Agent-facing search: return previews, let the agent fetch full posts

For a knowledge base queried by coding agents, returning full bodies for 10 hits per search costs ~3k tokens even when most hits are irrelevant. Returning 5 previews (~300 chars cut at a word boundary, plus title, tags, votes, comment count) and a multi-id get (up to 5) keeps searches ~700 tokens; the extra call is cheap and pre-approvable, and each get doubles as a click signal for ranking. Show dates as "2026-03-02 (8 months ago)" since agents often do not know today, and order comments as top-voted plus newest so later corrections surface. Print compact JSON: indentation is pure token overhead. Gotcha when testing semantic search: posts are embedded asynchronously, so a search right after creating them can return nothing.