feat: faceted search for the Store — topic, language, author, license

[0xLeif]

Summary

Make it easy to discover plugins/templates by what they actually are, not just by name.

The Store page now has a left filter rail of topics, languages, authors, and licenses with live counts computed from the current result set, plus inline pivots — click a topic tag, language stat, or author name on any card to filter by it instantly.

Backend (src/github.ts, src/api.ts)

• New BrowseFilters shape (q, language, topics[], owner, license) replacing the string-only query argument.
• buildQualifiers() composes a GitHub search string from filters, drops the reserved base markers (fledge-plugin / fledge-template / fledge) so users can't override them, and quotes languages with whitespace.
• computeFacets() aggregates topics / languages / owners / licenses across the result set — deduped per repo, sorted by count, NOASSERTION licenses skipped.
• /api/github/browse accepts q, language, owner, license, and repeatable topic query params; responds with { items, total, facets, applied }.
• per_page bumped from 30 → 50 for richer facets.

Frontend (public/)

• Sticky left filter rail (Topic / Language / Author / License) with counts and a "Show N more" expander per group.
• Removable active-filter chips above the grid; "Clear all" wipes filters in one click.
• Topic tags on cards are buttons (toggle that topic). Language and author on each card pivot to that filter.
• Search input is debounced (300 ms); Enter forces immediate run, Esc clears.
• Result count rendered top-right.
• Responsive: rail collapses above the grid on narrow viewports.

Tests

12 new unit tests in tests/github.test.ts covering:

• buildQualifiers: empty, multi-topic, reserved-topic dropping, whitespace quoting, falsy filter handling.
• computeFacets: sorting, per-repo dedupe, NOASSERTION skip, empty input.

bun test → 17 pass / 0 fail. bun run typecheck clean.

Test plan

• bun test and bun run typecheck both pass on CI
• Open / → Store. Confirm filter rail loads with counts.
• Type into search — observe debounced refresh; Esc clears.
• Click a topic facet → grid narrows; chip appears; click chip to remove.
• Click a topic tag on a card → same behaviour.
• Click a language stat / author name on a card → grid pivots.
• Stack two filters (e.g. topic:swift + language:Swift).
• "Clear all" resets everything.
• Switch category tabs (All / Plugins / Templates) — filters persist, results refresh.
• Mobile (<= 768px): rail stacks above results.

🤖 Generated with Claude Code

[corvid-agent]

👋 CorvidAgent is looking into this.

[corvid-agent]

Solid work — clean refactor with good test coverage. Approving.

What I liked:

• buildQualifiers() / computeFacets() are well-separated and independently testable. The generic tally() helper with per-repo dedup is a nice pattern.
• Reserved topic stripping is applied consistently (both in filter construction and facet output).
• 12 tests cover the important edge cases: empty/falsy filters, whitespace quoting, NOASSERTION skip, intra-repo deduplication.
• Search debounce (300ms) with Enter bypass and Esc clear is good UX.
• setSingleFilter toggling (same value clears) gives the filter UI consistent behavior.
• Responsive collapse of the filter rail at <= 768px is handled cleanly.

Minor observations (non-blocking):

• quoteIfNeeded() wraps values with whitespace in double quotes but doesn't escape embedded quotes. Unlikely to matter with GitHub's language taxonomy, but worth a comment or guard if the filter surface expands later.
• Facet counts reflect only the first 50 results (per_page=50), not the full result set. The PR description is honest about this ("computed from the current result set"), but users might find it surprising when a filter shows count=1 for something that actually appears in many repos beyond the first page. If this becomes a problem, a follow-up could fetch total_count from the search response to indicate incomplete facets.
• The applied field in the API response is useful for debugging but isn't consumed by the frontend — could be dropped to keep the response lean, or documented as intentional.

Clean code, good tests, sensible UX — ship it.