Warn about repeated triplets only when chance cannot explain them

[snarles]

load_data now issues warnings when there are duplicated rows in excess of what would be be typically seen in uniformly sampled triplets.

The check reports the observed repeat count, the count expected under an explicit null, and the probability of seeing at least that many, and warns only when that probability is below 0.2.

Null: the N trials are drawn uniformly from the K = C(M,3) possible triplets, with trial identity the unordered set {i,j,k}. The statistic is W, the number of colliding pairs. The collision indicators are pairwise uncorrelated, so W matches Binomial(C(N,2), 1/K) in both moments exactly, which is what makes the binomial reference right rather than merely convenient. One-sided: a deficit of repeats just means sampling without replacement, which main_tripletize does deliberately.

The expected counts are written with log1p/expm1 throughout. This is not cosmetic: the textbook form of the expected number of repeated triplets cancels catastrophically, returning a negative count at M=5000/N=100 and a value five orders of magnitude too large at M=20000/N=50. The p-value likewise underflows float64 well before the evidence stops being readable, and binom.logsf returns -inf there, so a log-scale Poisson tail keeps the reported log10 p finite.

Data is never altered or de-duplicated, and the check never raises.