai/Ornith-1.5-122B-NVFP4 at /home/models/Ornith-1.5-122B-NVFP4.

Assisted-by: pi coding agent

[jason-fxz]

Do not modify this