fix(tests): repair the joint_cross_attn real-Cosmos GPU test

[wayrise]

test_joint_cross_attn_forward_on_real_cosmos has not actually run since commit 4d8df56 ("Document OpenWAM usage and unify component keys"), which replaced configs/model/video_backbone/cosmos_predict25.yaml with cosmos_predict25_2b.yaml. The test's Hydra override was missed, so it raised MissingConfigException before any model code ran:

$ pytest tests/test_cosmos_predict25_joint_cross_attn.py
hydra.errors.MissingConfigException: In 'model/video_backbone': Could not find 'model/video_backbone/cosmos_predict25'

It is the only stale reference left — assets/openwam_usage_docs/train-and-deploy.md and the openwam-serve --help text already use cosmos_predict25_2b, and model_loader.py gates on the cosmos_predict25_ prefix.

The failure is invisible without the Cosmos-Predict2.5-2B checkpoint mounted, so CI skips it and it only surfaces on a box that has the weights — which is how it went unnoticed.

Two lines

1. The config name. model/video_backbone=cosmos_predict25 → cosmos_predict25_2b.

2. The stub_reason1 fixture. With Hydra resolving, the test then hit FileNotFoundError from the live Reason1 text encoder, against the config's /path/to/Cosmos-Reason1-7B placeholder. This was the one real-Cosmos GPU test not requesting stub_reason1, the fixture that exists precisely so these tests skip the 16 GB Qwen load — its closest sibling test_tri_system_forward_on_real_cosmos uses it. The test never needed a real encoder anyway: it passes context to forward() directly.

Verification

Run against real Cosmos-Predict2.5-2B weights on an H200 (OPENWAM_COSMOS25_2B pointed at the checkpoint):

tests/test_cosmos_predict25_joint_cross_attn.py   5 passed in 47s

The GPU case now genuinely loads the DiT and exercises the forward, rather than erroring or skipping. Full suite with the checkpoint visible:

1981 passed, 12 skipped

versus 1 failed, ... 20 skipped before — the 8 fewer skips are the Cosmos GPU tests the checkpoint unlocks. ruff check and ruff format --check clean.

🤖 Generated with Claude Code

[xueminchi]

Reviewed the two-line test repair and the related config, fixture, and model-loading/forward paths. The Hydra override matches the current cosmos_predict25_2b config. Reusing stub_reason1 is appropriate here: context is supplied directly to forward, and the fixture replaces only the text encoder, leaving the real Cosmos DiT load and the existing assertions intact. No blocking issues found. CI lint and tests passed; I did not independently rerun the checkpoint-backed GPU test.