Exact semantic-cache lookup before classifier fallback.
This plug runs early because a trusted cache hit is usually cheaper and more stable than asking a classifier or an LLM. Cache misses are recorded as local metadata so later plugs can include that context in traces and fallback logs.