We need at the very least to document this problem and possibly introduce diagnostics for this case. We might put a flag somewhere in the output file that specifies that it has been produced by a multi-thread Snapshot and therefore entries are scrambled w.r.t. the original dataset. TTree or RDataFrame could then check that flag and prompt a warning if things look fishy.
One case in which this might result in undesired warnings is when two files that have been produced by two multi-thread Snapshots in the same event loop are used as friends of each other. To be checked if the on-disk dataset PID can help recognize this case.