fix(web): stop search drops major hubs like "I. P. Pavlova"

[i11v]

Problem

Searching a stop name surfaced the wrong stops and could omit the one the user wanted. Reported for "Pavlov": the hub I. P. Pavlova never appeared in the results.

Root cause

searchStops truncated its candidate list with slice(0, limit) (limit = 10). That list is expanded: every stop contributes a grouped row plus one row per platform, and for a prefix match all of those rows share the same score (100).

So a handful of multi-platform prefix matches consumed the entire budget before lower-tier matches of other stops were reached. For "Pavlov", the rural Pavlov* stops and their A/B/… platform rows filled all 10 slots, and the word-boundary match I. P. Pavlova (score 80, because "Pavlov" sits after the "I. P." initials) was cut off entirely — it was never in the list to begin with.

Verified against a verbatim copy of the matcher: the current matcher only returns genuine substring matches, so the other reported symptom (top results cluttered with geographically-nearby stops that share no letters with the query, e.g. Náměstí Bratří Synků) cannot be produced by this code — that comes from a stale, PWA-cached older bundle. Shipping this fix re-hashes the JS bundle and busts that cache.

Fix

Cap the search by distinct stops, not expanded platform rows, so one stop's platforms can't crowd out another stop. Each included stop still forwards all of its platform rows (needed for platform-scoped queries like "Anděl A"). The result budget is 12 distinct stops.

With the fix, "Pavlov" now includes I. P. Pavlova; when geolocation is granted, the existing proximity scorer promotes the nearby hub toward the top.

Changes

• packages/web/src/lib/matcher.ts — cap searchStops by distinct nodes instead of raw candidate rows.
• packages/web/src/components/search.tsx — request up to RESULT_LIMIT (12) distinct stops for a text query.
• packages/web/test/matcher.test.ts — regression test: a word-boundary match isn't evicted by another stop's grouped + per-platform rows.

Verification

• vitest run packages/web — 36 passed (incl. new regression test)
• oxlint, oxfmt --check, tsc -p packages/web — all clean

🤖 Generated with Claude Code

---

Generated by Claude Code

[github-actions]

🚀 Preview deployed: https://preview-29.tablo.run
_{fallback: https://tablo-pr-29.i11v.workers.dev}