Gate media reranking behind bounded cascade

[brianmeyer]

Summary

• Keep media reranking opt-in, then add an explicit ambiguity gate so expensive VL reranking is skipped when RRF has a clear winner.
• Enforce separate media-query and media-result top-K caps before reranker calls.
• Prefilter empty media candidates that lack text or a resolvable media path so they do not consume reranker budget.
• Document the new media rerank controls in env var and release docs.

Research grounding

• SentenceTransformers retrieve-and-rerank guidance: https://sbert.net/examples/sentence_transformer/applications/retrieve_rerank/README.html
• SentenceTransformers quickstart top-k cross-encoder notes: https://github.com/huggingface/sentence-transformers/blob/main/docs/quickstart.rst
• Vespa phased ranking rerank-count/top-k design: https://docs.vespa.ai/en/ranking/phased-ranking.html

Tests

• python3 -m pytest -q tests/test_search_pipeline.py tests/test_cross_modal_pipeline.py
• python3 -m pytest -q