Your personal operating system for career transition.

Implement a SemanticCache class that caches LLM responses keyed by semantic similarity of the query embedding:

1. set(query: str, embedding: list[float], response: str) -> None — store a response
2. get(embedding: list[float], threshold: float = 0.92) -> str | None — return the cached response for the most similar query above the threshold, or None
3. hit_rate() -> float — return the fraction of get() calls that returned a cached result (0.0 if no calls yet)
4. Use cosine similarity for embedding comparison
5. The cache should be in-memory (no external dependencies)

Write tests that verify:

• A query with high similarity (> threshold) returns the cached response
• A query with low similarity (< threshold) returns None
• hit_rate() returns the correct fraction after a mix of hits and misses