docs(tool-registry): correct what tool_search returns `min_score` never leaves the model empty-handed by falling back to a single top match: scores are normalized against the best match, so the top result is always 1.0 and a cutoff only trims the tail. The escape hatch in the filter fires when *nothing* matched (max score 0.0) and returns the first `limit` tools in tool_id order. `keywords` do not simply boost the keyword channel — they replace the query string as the phrase that channel matches on. The search-tool cache is a bounded HashMap that evicts an arbitrary entry, not an LRU, and it backs the /v1/tools endpoints rather than per-model tool preparation.