bug(core): the embedding cache is keyed by the embedder's name, so it serves vectors of another model #1536

Problem

CachingEmbedder (embed/cache.rs) keys every entry on inner.name(), and the names are constants: "openai", "ollama", "embeddinggemma-q4"; FallbackEmbedder reports "fallback(onnx→cloud)". The key ignores the model, the dimensions and the endpoint. So switching text-embedding-3-small to -large, changing dims, or pointing at another endpoint serves the vector the previous configuration produced, silently, and a cloud vector produced through a fallback is cached under the primary's name. There is no dimension check on read.

Reproduced

A mock embeddings endpoint that records requests, same text, same UTEKE_HOME, [embedding] backend = "openai" with model = "model-a" and then model = "model-b": the second uteke remember never reaches the endpoint (one request in total) and stores the model-a vector.

Direction

The embedder owns an identity (what produced the vector: model, dimensions, endpoint) and the cache is keyed by it. Existing cache rows were written under the old key and simply stop being hit; the cache is a cache.

Behaviour impact

Local only, no wire change. After the upgrade the first embeddings are recomputed once. Recall results stop being wrong after a model switch.

Acceptance

• A cache entry is never returned for a different model, dimension or endpoint (test with two fake embedders that differ only in one of them).
• A vector read from the cache whose length is not the embedder's dimension is ignored.
• A fallback result is not cached under the primary's identity.
• Reproduced with the mock endpoint before and after (the second run reaches the endpoint).

Part of the embedding deepening epic.