Pass 2 sharing and Stage 1 coverage parity with the retired MajorPass

[zzylol]

Follow-up to #581 (#509 stage pipeline replaces MajorPass, #572 update of 2026-10-04). The facade now runs Stage 1 (logical_candidates) → Stage 2 (physical_candidates) → Stage 3 (plan_selection). These temporary regressions were accepted and must be closed.

Pass 2 sharing

Stage 1 lists one target's local alternatives only. The pipeline merges identical input sub-DAGs and identical summary producers after composition, but it does not plan sharing across queries:

• one state sized for the strictest consumer;
• one state that serves several statistics, such as p50 and p99 KLL and UnivMon moments.

Stage 1 coverage

MajorPass strategies that have no Stage 1 counterpart yet: rollup, avg → sum/count rewrite, top-k limit reuse, exact composition, shared sub-DAG and accuracy reconciliation, Hydra grouping, maintained populations.

Selection differences

Stage 2 runs every node at query time. Under Stage 3's analytical pricing, a query-time summary therefore never costs less than the raw aggregate it replaces, so the pipeline selects raw plans where MajorPass selected summaries. This will change once Stage 2 plans materialization.

Ignored tests (#[ignore = "...: #<this issue>"])

crates/planner/tests/summary_sharing.rs:

• Pass 2: quantiles_share_one_producer_sized_for_the_strictest_consumer, cross_series_p50_and_p99_share_one_producer, sql_p50_and_p99_share_one_producer
• raw plan selected: quantiles_with_equal_params_share_one_producer, different_producers_are_not_shared, identical_ungrouped_queries_share_their_producers, identical_sql_percentiles_share_one_producer

Changed expectation: crates/integration-tests/tests/operator_design_examples.rs batch_planning_selects_and_executes_each_plan (was batch_planning_replaces_and_shares_summary_operators): Stage 3 selects the raw SUM.

Legacy selection still in replacement.rs

Nothing in the facade uses it any more. The next PR (D2) moves it, and a later one deletes it once the parity above holds:

• GlobalSelection and its impl (assemble_selected_dag, ...);
• CandidateLogicalASAPDAGs::global_selection, global_selection_with_recurrence, global_selection_impl;
• search_workload, search_workload_with, search_workload_with_targets, cse_workload, search_cse_workload_with, default_strategies_with_evidence.

PlanningModels.cost is unused by the pipeline

Stage 3 prices plans analytically (plan_selection::price), so the stage pipeline never reads PlanningModels.cost; only the legacy search above does. This is documented on PlanningModels. Either Stage 3 takes a cost model, or the field goes away with the legacy search.