Minimax Music 3 degrades audio at ~3 seconds

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Your question

After experimenting with various samplers, schedulers etc. and not being able to find the source of the issue, this is what I experience and I am out of Latin excuses:
Linux, 128GB RAM, RTX 5090
Updated ComfyUI to latest version. Loaded TEMPLATE for Minimax Music 3. NO CHANGES to the graph except for: FP32 diffusion model and fp16 text encoder (tried BF16 as well to no audible change).
The resulting sound starts fine and at roughly 3 seconds in it degrades like a loss of bit bandwidth. It keeps this kind of quality loss throughout the length of the generated audio. I tried non-tiled audio generation and tiled, no change. As mentioned above, I tried various precisions for the models, no change. I also tried various samplers, seeds etc - but even the STOCK TEMPLATE shows this exact behavior.
I am out of ideas what to try out ... any pointers would be highly appreciated.

### Logs

```powershell

```

### Other

_No response_

No files, tests, or entry points are named, and the Logs section is empty. Start by reproducing the stock Minimax Music 3 template on the reported Linux/RTX 5090 setup, comparing tiled and non-tiled generation and the listed precision and sampler variations. Done means the degradation is isolated to a reproducible cause or the issue is narrowed to an actionable component.

Written by the indexing model from the issue text.