nx_modality_router.nx

nx_modality_router.nx -- H10 multi-modal model sharing (bits-up). Per NISHI_ELDER_AI_OFF_DOCKER_2026_05_20.md §2.1 H10: one shared backbone trunk serves multiple modality-specific heads concurrently -- text response, image prompt conditioning, audio embed, vision token embed. Stable Diffusion shares one UNet across N LoRA + ControlNet adaptations; Llava + Flamingo share one LLM trunk across text + vision modalities; the operator's "multiused the llm for response and image gen" cardinal IS this pattern. V1 mechanics: two registries -- backbones (one trunk weight reference each) + modality heads (small per-modality projection layers that attach to a backbone). Registering a head ACQUIRES the named backbone (bumps refcount); unregistering RELEASES it. A backbone with refcount > 0 cannot be evicted. resolve_backbone_ for_modality() walks the modality -> backbone link in one call. Composition path: - register_backbone(id, weights_data, hidden_dim, n_layers) ONCE per loaded base model - register_modality(id, backbone_id, head_data, kind) per modality the consumer wants to serve; this acquires the backbone - resolve_backbone_for_modality(modality_id) at request-time gives the consumer the backbone pointer + the head pointer in one call, so the forward-pass kernel can run trunk + head in sequence without a second lookup - unregister_modality on shutdown releases the backbone; evict_backbone only succeeds when no modality references it Pure substrate logic. No Linux features. Composes with shipped H9 nx_lora_pool (modality head's "head_data" pointer can itself be an adapter_id resolved through the LoRA pool, letting one backbone serve N modalities each with M fine-tuned adapter variants). V1 honest scope: - Sealed-enum modality kinds: TEXT, IMAGE, AUDIO, EMBED. A V2 could add VIDEO + PROPRIOCEPT etc.; the sealed-enum is