Coverage lost with the legacy candidate_selection tests (follow-up of #580)

[zzylol]

Follow-up of #580. The legacy selection (asap_plan_selection::candidate_selection: global_selection, cost_sorted, recurrence_profiles) and its legacy-only test files are being deleted (Q52, approved 2026-10-04). Some behaviours were covered only by those tests, and the #509 stage pipeline (plan_stages) doesn't cover them yet. Listing them here so the coverage can be rebuilt on the stage pipeline, or explicitly dropped.

Accuracy propagation and exact composition

• Max / avg over a quantile composes at query time, with accuracy propagated through the outer fold (integration-tests/tests/exact_composition.rs: max_and_avg_over_quantile_compose_at_query_time_with_statistics, avg_over_quantile_keeps_the_sum_over_count_rewrite_as_a_competitor, promql_max_by_zone_over_quantile_over_time_composes, identity_and_genuine_multi_row_folds_both_compose, a_shared_inner_summary_is_materialized_once_for_several_outer_folds, outer_summary_over_an_exact_function_composes_at_ingestion_time, every_exact_accumulator_is_finalized_before_an_outer_sketch, root_target_rejects_unproven_composition, custom_accuracy_rule_survives_root_target_and_materialization).
• DDSketch quantile ratio: the shared relative-error target, domain proofs, signed evaluations, population size (promql_to_post_asap.rs: ddsketch_quantile_ratio_meets_the_shared_relative_error_target, ddsketch_ratio_without_domain_proof_is_uncertified, ddsketch_ratio_rejects_unsafe_domains, ddsketch_ratio_rejects_one_invalid_domain_when_the_other_is_missing, ddsketch_ratio_bound_holds_for_signed_pinned_sketch_evaluations, ddsketch_ratio_requires_a_supported_population_size).
• Approximate children are never relabelled exact through binary arithmetic (promql_binary_arithmetic_never_relabels_approximate_children_as_exact).

CSE share-vs-recompute by recurrence and consumer count

• cost/recurrence.rs: plan_selection_uses_recurrence_profiles_for_cse_choices and the recurrence_profiles_* tests (mixed intervals across roots sharing a sub-DAG, invalid rates, unreachable sites, repeated-reference multiplicity).
• candidate_selection.rs: cost_sorted_orders_shared_sub_dag_candidates_by_cse_share_decision, effective_consumer_count_corrects_a_nested_groups_share_decision, effective_repetition_materializes_a_cse_choice_for_a_single_edge_child, shared_ancestor_keeps_a_single_use_cse_descendant_selected, estimate_cost_does_not_scale_with_the_readers_own_consumer_count.

Rollup vs CSE

• global_selection_compares_a_logical_rewrite_with_the_cse_choice, mixed_rewrite_group_keeps_and_selects_its_explicit_cse_pair, global_selection_propagates_uses_through_the_selected_rewrite, costed_logical_candidate_beats_uncosted_legacy_cse_choice.

Evidence-ranked physical cost

• mixed_candidate_ranking_uses_explicit_candidate_costs, cost_sorted_pairs_each_candidate_with_its_own_estimate_cost, complete_plan_costs_reject_unbound_cse_arms, exact_composition.rs: missing_cost_statistics_preserve_the_conservative_retain_exact, unknown_runtime_capability_keeps_candidate_but_prevents_selection. See also Stage 3: evidence-driven cost (storage I/O, hand-off bytes) #615 (storage I/O and hand-off evidence).

Missing accuracy evidence is never selected

• selection_never_chooses_a_candidate_missing_accuracy_evidence, grouping_cost_cannot_resurrect_unprovable_hydra_candidates, explicit_summary_infeasibility_prevents_selection, nothing_is_selected_when_the_root_target_rejects_every_summary. The frontend-promql tests for Hydra counts and uncalibrated UnivMon estimates were ported to plan_stages, so this is partly covered for those families.

Reconciled tighter-target propagation

• global_selection_propagates_reconciled_consumers_to_the_tighter_group, cost_sorted_and_global_selection_handle_a_shared_looser_target, cost_sorted_and_global_selection_handle_a_single_consumer_looser_target.

Also legacy-only

• integration-tests/tests/sql_to_post_asap.rs and promql_to_post_asap.rs pinned the post-ASAP shape that ASAPStrategies + global_selection produced for many SQL and PromQL queries (temporal reuse, joins, filters, top-k, binary arithmetic, exact workloads). Equivalent shape pins on Stage 1 candidates or the selected plan don't exist yet for most of them.