Attribution for src/tether/models/third_party/dreamzero/ (Lift #7)
===================================================================

This directory contains vendored code from NVIDIA Research's DreamZero
World-Action Model, accessed via FluxVLA's Apache-2.0 integration.

Source chain:
1. Original: NVIDIA Research DreamZero (github.com/dreamzero0/dreamzero)
   License: Apache License 2.0
2. Integrated by: FluxVLA (LimX Dynamics, Apache-2.0)
   Path: fluxvla/models/third_party_models/dreamzero/
3. Vendored here: src/tether/models/third_party/dreamzero/

Modules vendored (~9.5K LOC):
- modules/wan_video_dit.py — Wan2.1 DiT backbone
- modules/wan_video_dit_action_casual_chunk.py — Causal chunked DiT for action
- modules/wan_video_text_encoder.py — T5 text encoder
- modules/wan_video_image_encoder.py — CLIP image encoder
- modules/wan_video_vae.py — Video VAE
- modules/flow_match_scheduler.py — Flow matching noise scheduler
- modules/flow_unipc_multistep_scheduler.py — UniPC sampler
- modules/attention.py — Flash attention wrappers
- modules/cudnn_attention.py — cuDNN attention fallback
- modules/action_encoder.py — State/action tokenizer
- modules/utils.py — Utility functions
- modules/wan2_1_submodule.py — Wan2.1 submodule loader
- modules/wan2_1_attention.py — Wan2.1 attention
- modules/wan_video_camera_controller.py — Camera control conditioning

Apache-2.0 license text: see src/tether/kernels/LICENSE-Apache-2.0.txt
