Avoid null-restrict evaluation for predicates that reference non-join columns in PushDownFilter

[kosiew]

Which issue does this PR close?

• Part of perf: push_down_filter is pathologically slow for some plans #20002

Rationale for this change

PushDownFilter can spend a disproportionate amount of planning time inferring predicates across joins. One expensive path is is_restrict_null_predicate, which falls back to compiling and evaluating the predicate against a null-filled schema to decide whether a predicate is null-rejecting.

For predicates that reference columns outside the join-key set, that evaluation cannot succeed with the synthetic null schema built for join columns only. In practice, callers already treat evaluation failures as non-restricting, but we still pay the full cost of the physical-expression compilation and evaluation path first.

This change adds a cheap guard to detect predicates that reference columns outside the allowed join columns and returns false early. That preserves the existing behavior while avoiding unnecessary work in a hot optimizer path.

What changes are included in this PR?

This PR makes two focused changes:

1. In is_restrict_null_predicate, collect the join columns into a HashSet and add a fast-path check that verifies whether the predicate only references those columns.
2. If the predicate references any non-join column, return Ok(false) immediately instead of attempting null-evaluation.

Additionally:

• The evaluated join-column set is reused for the fallback evaluate_expr_with_null_column path.
• InferredPredicates::insert_inferred_predicate is simplified to use .unwrap_or(false) when consuming is_restrict_null_predicate, which matches the prior effective behavior of treating errors as non-restricting.
• A regression test is added for a predicate like a > b, where b is outside the join-key set, to verify the fast path returns false.

Are these changes tested?

Yes.

A test case was added to cover the scenario where a predicate references a column outside the join key set:

• a > b now explicitly verifies that is_restrict_null_predicate returns false.

This exercises the new early-return path and protects against regressions in predicate analysis behavior.

Are there any user-facing changes?

No.

This change is an internal optimizer performance improvement and does not change public APIs or intended query results.

LLM-generated code disclosure

This PR includes LLM-generated code and comments. All LLM-generated content has been manually reviewed and tested.

[kosiew]

run benchmark sql_planner_extended

[adriangbot]

🤖 Criterion benchmark running (GKE) | trigger
Linux bench-c4067930471-314-dhsls 6.12.55+ #1 SMP Sun Feb 1 08:59:41 UTC 2026 aarch64 GNU/Linux
Comparing push-down-02-20002 (3d3945c) to ab28234 (merge-base) diff
BENCH_NAME=sql_planner_extended
BENCH_COMMAND=cargo bench --features=parquet --bench sql_planner_extended
BENCH_FILTER=
Results will be posted here when complete