List Benchable Documents

Guideline documents already in the platform, so a run can be started without uploading.

Deliberately not scoped to the caller's company: tuning a prompt means running it over the guideline documents that exist, whoever they belong to. That is only acceptable because this router is not mounted outside DEVELOPMENT — see backend/api/internal/__init__.py.

Restricted in SQL to the types is_guideline_document accepts, because the alternative is a list where limit is spent on vendor contracts that happen to sort first and the guideline documents never appear. A guideline stored under a type that hides it here is still benchable by uploading it.

No text search, deliberately. The picker is a Tremor SearchSelect, which always filters its own children client-side with a literal substring test on the rendered label; a server-side ILIKE on top of that is a second filter with different rules, and the two disagreed on any query with edge whitespace or a LIKE metacharacter — the server matched rows the client then hid, showing no dropdown and no reason. The whole (bounded) list goes to the client and it filters once.