I dug in a little more and discovered that we've been looking at
gfx_content_paint_time a little bit wrong. A better condition for checking "has content been painted" is
gfx_content_paint_time IS NOT NULL AND gfx_content_paint_time > 0. Narrowing the focus to nightlies since 2020-01-08 which have the fixes in bug 1604862), I see that over 20% have at least one ping where Fenix gfx painted, but didn't see a compositor: https://sql.telemetry.mozilla.org/queries/67608/source
I'm calling this condition of reporting paint times but not reporting a
gfx_status_compositor_last_seen as "weird".
So this is over 20% of weird clients from a little over 6% "weird" pings. Maybe this is a temporary condition that clients pass through?
Assuming I've not failed to write this analysis properly, it looks as though only 5 times in pings since 2020-01-08 (and in builds built since 2020-01-08) have ever stopped telling us what compositor they saw, if they had previously told us about one: https://sql.telemetry.mozilla.org/queries/67616/source
This suggests that the clients and pings telling us they don't know of any compositors are really sticking by their stories. This also suggests that "user"-lifetime metrics are working just fine as far as their "never cleared" behaviour is concerned.