pytest 9's monkeypatch reads the previous value from the instance __dict__.
torch._dynamo.config serves its flags through __getattr__, so the first
test's teardown deleted the disable entry and every later setattr on it
failed: 28 setup errors in the Unit lane's test_lora_side_branch.py. The
same cause, not torch 2.12, is behind the TF32 teardown errors.
- Use torch._dynamo.config.patch(disable=...) through fixtures in the LoRA
side-branch and merge-precision tests.
- Correct the tf32_disabled fixture docstring.
- Note that the Qwen-Image-2.1 block references follow the golden lane's
FA2 runtime.
- Record the pitfall in .agents/lessons.
Aryan Kumar
27180f14e2