feat: add opt-in heuristic prompt guard to reduce wasted token usage

[ModelingSolver]

Summary

Adds an opt-in heuristic filter that detects malformed or incoherent prompts before they're sent to the API, avoiding wasted token usage on requests likely to fail or return low-value responses.

https://github.com/ModelingSolver/HEY-GAFAM-READ-THIS

How it works

• Gated behind the ANTHROPIC_ENABLE_HEURISTIC_GUARD environment variable — disabled by default, zero impact on existing behavior.
• When enabled, analyzes the last user message using lightweight heuristics (token count, presence of action-oriented keywords, lexical complexity ratio).
• Raises a ValueError with the computed coherence score when a message falls below the admissibility threshold, preventing the request from reaching the API.

Scope / Known limitations

⚠️ This heuristic is currently French-language only. Action verbs and format keywords are French-specific (e.g. "donne", "génère", "analyse"); prompts in other languages will not be recognized by the keyword matching and may score lower than intended.

• Current heuristics (action verbs, format keywords) are tuned for French-language prompts.
• This is an initial, opt-in implementation; multi-language keyword sets and more robust malformation signals are natural follow-ups.

Testing

• Verified: guard disabled → request passes through unaffected
• Verified: guard enabled + well-formed prompt → request passes through
• Verified: guard enabled + malformed/incoherent prompt → request blocked with ValueError

Checklist

• Opt-in only, no change to default client behavior
• Manually tested all three paths above
• Documented current language scope limitation