# Multi-Diffusion Fusion Integration Plan

To accommodate the new wave of Chinese long-video AI generation models (e.g., Wan, Kling, HunyuanVideo) which leverage multi-diffusion fusion for extended temporal consistency, we need to wire up a dedicated generation pipeline inside the core.

## User Review Required
Please review the proposed integration approach below.

## Open Questions
> [!WARNING]
> 1. **Engine Preference**: Do you have a specific target engine in mind (e.g., Kling AI, HunyuanVideo API), or should I build a generic multi-diffusion fusion stub that can interface with any of them?
> 2. **Execution Blocking**: Should the video generation block the `ArchitectBlade`, or should it be offloaded as an asynchronous background task while the mesh continues its cycle?

## Proposed Changes

### `services/engines/media/`
#### [NEW] [diffusion_fusion.py](file:///C:/AxeonPrime/services/engines/media/diffusion_fusion.py)
- Create a new `MultiDiffusionFusionEngine` class.
- Implement methods for `generate_long_video(prompt: str, duration_sec: int)`.
- Enforce temporal consistency checks and multi-frame latent blending (simulated API wrappers).

### `core/execution_blades/architect.py`
#### [MODIFY] [architect.py](file:///C:/AxeonPrime/core/execution_blades/architect.py)
- Import the new `MultiDiffusionFusionEngine`.
- Update `_handle_media_gen` to detect if the request is for long-form video.
- Initiate the diffusion fusion generation task and embed the tracking ID into the returned `DecisionProposal`.

## Verification Plan
1. Send a simulated long-video generation request to the Architect Blade via WebSocket or direct unit test.
2. Verify the intent triggers the `MultiDiffusionFusionEngine`.
3. Verify the generated `DecisionProposal` contains the correct multi-diffusion context before being passed to the `ConsensusEngine`.