branch-4.2: [fix](local shuffle) Populate distribution keys for select nodes #67941

[924060929]

Cherry-picked from #67941

[924060929]

run buildall

[924060929]

/review

[github-actions]

Approval opinion: no blocking or otherwise actionable issue was found on exact head c1e29eb44a8e986831bb08de0aa115d5d5fde51e.

Review checkpoints:

• Goal and correctness: centralizing the three translator-created SelectNode paths preserves the prior node IDs, Nereids mapping, predicates, projections, limits, and fragment-root behavior while supplying the missing output and child distribution metadata. SelectNode still forwards the parent's exact requirement; the child expression list now consumed by PlanNode.enforceRequire is resolved after child translation, so inserted exchanges receive bound child slots. A window placement on v cannot satisfy the join requirement on k, while already matching hash placement and exact GLOBAL/LOCAL/BUCKET flavors retain their existing satisfaction behavior.
• Scope and call-chain coverage: all three changed files and every Select construction branch were reviewed, along with physical property projection/request derivation, analytic/sort/join/set-operation enforcement, local-exchange serialization, null/ANY properties, aliases and consecutive projects, serial-source handling, and FE/BE planner boundaries. No unresolved candidate remains.
• Tests: the new three-task regression enables FE local-shuffle planning and exercises both mismatched (PARTITION BY v, join on k) and aligned (PARTITION BY k, join on k) cases. Its count and independent sums are a meaningful oracle for the historical row-loss failure. Direct plan-shape/key assertions and stacked-alias/serial variants would be useful follow-up coverage, but no production defect was established from those gaps.
• Configuration and compatibility: no configuration definition, protocol, persisted format, transaction/data-write path, rolling-upgrade contract, or documentation surface changes. Existing session variables only select the regression path.
• Concurrency, lifecycle, and observability: no shared-state, locking, ownership, cleanup, cancellation, logging, or error-propagation behavior changes.
• Performance: metadata extraction is planner-only and bounded by the distribution-key/output sets. Static tracing found no new unconditional repartition or exact-flavor downgrade.
• Validation status: this was a static review only; the runner contract prohibited builds and tests. Current CI shows the FE style/license and listed unit/coverage contexts successful. ShellCheck is currently failing on repository-wide shfmt findings in shell files outside this PR's three changed paths, so it is not treated as a finding against this patch.
• Review convergence: two independent complete-review agents and one separate risk-focused agent all returned NO_NEW_VALUABLE_FINDINGS; the main agent independently adjudicated every recorded risk and the final changed-file/unresolved-candidate sweep was clean.
• User focus: no additional review focus was supplied.

[924060929]

Superseded by #68199, which backports the whole local shuffle series as one PR.