I think the change to derive `clause_relids` from the child EMs makes sense, especially for the lateral-reference case described in the patch. One thing I may be missing, though: after replacing `rinfo->clause_relids`, should we also update `rinfo->num_base_rels`? `make_restrictinfo()` computes `num_base_rels` from the original `clause_relids`, but this patch subsequently replaces `clause_relids` for child clauses. In cases involving lateral references and nulling outer joins, the two sets can be different, so `num_base_rels` could potentially become stale. Would it be safer to recompute it from the new `clause_relids` here, using the same definition as in `make_restrictinfo()`? I am not sure whether there is an invariant elsewhere that makes `num_base_rels` unaffected in these cases. I t might also be useful to add a regression case specifically covering this metadata, in addition to the existing plan/execution coverage.
