Add search, sort, filter, and windowing to the analysis catalog - #253

Closes #231.

The catalog listed every analysis in one fixed order with no way to narrow it, so a draft of any size was navigable only by scrolling. The query core (#192, PR #211) already backed all of this — applyCatalogQuery, deriveFacets, CatalogSort, and CatalogFilters were built and tested. This is the UI that varies them: the panel's fixed query literal becomes state the controls drive, and nothing in the query core changed.

Search, sort, and filter state is ephemeral useState inside the panel. The panel is mounted only while it is open, so closing it clears the query by construction rather than by a reset effect — a filter that survived a reload would leave rows missing with nothing on screen saying why. Open/closed and width stay tab-scoped in useWebViewState.

Filters sit behind one control

A panel narrow enough to need filtering should not itself be filled with filter controls, so all of them live in a popover whose trigger reports how many are active.

All four groups ship. Against today's data only books ever raises a control: nothing writes part of speech, confidence, or features yet, so deriveFacets correctly yields nothing for them and no control appears. Used nowhere ships too, and matches nothing until PT9 import (#150) or the catalog's own delete/merge paths land — detachTokenAnalysisLink drops a payload with its last link, so no current write path can produce a zero-usage row. Both are the outcome #231 decided on and both are tested for, not gaps.

Facets are derived from every row the draft holds rather than from the rows a filter left standing. A facet judged against its own selection's survivors would collapse to that selection, leaving nothing on screen to widen it back by.

Carrying no value is a choice of its own

CatalogFacets lists the absent value as undefined, which is what lets a reader ask which analyses are still missing a field as readily as which carry a given value. The platform MultiSelectComboBox speaks strings alone, so that choice needs a spelling: \u0000untagged, a leading NUL being one no part of speech, confidence level, or feature value can collide with. Values are read back through a map rather than compared against the sentinel, so undefined is recovered as the choice it is. Worth a look — it is the one place the control's vocabulary and the filter's diverge.

Books is the exception: a usage names a book by construction, so that facet never offers an untagged choice, and the selection is filtered before it reaches CatalogFilters.

Windowing

useRowWindow mounts a growing leading slice of the listing and extends by a chunk each time an end-of-list sentinel comes within reach. Grow-only and anchored to nothing — it never culls from the top and never adjusts the scroll position. Deliberately not useSegmentWindow: that hook is anchored around a scripture reference, and a row list has no counterpart to hold still, so it needs none of that geometry bookkeeping.

Two details worth the reviewer's attention:

• The window starts over when rows is a different array, adjusted during the render that first sees the new listing rather than in an effect afterwards — an effect would let one frame paint the new rows at the old, grown count before shrinking back. Keyed on array identity rather than length, because a query can narrow a listing to a different set of rows of the same size.
• The observer re-subscribes on every count change, not just on the elements'. An IntersectionObserver reports only intersection transitions: after an extend the sentinel node is unchanged and may still sit inside the arming margin, where a stale observer would stay silent however far the reader scrolls. A fresh observer re-delivers the current state, extending once per delivery until the sentinel is pushed clear.

Not in the issue: a listing narrowed to nothing says so

Reusing "No analyses recorded yet" for a query that matched nothing would tell readers their draft is empty when they have merely mistyped, and send them looking for lost work. There is now a second message for that case, and both go through the platform EmptyState rather than a hand-rolled paragraph.

Search semantics are unchanged and accepted as-is

applyCatalogQuery matches the whole trimmed, folded query as a single substring against a per-row blob. So a multi-word query never matches across fields, there is no match weighting, and ordering is the chosen sort key alone. The placeholder promises "Search forms and glosses" and nothing more. Multi-term search, relevance ranking, and highlighting are query-core work, not UI work, and separate issues if wanted.

Mocks

Stubs the platform SearchBar, Select, MultiSelectComboBox, and EmptyState, each documenting where it diverges from the component it stands in for — notably that MultiSelectComboBox resolves an entry by label as the real component's own select handler does, so a stub test cannot pass on a collision the real component would drop.

Testing

Covers all eight behaviors #231 lists. Full suite passing, 100% coverage, lint clean.

This change is 

Summary by CodeRabbit

• New Features

  • Added an Analysis Catalog panel with searchable, sortable, and filterable analysis entries.
  • Added filtering by book, part of speech, confidence, features, gloss availability, and usage.
  • Added expandable analysis details, usage navigation, selection persistence, and localized empty states.
  • Added virtualized loading for improved performance with large catalogs.
  • Added English localization for catalog controls, labels, messages, and Save As states.

• Tests

  • Added comprehensive coverage for catalog interactions, filtering, sorting, navigation, localization, and empty states.