Was finally trying to get to the bottom of
PHOENIX-1972. Ultimately stumbled onto the subtlety that the ResultSetResponse from ColumnsRequest isn't run through finagle.
I believe this ultimately causes the ColumnMetaData to be of Types.BIGINT and Rep.PRIMITIVE_LONG instead of Types.BIGINT and Rep.NUMBER.
This ultimately pushes us to the LongAccessor instead of the NumberAccessor (which is really the BigNumberAccessor) that correctly handles the "cast" from an Integer to a Long that the LongAccessor does not.
My only concern so far is that I haven't also been able to get this happen via hsqldb in a test. I've only been able to verify it via Phoenix.