Upgrade to Opus on distributional / reform questions in _select_chat_model

[vahid-ahmadi]

Symptom (live test 2026-05-28)

The decile-impact reform question routed to Haiku 4.5 (visible in stream: "model": "claude-haiku-4-5"). Haiku is fast per turn but semantically weaker on API derivation — it needed many guesses to converge on the reform shape, and ultimately never converged within the iteration budget. Opus would likely have nailed it in 2–4 iterations.

Diagnosis

_select_chat_model (backend/routes/chatbot.py:153) routes by message length / token count. That heuristic doesn't capture cognitive difficulty — distributional simulation questions are short to ask but reasoning-heavy.

Fix

Add a keyword/intent check to upgrade to Opus when the latest user message contains any of:

• decile, distributional, winners, losers
• reform, increase the X, change the X by
• 1pp, 1 percentage point, 1%, points
• marginal rate, effective rate (paired with by/from)
• poverty, inequality, gini

Tune the list against real usage rather than guessing — start small, watch which queries the agent fumbles, expand.

Constraints

• Don't change pricing visibility (the cost UI already shows per-message £).
• Cache hits still preferred where possible — keep the routing decision cheap (no extra LLM call for routing).
• Should compose with Plan mode and Charts mode (no override conflict).

Open questions

• Should charts_mode=True also bias toward Opus? Charts often imply distributional analysis. Probably yes.
• Should Plan mode always upgrade? Plan turns generate clarifying questions, low cognitive load — probably stay on Haiku.