Read a split

A read of one Table version, sent whole with every plan and read request. columns selects top-level columns, or all of them when absent; projection computes the output columns instead. A filter clause that the service cannot enforce exactly fails the scan before any rows are read. limit caps the rows each split returns, in stored order. An emit returns a top-k or aggregates of the output rows instead of the rows themselves. rowRange selects half-open stored positions in this exact version, before filtering and deletion removal, concatenating fragments in stored order. Positions are not stable across rewrites. Empty ranges return no rows; reversed ranges are invalid. rowIdColumn adds a last output column of that name holding each row's row ID: the Table's stable identity of the row, a uint64 that survives rewrites and compaction and is never reused. rowIds keeps only rows whose row IDs fall in its ranges: planning prunes files whose row IDs cannot match, and reads enforce it exactly. A top-k can return and order by the row ID column. Aggregates take no rowIdColumn.

Stored rows of one file of the scanned version. Stored rows include rows that a later version deleted; a read skips them. Resume a split by raising rows.start to the last row a read reported, and share one between readers by dividing rows.