feat(inference): Cosmos3-Edge num_frames default + Edge GPU smoke tests

[lfengad]

What

Adds a Cosmos3-Edge–specific num_frames inference default and two Edge GPU smoke tests, mirroring the existing Nano suite and wiring both into the current inference-smoke CI jobs.

num_frames default for Edge

• args.py: default num_frames to 121 for Cosmos3-Edge video generation (keyed on the reasoner vlm model_name), following the existing per-model _RESOLUTION_SHIFT_DEFAULTS pattern. Other models keep the 189 modality default. The default is scoped to plain video generation only — it is not applied to image modes (single-frame), action modes (own default), or the reasoner (reports VIDEO vision_mode but uses num_frames as an inert 1).
• resolution (480, from model config) and fps (24, shared) were already correct for Edge, so only num_frames needed a model-specific default.
• args_test.py: test_edge_num_frames_default — Edge=121 / Nano=189 / explicit-override, plus regressions proving the scope: Edge text2image→1, Edge policy→189 (action untouched), Edge reasoner→1. All build on CPU (no GPU/assets).
• docs/inference.md: note the Edge default.

Edge GPU smoke tests (mirror Nano)

• tests/edge_inference_smoke_test.py (8-GPU): t2v + policy + forward_dynamics. Edge has no sound (sound_gen: false), so t2vs → plain t2v; the t2v sample gets the stronger whole-clip content check (_assert_video_has_content) and asserts the Edge 480/121/24 generation defaults.
• tests/edge_reasoner_inference_smoke_test.py (4-GPU): image-conditioned reasoner. Golden-free smoke that validates the generated reasoner_text is non-empty, coherent (min length + lexical diversity), and a complete description — it must reference both the dominant robotic subject and the table/lab setting.
• gpu-tests.yml: both Edge tests added as steps in the existing generator-inference-smoke / reasoner-inference-smoke jobs (reusing checkout + uv sync + input cache).

Verification

Run on a 4×GB200 node with the repo .venv (container guardrails enabled):

• Unit test_edge_num_frames_default: passes on CPU (6 cases incl. image/action/reasoner scope regressions).
• num_frames=121 default, end-to-end: a real Edge t2v run (no --resolution/--num-frames/--fps; t2v.json pins none) resolved to resolution=480, num_frames=121, fps=24 and produced a valid 121-frame, 832×480 video (pixel std 66) — confirming the default takes effect in real inference, not just at build time.
• Reasoner smoke: passed on 4×GB200. Edge produced an accurate description — "two robotic arms with mechanical hands … on a table with wooden structures … in a lab or research environment …" — hitting both the subject and scene anchors, so the content check matches real output.
• Not run locally (4-GPU node): the 8-GPU generator smoke's policy / forward_dynamics paths — first real execution is on the CI 8-GPU runner. The reasoner smoke was verified on GB200; CI runs on H200 (greedy decode + broad anchors make a pass expected, not hardware-proven).

🤖 Generated with Claude Code