feat(cli): add refusal detection and anti-refusal prompt hardening

[markm39]

Summary

Prevents the LLM from refusing to attempt hard problems (FrontierMath, IMO, competitions).

Prompt hardening:

• System prompt: "MANDATORY: You MUST attempt every proof. NEVER refuse."
• API instructions: "NEVER refuse to attempt a proof"
• Branch agent prompts (Planner, Prover): explicit anti-refusal language

Refusal detection + forced retry:

• New refusal.rs module with two-tier pattern matching:
  • Strong signals (always trigger): "open problem", "FrontierMath", "cannot be proven", etc.
  • Weak signals (only trigger with no tool calls + no Lean code): "I apologize", "too difficult"
• When detected, injects a system message forcing the model to write code
• Max 2 retries before accepting (escalating from constructive to forceful)
• Applied at both main agentic loop and branch loop break points

Test plan

• cargo clippy --workspace -- -D warnings
• cargo test --workspace (32 tests pass, including 7 new refusal tests)
• Unit tests: strong refusal, weak refusal, tools override, lean code override, normal text