MiniMax H3 AudioVAE input is center-cropped, dropping leading audio samples and producing incorrect latent lengths

### 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)

### Expected Behavior

MiniMax H3 audio should reach MiniMaxH3AudioVAE.encode() without generic center-cropping.

The H3 AudioVAE already handles non-aligned audio lengths itself by right-padding to the next 800-sample boundary, so the beginning of the input waveform should be preserved.

### Actual Behavior

ComfyUI's generic VAE preprocessing first center-crops H3 audio down to a multiple of 800 samples.

For example, an input of 437333 samples is cropped to 436800 samples before reaching MiniMaxH3AudioVAE.encode().

This removes approximately 266 samples from the beginning (~8.3 ms at 32 kHz) and causes the encoder to produce 546 audio latent steps instead of the 547 steps its own right-padding behavior would produce.

### Steps to Reproduce

Start the latest ComfyUI with all custom nodes disabled.
Create or use a 32 kHz WAV file whose sample count is not divisible by 800.
For example, a 437333-sample file is about 13.67 seconds long.
Load minimax_h3_audio_vae_fp32.safetensors with the stock VAELoader.
Use the audio as a reference input to the stock MiniMax H3 Reference to Video node. This node passes reference audio through the normal ComfyUI H3 AudioVAE encode path.
Trace/log the waveform shape entering MiniMaxH3AudioVAE.encode().

For a 437333-sample input, the AudioVAE receives only:

436800 samples

rather than the original 437333 samples.

The missing 533 samples are removed by the generic VAE preprocessing before the H3 AudioVAE is called. Because this is a centered crop, roughly 266 samples are removed from the beginning.

MiniMaxH3AudioVAE.encode() itself is designed to receive the complete waveform and right-pad it to an 800-sample boundary.

