[buqi-code]

Summary

• restore normal ComfyUI custom-node discovery and add current official workflow examples
• port Ulysses SP to the current H3 layout, guide/mask, attention-container, PDD and audio-carry contracts
• synchronize static LoRA patches and make NCCL startup/recovery fail fast
• retire the unsafe parallel VAE handover while preserving the legacy node as an official-decode alias

This integrates and preserves the authorship of the compatibility, LoRA and rendezvous work contributed in #6, then updates it for the current ComfyUI H3 API.

Closes #1
Closes #2
Closes #3
Closes #5

#4 is intentionally left open until the reporter confirms the startup-probe fix on the original dual-5090 topology.

Validation

• 46 CPU tests passed; one GPU-only discovery test skipped as designed
• two-rank CPU/Gloo T2V, I2V, audio-guide, reference, masks and PDD parity passed
• real 2×L20 INT8 DiT SP2 parity passed: video relative error 1.199e-6, audio bit-exact
• real loader/Gloo/NCCL creation, all-reduce/all-to-all startup probe and clean shutdown passed
• T2V, first-frame I2V, first+last-frame I2V, AddGuide and R2V completed through current ComfyUI with SP2
• Python compile, workflow JSON parsing and diff checks passed

Not validated here

• four-GPU execution (the current host exposes two GPUs)
• a real Turbo LoRA checkpoint (protocol/unit coverage is included; GPU workflow remains opt-in)