paged-palette

The dashboard search palette keeps node/session ranking local but obtains bounded Issue/Eval planes from paged-review instead of graph row arrays.

The one graph/session search palette still ranks node prose from the lite corpus and live session identity locally. Its Issue and scenario planes are demand data: while the palette is open, the debounced text drives page 1 of /api/issues and current /api/evals; each contributes at most 25 matching rows and the palette interleaves them with node/session hits exactly as before. /api/specs/lite contains node prose only, never scenario declarations that recreate the Eval list. Opening the palette therefore pays bounded review rows, while never opening it pays none.

Node and session planes still use the shared local lexical ranker. Issue and scenario rows have already passed the shared server matcher, so the palette preserves each response's stable order instead of applying corpus IDF a second time to a match-only page. In particular, an exact query returning one review row remains one visible hit; it cannot collapse to zero because every document in that filtered slice contains the query.

Every selectable palette row is a real matched node, session, Issue, or scenario. The review responses' pagination totals are transport metadata, not searchable entities: the palette never appends synthetic "showing N of M" / see-all rows to its ranked results. The canonical Issues and Evals lists remain available through the dashboard's normal navigation rather than masquerading as matches.

An Issue hit routes to its detail. A measured Eval hit routes to its detail; a blind scenario routes to the canonical node-filtered Evals list because it has no result detail. Plane boost, keyboard ownership, and selection routing remain session-search's single shared behavior.