feat: expose residual_expression and post_join_filter on join builders

[nielspardon]

Completes the remaining work for #190: surface the two Expression fields spec #1044 added to HashJoinRel and MergeJoinRel — residual_expression and post_join_filter — on the ergonomic physical-join builders (hash_join / merge_join) and the corresponding DataFrame methods. They were previously only settable on the raw proto.

Binding semantics

Each expression is resolved against the schema the spec prescribes:

• residual_expression is "evaluated on each candidate key-match" (both rows present), so it binds against the combined left+right schema.
• post_join_filter is "applied to each output record after join-type-specific output formation... semantically equivalent to placing a FilterRel directly above this join", so it binds against the join output schema, reusing the existing _join_output_struct / join_output_names helpers.

Also fixes the logical join builder

The logical join builder previously bound post_join_filter against the combined schema too. For inner/outer/left/right/single/mark joins the output equals (or prefixes) the combined schema, so behavior is unchanged. For right_semi / right_anti (output is the right side only) a filter on a right column now uses the correct output-relative field index instead of a dangling combined-schema index, and a filter referencing a dropped side fails fast instead of emitting an invalid reference.

Notes

• The optional predicate/extension arguments are keyword-only, matching the logical join builder's convention.
• New tests cover inner + right_semi output binding, residual_expression combined binding, fail-fast on a dropped-side reference, keyword-only enforcement, and byte-identical DataFrame↔builder parity across inner/right_semi.

Closes #190

🤖 Generated with AI