Summary

Two recipe modules under src/megatron/bridge/recipes/qwen_vl/ had no unit-test coverage:

L0 functional scripts are slow and GPU-bound — they don't catch the kind of regression that issue #3177 was filed about ("Kimi K2 had a recipe but no recipe test, so we missed an MCore API breaking change").

This PR adds 44 parametrized unit tests modelled after test_glm_45v_recipes.py and test_kimi_k2.py. Each recipe is built with a stubbed AutoBridge (no HF Hub I/O) and a no-op apply_flex_dispatcher_backend (no torch.cuda dependency), then asserted against the documented contract.

Refs #3177. Tests-only — no production code changes.

What's covered

tests/unit_tests/recipes/qwen_vl/test_qwen3_vl.py — 16 tests

tests/unit_tests/recipes/qwen_vl/test_qwen35_vl.py — 28 tests

Implementation notes

• AutoBridge stub — _FakeAutoBridge.from_hf_pretrained() returns a permissive provider that absorbs arbitrary setattr calls. Only the attrs the recipes read (vocab_size, apply_rope_fusion) get explicit defaults.
• No num_moe_experts on the fake provider — that means even without the patch on apply_flex_dispatcher_backend, the function would short-circuit on its own first guard before touching torch.cuda.get_device_properties(0). The fixture additionally patches the helper to a no-op so the test stays independent of that internal early-return.
• Two-module patch — qwen35_vl pretrain delegates into _qwen3_vl_common, which references the qwen3_vl module's AutoBridge symbol. The qwen35_vl fixture patches both modules so the delegation path stays HF-free regardless of which symbol the recipe touches.
• Energon variant — qwen3_vl_8b_peft_energon_config calls AutoTokenizer.from_pretrained and Qwen3VLProcessor.from_pretrained directly inside _make_energon_dataset. The dedicated test mocks both at the recipe-module level and asserts the resulting dataset is an EnergonProvider.

Test plan

• python3 -m ast parse of new files
• ruff check clean
• ruff format --check clean
• CI: L0 unit tests pick up the new modules automatically (no workflow file changes needed). Both modules live under tests/unit_tests/recipes/qwen_vl/, which is already covered by cicd-unit-tests-core.

Risk

Zero — tests only. No production code changes.

Notes for reviewers

• I considered combining the qwen3_vl and qwen35_vl modules into a single test file but kept them split because (a) qwen35_vl needs to patch two modules' AutoBridge (the inheritance through _qwen3_vl_common) and the autouse fixture is cleanest when scoped to one test class per module, and (b) the recipe families have different attribute-touching surfaces (qwen35_vl sets MTP, qwen3_vl uses the flex-dispatcher helper), so split files keep each test focused.
• The energon variant test is the only one that needs to mock AutoTokenizer / Qwen3VLProcessor at the recipe-module level. All other recipes go through AutoBridge only.
• The recipe family also includes qwen25_vl.py — that one already follows the "flat" pattern this issue asks for, but is not yet covered by unit tests either. I'd happily land that in a follow-up if the maintainers want — it's the same pattern.