A Spearman over exact zeros is tie-degenerate, and argsort-of-argsort silently picks one value from a range

def avgrank(a):                     # average ranks, ties shared
    a = np.asarray(a, float)
    o = np.argsort(a, kind='mergesort')
    r = np.empty(len(a), float); i = 0
    while i < len(a):
        j = i
        while j + 1 < len(a) and a[o[j+1]] == a[o[i]]: j += 1
        r[o[i:j+1]] = (i + j) / 2.0; i = j + 1
    return r

scipy.stats.spearmanr and scipy.stats.rankdata already do this; the hand-rolled argsort(argsort(...)) does not, and the two differ only when ties are present, which is why it passes every smoke test.

Three rules follow:

1. Count the exact ties and print the count beside the rho. n_exact_zero = 3 of 8 is the warning label.
2. When ties are present, report the tie-corrected value and the arbitrary-tie-break range. The range is cheap to enumerate at these n and it is the honest error bar on the instrument, separate from the sampling error bar.
3. Do not read a tie-broken rho as more extreme evidence. Here the arbitrary choice moved the value by 0.19 in rho, and in the favourable direction.

It joins the family where the instrument supplies an answer the data did not:

• a fallback regex that finds an accession inside a subunit map and reports it as the entry's own (rule 69b.3);
• a flat control reaching Spearman -0.619 on values that are all <= 0.09x, which is why this beat put an absolute floor on every locus entering a correlation;
• and this: a tie-break rule inventing an ordering among identical zeros.

In all three the code runs, nothing errors, and the number is a property of the procedure.