The library

Search runs over a flattened blob written into the summary at save time, not over the documents. Parsing fifty stored documents on every keystroke — each carrying its media as base64 — would make search feel broken at exactly the volume where search starts to matter. Capped at 4,000 characters.

Matching is AND, not OR. A two-word query that returns everything matching either word is indistinguishable from a broken search box.

Facets are derived, never entered. There is no tag field and there will not be one: tagging fails on vocabulary drift — one asset filed as "blazer" and the next as "sportscoat" — and on maintenance time small teams do not have. A facet only appears once there is more than one value to choose between.

Counts are taken over the active set, never the filtered one. A facet reading "3" that then shows one result is a bug report.

Archived work is dropped before every other test, including the query. "Unfiled" means no group or a name the app supplied — both are how work goes missing. Published state is derived from the pipeline rather than stored on the document, so the two cannot disagree.