fix(chat): broaden over-specific prediction-market queries on zero results

[VickyXAI]

Problem

2026-07-17 production log audit: a user asked about betting on "the world cup final game ... spain against argentina". Franklin's search_prediction_markets passed the model's verbose query verbatim to Predexon, got 0 results, and the model hallucinated ("the next World Cup will be in 2026").

Live probes confirmed the root cause — Predexon /markets/search is strict keyword AND-matching over market titles:

• q='World Cup final game' -> 0 results
• q='World Cup final Spain vs Argentina' -> 0 results
• q='World Cup' -> 10 markets, incl. Polymarket 'Will Spain win the 2026 FIFA World Cup?' ($58M vol) and 'Will Argentina win?' ($85M vol)

At least 10 of 124 franklin-try conversations show this zero-result miss pattern on over-specific queries.

Fix

1. Auto-broadening retry (src/hooks/use-franklin-chat.ts): when the first search returns 0 results, retry once with the query stripped of qualifier/filler words (final, game, odds, vs, winner, betting, ...) down to its 1-3 strongest entity keywords. If the retry finds markets, return them with a note naming the broadened query so the model knows what actually matched. All-filler queries ("hottest markets right now") broaden to nothing and skip the retry, so a paid retry only fires when there are real entity keywords to search.
2. Model guidance: tool schema description + FRANKLIN_TOOLS_PROMPT now state up front that search is strict keyword AND-matching over market titles — query with 1-3 entity keywords (e.g. "World Cup", "Fed rate cut"), never full sentences or qualifiers.

Broadening behaviour on audit queries:

• "World Cup final game" -> "World Cup"
• "World Cup final Spain vs Argentina" -> "World Cup Spain"
• "Fed rate cut odds" -> "Fed rate cut"
• "hottest markets right now" -> "" (no retry)

Verification

• npx tsc --noEmit clean
• npx eslint clean on both changed files
• No CJK characters introduced (repo is English-only)