fix(minimax-h3): REF canvas renders COHERENT + drain the scratch pool at denoise→decode

[localai-bot]

What

Two things, both grounded in a full REF-canvas run (1344x768 / 124f) that §8.4 had left unrun ("~2.85 h, so it was not run").

1. The REF canvas renders COHERENT — measured

The §8.4 white latent does not reproduce at the REF canvas on the matching partition.

Stage	Metric	Value	Reference
init noise	adj-cell cosine	0.0019	white by construction (control)
VAE-input latent	adj-cell cosine	0.8924	§8.4 white 0.06 / real encode 0.789 / §8.6 coherent 0.9467
decoded frames	period-16 seam ratio	1.15	== the known-good 864x480 render

The earlier grids were task/partition, measured on one canvas and prompt: ref2va-on-FL2VA 2.28 → t2va pre-guard 1.87 → t2va on the guarded binary with Gaussian init noise 1.15. The artifact scales with canvas, which is why small canvases looked acceptable and the REF canvas did not.

2. The decode OOM — found and fixed

The 50-step denoise completed, then VAE decode took the box down: NV_ERR_NO_MEMORY, then minimax-h3-gen invoked oom-killer. Twice.

The scratch pool is uncapped on this platform (cuda.cpp:98, device_pool_cap_bytes = 0), so every block a DBuf returns across 50 denoise steps is retained for the life of the process, keyed by size class. The decode allocates different classes, reuses none of them, and cudaMallocs on top of the whole retained set.

DevicePool::Drain releases them and reports the bytes; MiniMaxH3GenerateT2va calls it once at the denoise→decode boundary.

[h3] drained 10.25 GiB of denoise scratch before VAE decode
  wrote 124 frames (1344x768 @ 24 fps) + audio.wav (32000 Hz)
[DevicePool] hits=98475 misses(cudaMalloc)=75 hit-rate=99.92% distinct-classes=8

10.25 GiB released, render completed with audio, no reboot — and the pool still served 99.92% from the free list, so draining at a phase boundary costs nothing of the within-phase reuse the pool exists for. Per-step forward time was unchanged at 176.3 s.

Alternatives rejected

• Capping the pool platform-wide — charges every CUDA model the cudaMalloc/cudaFree sync storm the pool was built to avoid, to fix a problem that appears at one phase change.
• --keep-quant as a no-code workaround — is not one. OOM-killed the box ~8 min in, during weight staging, before a single denoise step.

Safety

free_ only ever holds blocks a DBuf already returned, so nothing live is touched. Under VT_POOL_BYPASS the free list is always empty and Drain is a no-op. The drain is called from the H3 pipeline only; no other model's path changes.

Also

Corrects a comment claiming decoder_tiling "defaults false" while minimax_h3.h has bool decoder_tiling = true. Tiling composes with temporal chunking by default and is required, not optional — the ViT3D's RoPE is length-normalized over the grid it is handed. The stale comment cost this row a wrong hypothesis.