Summary

• add pinned local text, CLIP media, and CLAP audio backends
• decode real image/audio/video/3D content without descriptor or precomputed-vector shortcuts
• isolate incompatible embedding spaces and fuse only ranked evidence
• add bidirectional text/media projections with provenance, rollback, and logical deletion

Validation

• 47 targeted multimodal tests passed
• Ruff passed
• package build and twine check passed
• real pinned model smoke passed for text, PNG, MP4, OBJ, and WAV

Claim boundary

This PR implements and validates the real local encoder/runtime contract. It does not claim the 1000-asset multimodal admission benchmark has passed.