Exadata Excessive single block reads on UNDO long after COMMITs

DB/Grid version 19.7; Exadata version 19.2.15.0.0.

Getting seemingly random offloading failures that instead do single block cell reads even when every table in the query had a COMMIT on it at least 1 hour earlier. Trying to narrow down reason for UNDO activity; still not sure how it can be because of CRs from UNDO. Happens almost every night on a random query.

ASH report shows UNDO as the culprit. When the query is run a few hours later, takes seconds.

Looking at the Top Events, I can find the cell number and ASM disk number from P1 and P2, but that doesn't get me much closer to a root cause: