[MamadouSDiallo]

mean, total, prop, ratio, quantile, median take overall=True (with by=): the result is the by= rows plus the by=None rows, exactly the two calls' numbers, listed last, null in the domain columns, printed as "All". Lists (several y or p) get it per member. Serialized payloads mark the rows with by=None.

Covariance: the overall rows come from their own call, so the combined result is marked unfilled and contrast() refuses with a pointer to the by= result (same rule as multi-variable lists). Joint domain/overall covariance would need kernel support.

overall=True without by= raises OVERALL_WITHOUT_BY.

[MamadouSDiallo]

Closing: two calls glued together, with no joint covariance. If an overall row comes back it should be one kernel pass with the overall as an extra domain, so domain-vs-overall contrasts work.