fix(comfyui): T2V workflow pairs 14B models with the 5B VAE, breaking every run at VAEDecode

[mandaar17]

Problem

The bundled T2V workflow tools/_comfyui/workflows/wan22-t2v-4step.json pairs the WAN 2.2 14B FP8 diffusion models (wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors / wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors) with the wrong VAE: its VAELoader references wan2.2_vae.safetensors, which is the WAN 2.2 5B model's VAE and expects 48-channel latents. The 14B models produce 16-channel latents, so every text-to-video run through this workflow fails at the VAEDecode node.

Repro

1. Install the models listed in _REQUIRED_MODELS_T2V (tools/video/comfyui_video.py).
2. Run any T2V generation through comfyui_video (or queue wan22-t2v-4step.json directly in ComfyUI).
3. Sampling completes, then VAEDecode fails with:

Expected tensor to have size 48 at dimension 1, but got size 16

Fix

• wan22-t2v-4step.json: vae_name → wan_2.1_vae.safetensors, the VAE the WAN 2.2 14B models require. This matches the sibling wan22-i2v-4step.json, which already uses it with the same 14B FP8 model family.
• tools/video/comfyui_video.py: update _REQUIRED_MODELS_T2V to list wan_2.1_vae.safetensors so the preflight/required-models check matches the VAE the workflow actually loads (previously it would demand a model the workflow no longer uses, and not check for the one it does).

Verified locally: with the corrected VAE, the same workflow decodes and renders T2V output successfully on a RTX 4090.

🤖 Generated with Claude Code

[calesthio]

Thanks for the focused fix. I reviewed this against docs/PR_REVIEW_GUIDE.md and checked the PR description, changed files, and current review/comment state. The scope is tight: the bundled WAN 2.2 T2V workflow and the required-model preflight list now both reference wan_2.1_vae.safetensors, matching the 14B FP8 model family and the sibling I2V workflow.\n\nValidation run locally on the PR head:\n- python -m py_compile tools/video/comfyui_video.py -> passed\n- JSON/workflow consistency check -> VAELoader node 4 uses wan_2.1_vae.safetensors\n- Required-model consistency check -> tools/video/comfyui_video.py requires wan_2.1_vae.safetensors and no longer requires wan2.2_vae.safetensors\n\nI did not find unrelated scope, dependency, security, or provider-discovery risk in this diff. The live ComfyUI decode path still depends on having the WAN models installed, but the PR author's RTX 4090 verification plus the workflow/preflight alignment are enough for this narrow fix.