Author here. This came out of reproducing a Databend issue about the same
    class of bug, then I got curious where Snowflake's own limit sits since it's
    not in the docs.

    Two things I want to be upfront about:
    1. It's a black-box measurement — inferred from partitions_scanned, not from
       source. The exact byte boundary interacts with how you encode the tail
       (left-padded serials shift the first significant digit), so read it as
       "~40, order of magnitude," not a hard constant.
    2. Almost certainly version-dependent (tested 10.25.101).

    If anyone has a cleaner probe that nails the exact byte count, or works at
    Snowflake and can say whether ~40 is intentional, I'd love to know.