test(models): Add AutoBridge registration matrix

[yaoyu-33]

Summary

• Add a curated unit-test matrix for AutoBridge model registration.
• Cover baseline and high-churn families: Llama, Qwen3-VL, Qwen3-Omni, Exaone4, GLM 4.5, Sarvam MoE, MiMo V2 Flash, and Step3.5.
• Assert each bridge module imports, the architecture is advertised by AutoBridge.list_supported_models(), and AutoBridge.from_hf_config() dispatches to the expected bridge without weights or model construction.

Validation

• uv run --no-sync ruff check tests/unit_tests/models/test_autobridge_registration_matrix.py
• uv run --no-sync ruff format --check tests/unit_tests/models/test_autobridge_registration_matrix.py
• uv run --no-sync python -m py_compile tests/unit_tests/models/test_autobridge_registration_matrix.py
• git diff --check
• uv run --no-sync pre-commit run --all-files
• Remote standard-container verification: uv run --no-sync python -m pytest tests/unit_tests/models/test_autobridge_registration_matrix.py -q -> 8 passed

Local targeted pytest was attempted after submodule initialization but the workstation environment cannot resolve nvidia-resiliency-ext==0.6.0 for its platform; the remote container run above is the authoritative pytest validation.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

/ok to test cbece1d

[claude]

LGTM — clean, well-structured registration matrix test.

A couple of minor observations:

• Coverage: The test covers 8 of ~22 registered bridges. The PR description says baseline and high-churn families, which is a reasonable scope. Consider adding a follow-up TODO or comment to extend coverage to remaining families (e.g., Qwen2, Qwen3, Qwen3MoE, Mistral, OLMoE, DeciLM/LlamaNemotron, etc.).

• auto_map id for mimo_v2_flash: The minimal_config helper builds auto_map as modeling{case.id}.{case.architecture}. For mimo_v2_flash this yields modeling_mimo_v2_flash.MiMoV2FlashForCausalLM, which matches the real model auto_map in mimo_v2_flash_bridge.py:207. No issue, just noting the convention holds.

Suggested test cases: No perf tests impacted.