I was not looking at the timestamps for my estimate. If average block time is 10min (600s) , then a quota of roughly 1% of empty blocks found points to an average "empty block mining time" of 6s. Rough guesstimate, of course, but still way beyond the actual "assembly time" for a full block header.
You can't guess these things you have to look at the actual values, and the only value we have is the timestamp inside each block header.
Here are past 4 "empty" blocks and their previous block with their timestamps, the difference is 0 or 1 second:
672200 2021-02-25 23:56
672199 2021-02-25 23:55
672183 2021-02-25 21:38
672182 2021-02-25 21:38
671897 2021-02-24 01:57
671896 2021-02-24 01:57
671775 2021-02-23 02:46
671774 2021-02-23 02:45