v0.2.0-alpha.1

Changed

• Search the relevance-oriented core and long-tail discovery engine pools in
  parallel, with a 30% discovery share that rises to 50% for weak core results.
• Rank deduplicated media using SearXNG score, reciprocal rank, title coverage,
  engine consensus, media quality, host diversity, and balanced image/video
  scheduling.
• Use each Searchy user's language for SearXNG and add one English core fallback
  for weak non-English results. Operator-set LANGUAGE remains authoritative.
• Keep the proven SearXNG 2026.6.24-e3126b89e production digest after the
  2026.7.16-9f9c00819 candidate regressed DuckDuckGo Images under load. Pin an
  eight-second request budget and explicit core and discovery engine defaults.

Fixed

• Keep all contributing engines and positions during deduplication while
  preserving the primary engine used by existing analytics.
• Accept protocol-relative HTTPS media, upgrade known Bilibili HTTP hosts, and
  select the first safe HTTPS candidate instead of the first non-empty URL.
• Preserve valid width, height, and resolution metadata for quality ranking.

Privacy and operations

• Upgrade golang.org/x/text to v0.39.0 to fix reachable vulnerability
  GO-2026-5970 in Unicode normalization used through pgxpool.
• Keep query text out of logs, analytics, grid sessions, and benchmark output;
  cache and singleflight isolation now also include the resolved language.
• Add no database migration, proxy, Tor route, public SearXNG route, or new
  secret. SAFE_SEARCH=0, IMAGE_PROXY=false, and direct egress remain.
• Add a read-only 36-query synthetic benchmark covering exact, long-tail, and
  multilingual searches. Against the live candidate it improved average
  keyword relevance@10 by 18.6%, kept the exact-query group non-regressing,
  returned no empty cases, kept usable HTTPS at 100%, and measured a 1.5-second
  primary p95 even with one image core engine degraded.