Semantha prototype: versioned embedding inputs and model backend

[cuzzo]

Implement raw, normalized, and semantic-fact inputs; retrieval/clustering/similarity prompt spaces; EmbeddingGemma batching, 256/768 Matryoshka normalization, content-addressed reuse, and an offline deterministic test backend. Acceptance: golden inputs, finite normalized vectors, incompatible-space rejection, incremental reuse tests.

[cuzzo]

The versioned input spaces, batching, Matryoshka truncation/renormalization, content-addressed reuse, deterministic test backend, and Sentence Transformers adapter are implemented. PyTorch CPU and Sentence Transformers are installed. Keeping this open until google/embeddinggemma-300m is downloaded and encoded successfully; Hugging Face currently returns 401 for the gated model and requires accepted terms plus authentication.