Implement convergence speed analysis

[evanvolgas]

Implement convergence speed analysis for publication results section.

Metrics to Implement:

1. Time-to-X% oracle performance: How many queries to reach 80%, 90%, 95% of oracle quality?
2. Convergence rate: Quality improvement per query over time
3. Learning curve comparison: Plot quality over time for all algorithms
4. Convergence stability: Variance in learning trajectory

Implementation Location: conduit_bench/analysis/convergence_analysis.py

Required Functions:

def compute_time_to_threshold(
    results: BenchmarkResults,
    oracle_results: BenchmarkResults,
    thresholds: List[float] = [0.8, 0.9, 0.95]
) -> Dict[str, Dict[float, int]]:
    """Compute number of queries to reach X% of oracle performance."""
    
def compute_learning_curves(results: BenchmarkResults) -> pd.DataFrame:
    """Compute rolling average quality over time."""
    
def compute_convergence_stability(results: BenchmarkResults, window: int = 100) -> Dict[str, float]:
    """Compute variance in learning trajectory."""

Visualization Requirements:

• Learning curves (quality over time, all algorithms)
• Time-to-threshold comparison (bar chart)
• Convergence stability comparison

Success Criteria:

• All convergence metrics implemented and tested
• Integration with existing analyze CLI command
• Visualization functions ready
• Statistical significance testing (bootstrap confidence intervals)
• Documentation and examples added

Timeline: 4 hours

Dependencies: Issue #33 (experiment results)

Publication Impact: Addresses "how fast do bandits learn?" research question

Key Insight: LinUCB may be slower to converge (high-dimensional context) but eventually outperform, or UCB1 may converge faster but plateau lower.