Integrate a model into flashdreams

The ordered procedure for binding an external video model to the flashdreams framework. Read the flashdreams-integrations skill first for the architecture (layers, contracts, the cache tree) — this skill is the route, that one is the map.

Worked example throughout: integrations/hy_worldplay/ (HY-WorldPlay WAN-5B I2V), which reuses the integrations/wan22/ Wan 2.2 TI2V-5B recipe. It is the most complete reference integration; read it side-by-side. Match python-docstring-style.

The core bet: reuse, don't re-implement

Most modern video models are DiT-family. Before writing anything, find the closest existing flashdreams recipe (integrations/wan22, wan21, self_forcing, …) and subclass it. HY-WorldPlay is a Wan 2.2 TI2V-5B with three conditioner deltas — it adds ~3 small subclasses, not a from-scratch network. If your model maps onto an existing backbone, the job is config + checkpoint remap + deltas + verify, which is days–weeks. If it needs a novel network/attention/inference loop, it is much longer — say so up front.