[selmanozleyen]

A dataset split across files can keep each side's rep only where it lives, e.g. obsm/X_src only in the controls' file and obsm/X_tgt only in the perturbed one (a quadratic coupling). Before, every file had to carry every rep a stream asks for, even files the stream never selects rows from.

• Streamed reps: Source.rep(loc, needed) gives a file that lacks the rep and has no selected rows a never-read placeholder (an unwritten in-memory zarr array or empty CSR group when backed, a broadcast zero view or empty CSR matrix in memory), so annbatch still gets one backing per file. A missing rep in a selected file raises.
• In-memory streams and eval: _read_rows reads the rep only from files that hold selected rows.
• load_backed_adata / open_source skip a key a file does not have, leaving the decision to Source.rep.

[selmanozleyen]

Superseded: scfit already supports this through separate sources (a link stream on its own source_key binds to the primary via match_on). sckitflow will route each file to the target and/or control source by role, so no placeholder is needed here.