The same comparison, split by signup week, so a trend is visible rather than one pooled figure.

A single number hides the two things worth seeing: whether the gap is opening or closing, and whether either arm is drifting for reasons that have nothing to do with the programme (a change in acquisition mix moves both arms together, and pooling makes that look like a stable effect).

Same eligibility as the pooled readout, for the same reasons: durable arm labels rather than a recomputed hash, and closed windows only. Weeks with an empty arm are returned with nulls rather than dropped, because a gap in the series is information - it says the holdout was not running then.