Benchmark kNN discovery sensitivity to k and time-series length #55

Follow-up to #14 / #48 and the kNN estimator cleanup in #53.

The estimator-level experiments suggest that the Chebyshev kNN CMI estimate improves substantially as the sample size increases, but we have not yet characterized how that translates to causal-discovery performance.

I’d like to benchmark the discovery path across:

• multiple k values,
• multiple time-series lengths T,
• multiple random seeds,
• the Chebyshev kNN estimator used by the current API.

I’ll report graph-level metrics such as TP/FP/FN, precision/recall/F1, and runtime, and check whether recovery improves consistently as T increases or whether particular k values behave poorly.

I’ll base the experiments on the final kNN implementation once #53 is resolved, so the benchmark reflects the version that will actually ship.