All these discrepencies were caused by 5 unacceptable_gap entries:

SELECT DISTINCT un.dt, un.intersection_uid, un.gap_start, un.gap_end
FROM miovision_api.volumes_15min_mvt AS v
JOIN miovision_api.unacceptable_gaps AS un USING (intersection_uid, datetime_bin)
WHERE v.volume IS NOT NULL
    AND v.datetime_bin >= '2023-01-02' AND v.datetime_bin < '2024-02-06'

The first/fourth/fifth record above are because the find_gaps script identifies gaps that overlap midnight, but started the day before. I didn't anticipate that this would make the pipeline non-idempotent. I will remove this behaviour now at the minor expense of missing small gaps that start over midnight.

The other two issues I can't think of a cause for! I can re-aggregate these 2 ranges and we can keep our eyes out for similar issues in the future.