[benbellick]

Which issue does this PR close?

Closes apache#25345.

Rationale for this change

Grouped aggregation only requires each equality group to remain within one input partition. The preceding range-locality analysis PR can prove that property for audited transformations of concrete range boundaries, allowing aggregation to avoid a hash shuffle without claiming general partition-layout compatibility.

For example, hourly truncation of a timestamp range split at an hour boundary keeps every (key, hour) group local. A split inside an hour does not, so repartitioning remains required.

What changes are included in this PR?

Aggregation enforcement and validation now share one locality policy across combined and uncombined partial/final plans. Useful range partitioning is preserved, repeated optimization remains stable, grouping sets remain excluded, and joins cannot consume the weaker locality guarantee.

The SQLLogicTest removes repartitioning for aligned hourly date_trunc boundaries while retaining it for overlapping day buckets and date_bin.

This PR is stacked on the preceding range-transformation locality analysis and should be reviewed after it.

Are these changes tested?

Optimizer regressions cover uncombined partial/final validation, repeated enforcement after aggregate combination, grouping-set rejection, and join isolation. The targeted SQLLogicTest and full extended workspace suite pass.

A reproducible Criterion benchmark uses eight in-memory hourly range partitions and eight rows per group. It builds the data and physical plan outside the timed section, then measures execution with and without the otherwise unnecessary hash repartition. Two synthetic workloads showed approximately 35–36% lower execution time.

Are there any user-facing changes?

Eligible grouped aggregations can avoid hash repartitioning. Other operators continue to require ordinary distribution compatibility.

---

Note: This PR was developed with AI assistance. All changes have been reviewed, and I take full responsibility for this contribution.

[benbellick]

Combined into #3 so the range-locality analysis and its aggregation consumer can be reviewed as one complete optimization.