[H3] Periodic dark frames & frame stutter every 17 frames across all param combos

### Describe the bug

MiniMax H3 generates periodic dark frames and noticeable frame boundary stutter at exactly **17-frame intervals**. This occurs across **all** combinations of sampler/scheduler/step count/resolution/text-encoder tested. The official MiniMax SGLang-generated sample video (t2va_2k.mp4) does **not** exhibit this pattern.

### Environment

- **GPU**: RTX 3060 12GB | **RAM**: 32 GB | **OS**: Windows 10, Driver 560.94
- **ComfyUI**: v0.30.0 + cherry-pick #15268 #15334 #15243 #15377
- **PyTorch**: 2.13.0+cu126 | kitchen 0.2.26 | aimdo 0.4.11
- **Models**: fp8_scaled diffusion + nvfp4_awq TE + video_vae_fp16 + audio_vae_fp32 (all verified)
- **Workflow**: official T2V template, simple scheduler, res_multistep, 20 steps, 832×480

### Symptom (quantitative data)

**Frame brightness** every 17th frame (VAE clip_length=17 boundary) drops ~30 units (e.g. 157 vs 190):

```
Frame 0: 157(dark) 1-16: 183-194(normal)
Frame 17: 160(dark) 18-33: 185-195(normal)
Frame 34: 161(dark) ...
Frame 51: 154(dark) ... (gets darker each cluster)
```

**Frame-to-frame differences** show massive spikes at cluster boundaries:

| | Normal frames | 17-frame boundary | 4-frame (token) boundary |
|---|---|---|---|
| Mean abs diff | 2.2–3.5 | **26–34** | 6–8 |

**Official reference** (t2va_2k.mp4): consistent frame diffs 2.5–3.4, **no periodic spikes**.

### Tried (all same pattern)

- All scheduler/sampler combos: simple+res_multistep, karras+dpmpp_2m
- Steps: 20, 30, 40
- Resolution: 832×480, 512×768, 640×352
- Duration: 56 frames, 124 frames
- TE: int8_convrot vs nvfp4_awq
- Diffusion: fp8_scaled vs int8_convrot
- After Cherry-picking master fixes #15268 #15334 #15243 #15377

### Note

VAE post_quant_conv/quant_conv initially corrupt (all-zero tail from incomplete download), repaired and verified before testing.

No file or test is named in the report; first reproduce the official T2V template with the stated 17-frame VAE clip length and compare it with t2va_2k.mp4 using the reported frame-difference measurements. Done means identifying the source of the 17-frame darkening and eliminating the boundary stutter across the tested parameter combinations.

Written by the indexing model from the issue text.