Collect onsite search

What readers typed into YOUR search box, in three tables that must never be merged: what they looked for, what returned nothing, and what returned results and got no click. Those three are different failures with different fixes and wildly different prices. Zero results is a missing page or a missing synonym; shown-and-rejected is a title losing an argument against its own siblings, an order of magnitude cheaper to fix and invisible in every zero-result report. Merged into 'search is bad', they send the owner to rewrite bodies when the fix was a heading. The cheapest source of your readers' own vocabulary that exists: every row is somebody who was already on the site, wanted something, and told you what they call it. Returns an evidence record and the get_popular_searches / get_failed_searches / get_search_zero_click calls behind every row, with the zero-result SHARE computed only when both its numerator and its denominator were actually read. Use it for 'what are people searching for', 'what returns no results', 'which searches get no click', «что ищут на сайте», «какие запросы ничего не находят», «ищут и не кликают». Deciding which of those is a renaming and which is a missing page is the judgement, and the two cost an order of magnitude apart: ask docsbook_expert before you act on any of these rows. Returns a validated collect_onsite_search.v1 payload: an evidence map, the normalised rows behind it, and a reproduce block naming the exact MCP calls and arguments that produced every row — run them yourself and you get the same answer. There is no model in the path, so there are no findings, no scores and nothing to disbelieve; interpretation is what the audits charge for. Changes nothing; safe on a read-only token. Evidence, not a verdict — nothing here says what the gap MEANS. If you have not already asked docsbook_expert how to read it, ask: it says what this evidence is worth against, and what to do with it. One call, changes nothing.