Module tensor_roles 

Task #12 / STELLAR §A — tensor-name → engine GEMM-role mapping + the ternary policy.

The transcoder (p64_weight) needs to know what each tensor is to (a) populate the P64 manifest with real engine roles (so a container boots without a GGUF re-parse) and (b) apply the §A compression policy: ternary the FFN projections, keep attention + norms + embeddings at higher fidelity (ternary norms/attention would wreck coherence).

Two naming conventions are recognised:

• GGUF / llama.cpp — blk.{L}.attn_q.weight, blk.{L}.ffn_gate.weight, token_embd.weight, …
• Hugging Face safetensor — model.layers.{L}.self_attn.q_proj.weight, model.layers.{L}.mlp.gate_proj.weight, model.embed_tokens.weight, lm_head.weight, …

Structs

TensorRole

A resolved tensor identity: an engine role + its layer (P64_LAYER_GLOBAL for non-layer tensors).

Functions

name_is_ternary_eligible

True iff this tensor name resolves to an FFN projection (so it is ternary-eligible). A name we cannot classify is not ternarised (fail safe to high fidelity).

name_to_role

Map a tensor name (GGUF or HF convention) to its engine role + layer, or None if unrecognised.

ternary_eligible

The §A ternary policy: only the FFN projection weights (gate / up / down) are eligible for BitNet-1.58b ternary packing. Attention projections, norms, and embeddings stay at higher fidelity — ternarising them destroys coherence.