Bounded join and subplan planning

The bounded planner follows DataFusion's cross-join elimination for a narrow physical pattern: (A × B) JOIN C with equality links from both dimensions to C. When both dimension estimates are no larger than C, it can build each dimension and probe C through two linked joins instead of materializing the Cartesian intermediate. A projection and, when required, blocking Arrow exchange restore the original schema and distribution. Hints, residuals, NULL-safe equality and ordered joins retain the original plan. The SQL remains unchanged.

This allowed TPC-H Q9 at scale factor one to finish within the original 512 MiB native authority. Its repeated release comparison measured 27,230.35 ms stock versus 1,919.77 ms native (14.184×). A separate regression-check cohort measured Q2 at 3,427.85 ms stock versus 535.95 ms native (6.396×), replacing an earlier native Cartesian plan that lost to stock. These shared-host cohorts include full scans, transposes, JNI, blocking shuffles, sinks and completed job results; their variability and earlier failures remain in the TPC-H evidence.

The planner also preserves Flink's reuse of a shared unary aggregate subplan. It runs the host subplan reuser before native substitution, counts consumers of the admitted shared aggregate, and preserves that subtree's digest through Flink's clone-and-reuse cycle. The existing native share operator then gives each branch a retained Arrow buffer view. Other native nodes retain their single-consumer digest barriers. Both host reuse switches and native source sharing must be enabled. Scheduling, replay and memory accounting remain with their existing owners; no result cache is introduced.

Sharing reduced Q15's native median from 1,393.73 ms to 794.67 ms in separate cohorts, but the matching stock median was 596.21 ms. This remains a regression to optimize and is not claimed as acceleration. The complete 22-query goal remains open while its performance and validation work is unfinished.