# Shared Model Backbone

This package is the single integration point for modules that need Qwen layers.

## Contract

- `register_shared_model(model, tokenizer)` registers the already-loaded Qwen/GptOss model.
- `get_shared_adapter()` returns a `SharedBackboneAdapter`.
- Subsystems use the adapter for:
  - `get_embedding_layer()`
  - `get_layers()`
  - `get_layer(index)`
  - `encode_text(text)`
  - model `device`, `dtype`, `hidden_size`

## Rule

Modules must not instantiate fallback language models or random embeddings for normal operation. If the shared adapter is unavailable, they should skip optional work or return an empty result.