Tell a real benchmark regression from noise

A benchmark regressed somewhere in the last 200 commits, and benchmark numbers are noisy. Before a bisecting agent names a commit, and the person behind it, it has to tell a real slowdown from noise.

Write classify_bench_regression(baseline_runs, candidate_runs, noise_band_pct): the rule your agent follows to compare two commits.

• baseline_runs, candidate_runs (list of float): repeated timings for the known-good commit and for the commit under test. Higher is slower.
• noise_band_pct (float): the declared noise, in percent.
• Return "regression", "clean", "inconclusive" or "insufficient data".

Rules, in order:

1. Fewer than 3 runs on either side: "insufficient data".
2. Take the median of each side. With an even count, the median is the average of the two middle values. If the baseline median is 0 or less: "insufficient data".
3. The band is the larger of noise_band_pct and the baseline's own spread, (max - min) / baseline median * 100.
4. The effect is (candidate median - baseline median) / baseline median * 100.
5. An effect at or below the band, a speedup included: "clean".
6. An effect of at least twice the band: "regression".
7. Otherwise: "inconclusive".