Add cost estimation and context budget forecasting

[yanpgwang]

Summary

Two new features that make context-profiler reports immediately actionable:

1. Cost Estimation (pricing.py)

Adds $/request cost estimates based on model pricing tables.

• Supports Claude (Opus, Sonnet, Haiku), GPT-4o, GPT-4o-mini, GPT-4-turbo
• Loose model name matching (e.g., "claude-3-5-sonnet-20240620" → Claude 3.5 Sonnet)
• Cost appears in token_counter summary and diagnosis JSON
• Gracefully skips if model is unknown

2. Budget Forecasting (session_insights.py)

Predicts when a session will overflow the context window.

• Linear extrapolation from turn-over-turn token growth
• Model-aware context window sizes (Claude 200K, GPT-4o 128K)
• New budget_forecast field in session_insights:
  • growth_rate_per_turn
  • current_utilization
  • estimated_overflow_turn
  • context_window_tokens
• New diagnostic issue: CONTEXT_OVERFLOW_RISK
  • Triggers when overflow is projected within 2x current turn count
  • Severity: critical (>80% utilized), warning (>50%)

Testing

• 92 tests passing
• New test file: tests/test_budget_forecast.py (20 test cases)
• All existing tests unaffected (no regressions)

Example output

{
  "budget_forecast": {
    "growth_rate_per_turn": 10000.0,
    "current_utilization": 0.3906,
    "estimated_overflow_turn": 12,
    "context_window_tokens": 128000,
    "model": "gpt-4o"
  },
  "cost": {
    "estimated_input_cost_usd": 0.001978,
    "estimated_model": "GPT-4o"
  }
}