FLUX 3 Prompt Enhancement GGUF

The PUP model and system prompts are by Black Forest Labs, derived from
Qwen3.5-9B, and distributed under Apache-2.0.
Source model: black-forest-labs/flux-3-image-klein-pup
Revision: 8b069751eb5dcab23829d9c5eccb470eb8b8a4ed

The separately distributed Heretic GGUF variant is converted from
tuanlai/flux-3-image-klein-pup-heretic on ModelScope:
https://modelscope.cn/models/tuanlai/flux-3-image-klein-pup-heretic
Weight-file revision: 80efba65e01bb2e320dea412fa657d701491b443
Source SHA256: beee7e8b4aa5e5ff1b9342d70b46022f9508aa6ae5010c76427ced6a0f0a8f50
Its upstream model card has separate license declarations; see
docs/huggingface/LICENSES.md in the node repository or LICENSES.md on the Hub.
Heretic weights are not relicensed by this integration.

flux3_pe/bbox.py is copied from Black Forest Labs' Apache-2.0 reference code:
https://github.com/black-forest-labs/flux-3-image-klein
Revision: 45e57c9fb06f9751cdced29a65c5d2171003be05
Image preprocessing and request scaffolds follow that reference implementation.

T8star provides GGUF conversion, ComfyUI integration, and modifications to
layout validation, identity reconciliation, caching, and process cleanup.

The separately distributed llama.cpp runtime is MIT licensed. Its generated-token-only
penalty extension is recorded in patches/llama-generated-penalties.patch.
The optional Python backend uses the separately installed MIT-licensed
llama-cpp-python package: https://github.com/abetlen/llama-cpp-python
The image-generation model is a separate artifact with its own license.
