builders.plan.join and the physical equi-joins omit subtrees=, so a post_join_filter over a cached input raises #275

join() (src/substrait/builders/plan.py:695) and _physical_equi_join() (plan.py:1364) call _join_output_struct(...) without passing subtrees=, so the output schema they bind post_join_filter against is re-inferred from the raw root input with no shared-subtree scope. A ReferenceRel in either input then cannot resolve and inference raises.

lateral_join (plan.py:797) already does this correctly, reusing the left_ns / right_ns it inferred ~30 lines above via _join_struct_from_schemas(type_name, left_ns.struct, right_ns.struct).

Reproducer

import substrait.type_pb2 as stt
from substrait.builders.type import i64, string
from substrait.dataframe import col, lit, read_named_table


def schema(names, types):
    return stt.NamedStruct(
        names=names,
        struct=stt.Type.Struct(types=types, nullability=stt.Type.NULLABILITY_REQUIRED),
    )


left = read_named_table("a", schema(["store_id", "k"], [i64(), string()]))
right = read_named_table("b", schema(["fk", "amount"], [i64(), i64()]))

# Without post_filter this builds fine.
left.cache().join(right, on=col("store_id") == col("fk"), how="left").to_plan()
print("no post_filter: OK")

# With post_filter it raises.
left.cache().join(
    right, on=col("store_id") == col("fk"), how="left", post_filter=col("amount") > lit(5)
).to_plan()

no post_filter: OK
Traceback (most recent call last):
  ...
  File "src/substrait/builders/plan.py", line 695, in resolve
    struct=_join_output_struct(
  File "src/substrait/type_inference.py", line 593, in _join_output_struct
    left = infer_rel_schema(left_rel, registry=registry, subtrees=subtrees)
  File "src/substrait/type_inference.py", line 893, in infer_rel_schema
    raise Exception(
Exception: ReferenceRel subtree_ordinal 0 is out of range (0 shared subtree(s) in scope)

.hash_join(...) and .merge_join(...) fail the same way. Dropping post_filter succeeds, because nothing then re-enters _join_output_struct.

Measured on main at afa3c52. Pre-existing, not a regression — it reproduces identically on 45c8ca7.

Passing the already-inferred structs, as lateral_join does, both fixes the crash and drops a redundant recursive re-inference of both subtrees per join, which is the same quadratic re-inference tracked in #207. Shared-subplan support was added in #211.

🤖 Filed with AI assistance as a follow-up to review of #272.