### Debug Logs

```powershell
python main.py \
--listen \
--disable-all-custom-nodes \
--log-stdout
[INFO] setup plugin alembic.autogenerate.schemas
[INFO] setup plugin alembic.autogenerate.tables
[INFO] setup plugin alembic.autogenerate.types
[INFO] setup plugin alembic.autogenerate.constraints
[INFO] setup plugin alembic.autogenerate.defaults
[INFO] setup plugin alembic.autogenerate.comments
[INFO] Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope1_', 'apply_rope_', 'apply_rope_split_half', 'apply_rope_split_half1', 'apply_rope_split_half1_', 'apply_rope_split_half_', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'dequantize_w4a8_int8_weight', 'gemv_awq_w4a16', 'int8_linear', 'na3d', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'quantize_w4a8_int8_weight', 'rms_adaln', 'rms_rope', 'rms_rope1', 'rms_rope1_', 'rms_rope_', 'rms_rope_split_half', 'rms_rope_split_half1', 'rms_rope_split_half1_', 'rms_rope_split_half_', 'rotate_int8_convrot_weight', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8', 'w4a8_int8_linear']}
[INFO] Found comfy_kitchen backend triton: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope1_', 'apply_rope_', 'apply_rope_split_half', 'apply_rope_split_half1', 'apply_rope_split_half1_', 'apply_rope_split_half_', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'int8_linear', 'na3d', 'quantize_and_rotate_rowwise', 'quantize_int8_rowwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'rms_adaln', 'rms_rope', 'rms_rope1', 'rms_rope1_', 'rms_rope_', 'rms_rope_split_half', 'rms_rope_split_half1', 'rms_rope_split_half1_', 'rms_rope_split_half_', 'w4a8_int8_linear']}
[INFO] Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope1_', 'apply_rope_', 'apply_rope_split_half', 'apply_rope_split_half1', 'apply_rope_split_half1_', 'apply_rope_split_half_', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_embedding', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'dequantize_w4a8_int8_weight', 'gemv_awq_w4a16', 'int8_linear', 'na3d', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'quantize_w4a8_int8_weight', 'rms_adaln', 'rms_rope', 'rms_rope1', 'rms_rope1_', 'rms_rope_', 'rms_rope_split_half', 'rms_rope_split_half1', 'rms_rope_split_half1_', 'rms_rope_split_half_', 'rotate_int8_convrot_weight', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8', 'w4a8_int8_linear']}
[INFO] Found comfy_kitchen backend hip: {'available': False, 'disabled': False, 'unavailable_reason': 'PyTorch ROCm/HIP runtime not available', 'capabilities': []}
[INFO] Checkpoint files will always be loaded safely.
[INFO] Total VRAM 32111 MB, total RAM 256808 MB
[INFO] pytorch version: 2.13.0+cu130
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 NVIDIA GeForce RTX 5090 : cudaMallocAsync
[INFO] Using async weight offloading with 2 streams
[INFO] Enabled pinned memory 231127.0
[INFO] Using pytorch attention
[INFO] aimdo: /project/src/control.c:276:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 5090 (VRAM: 32110 MB)
[INFO] DynamicVRAM support detected and enabled
[INFO] Python version: 3.13.14 | packaged by Anaconda, Inc. | (main, Jul 9 2026, 14:36:31) [GCC 14.3.0]
[INFO] ComfyUI version: 0.34.0
[INFO] comfy-aimdo version: 0.4.15
[INFO] comfy-kitchen version: 0.2.31
[INFO] comfyui-frontend-package version: 1.51.9
[INFO] comfyui-workflow-templates version: 0.11.50
[INFO] comfyui-embedded-docs version: 0.5.10
[INFO] comfy-kitchen version: 0.2.31
[INFO] comfy-aimdo version: 0.4.15
[INFO] Asset seeder disabled
[INFO] No OpenGL_accelerate module loaded: No module named 'OpenGL_accelerate'
[INFO] Skipping loading of custom nodes
[INFO] Context impl SQLiteImpl.
[INFO] Will assume non-transactional DDL.
[INFO] Using RAM pressure cache.
[INFO] Starting server
[INFO] got prompt
[INFO] VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
[INFO] Requested to load MiniMaxH3AudioVAE
[INFO] Model MiniMaxH3AudioVAE prepared for dynamic VRAM loading. 576MB Staged. 0 patches attached. Force pre-loaded 401 weights: 539 KB.
[INFO] Model MiniMaxH3AudioVAE prepared for dynamic VRAM loading. 576MB Staged. 0 patches attached. Force pre-loaded 401 weights: 539 KB.
[INFO] Prompt executed in 1.21 seconds

python - <<'PY'
import av
from pathlib import Path

files = [
Path("input/h3_audio_crop_repro_437333.wav"),
max(Path("output/h3_crop_repro").glob("decoded*.flac"),
key=lambda p: p.stat().st_mtime),
]

for path in files:
container = av.open(str(path))
stream = container.streams.audio[0]

samples = 0
sample_rate = stream.codec_context.sample_rate

for frame in container.decode(stream):
samples += frame.samples

print(path)
print(f" sample_rate: {sample_rate}")
print(f" samples: {samples}")
print(f" duration: {samples / sample_rate:.9f} s")
PY
input/h3_audio_crop_repro_437333.wav
sample_rate: 32000
samples: 437333
duration: 13.666656250 s
output/h3_crop_repro/decoded_00001.flac
sample_rate: 32000
samples: 436800
duration: 13.650000000 s
```

### Other

I found this while debugging an H3 V2V workflow where the target required 547 audio latent steps but source-audio encoding produced only 546.

The initial workaround was to append one zero latent step. Tracing the encode path showed that the mismatch originates earlier: the generic VAE crop happens before the H3 AudioVAE gets the waveform, and also shifts the beginning of the audio.

A possible narrow fix is to disable generic VAE input cropping specifically for the MiniMax H3 audio VAE:

self.crop_input = False

This would allow MiniMaxH3AudioVAE.encode() to perform its existing model-specific right-padding without changing generic VAE behavior for other models.

Start at MiniMaxH3AudioVAE.encode() and trace the generic VAE preprocessing that receives the 32 kHz waveform before it. Reproduce with the 437333-sample WAV and --disable-all-custom-nodes; done means the full waveform reaches encode unchanged and the existing right-padding produces 547 latent steps without changing other VAEs.

Written by the indexing model from the issue text.