Joseph Doherty
8d51380ff1
A reverse read (start > end) with a small NumValuesPerNode resumed through the tie-cluster path, which drained the cluster and stepped FORWARD to T + 1 tick - outside a reverse window - so the read ended with GoodNoData after one sample. SliceTieCluster now steps down to T - 1 tick for a reverse read (terminal at the excluded end), and the node manager slices the over-fetched cluster in the read's own direction so the skip count is not applied from the wrong end. Forward paging is unchanged. Claude-Session: https://claude.ai/code/session_01Ef2ozxS7QPanGT8uoATbTe