feat(minimax-h3): default workflows; starter-model decision documented (PR 5/5)

[lstein]

Summary

PR 5 of 5 in the MiniMax H3 stack (on top of #141): default workflows, plus a documented decision on the starter-model entry. Two commits, deliberately separated so the license-sensitive part stays droppable.

Commit 1 — default workflows

• Text to Video - MiniMax H3: loader → prompt → denoise → latents-to-video with both video_latents/audio_latents and the video+audio VAE edges. Form exposes model, prompt, width/height, frames, steps. Notes cover the 17n+5 frame grid, fixed 24 fps, 32-px dimension grid, muxed AAC audio, and the absence of a negative prompt.
• First Frame to Video - MiniMax H3: adds minimax_h3_frame_conditioning. The backend's keyframe-coupling contract is enforced structurally: a single image node feeds first_image on both the prompt and frame-conditioning nodes, and shared width/height integer nodes feed prompt + frame conditioning + denoise — form edits cannot desynchronize them.
• Both files are covered by the default-workflow registry test (node types/versions/inputs validated against the live registry; glob count bumped 12 → 14). Defaults verified legal: 124 frames, steps 50, 1344×768.

Commit 2 — starter model: documented TODO, no entry

Neither expressible install source works, so no entry ships yet:

• A bare MiniMaxAI/MiniMax-H3 source downloads ~498 GB — filter_files keeps every model*.safetensors across all subtrees, including FL2VA/, Ref2VA/, and transformer_ref/.
• The ::subfolder+subfolder syntax keeps only files inside the listed subfolders and silently drops the root modular_model_index.json that the H3 probe hard-requires — the install would land as an unknown model.

The TODO in starter_models.py records both unblock paths (a root-config-aware filter_files, or an upstream FL2VA-only repo) and the MiniMax H3 Community License constraints (territory exclusions extend to outputs; "MiniMax H3" attribution). Until then, users install the model manually by pointing the Model Manager at a locally assembled fl2va folder (~42.5 GB working set).

Testing

• tests/app/services/model_install + tests/app/services/workflow_records + tests/model_identification: 146 passed, 0 failures; ruff clean.
• Both JSONs pass WorkflowValidator.validate_json, and every edge handle was checked against the actual invocation input/output models.
• Manual gate: load each default workflow in the Workflow Editor once and run it (the default-workflows README asks for this before release).

🤖 Generated with Claude Code