test: select plans through plan_stages instead of global_selection

[zzylol]

Stack: #574 → #620 → #618 → #621 → #627 → #625 → #628 → #632 → #634 → #616 → #617 → #622 → #624 → #629 → #630 → #631 → #633 → #635 → #636 → #637

Problem

Many tests used the legacy candidate_selection::global_selection (plus assemble_selected_dag) only to get a selected DAG to execute or inspect (Q52). That kept the legacy selection alive and tested plans the planner no longer selects.

Changes

These tests now take the plan the #509 stage pipeline selects (plan_stages, one root, run once over a continuously ingested source, with the built-in models). The selection goes through a small selected_dag(root, accuracy) helper in each test crate's existing shared module:

• executor/tests/common (reference-executor capabilities)
• integration-tests/tests/executor_models
• frontend-promql/tests/support.rs
• the cfg(test) fixture in executor/src/physical_planner/candidates.rs, which calls plan_stages inline

Ported: executor deployment_computation, planspace_series_identity_heap, precompute_candidates, promql_fallback, tumbling_pane_merge, physical_planner/candidates.rs; frontend-promql count_planning, univmon_candidates; integration-tests operator_sharing, precompute_raw_samples, sql_to_physical; planner summary_sharing.

Legacy search calls (search_workload*, ASAPStrategies) that these files use for other assertions are untouched. They belong to the legacy-search deletion (Q52 step 7).

Assertions that only held for legacy behaviour

• deployment_computation::per_series_average_divides_stored_sum_by_count: Stage 1 has no avg → sum/count rewrite (that was the legacy AvgToSumOverCountStrategy), so avg_over_time now selects the raw PromQL fallback. The test now plans sum_over_time(m[5m]) / count_over_time(m[5m]), which still divides stored per-series sum and count states and drops the name. The overflow assertion is dropped: it checked the avg accumulator's checked division, which a Stage 1 plan never reaches (inf / n is inf, not an error).
• promql_fallback::histogram_quantile_selection_keeps_the_exact_fallback: the legacy search kept histogram_quantile whole. Stage 3 keeps it exact (no sketch, still asserted for both targets), but under sum by (le, job) it selects an exact Sum accumulator below the quantile. This whole-expression fallback harness can't bind that plan, so only the raw-bucket query is executed (still 1.75).
• count_planning / univmon_candidates: "the chosen candidate never has missing accuracy evidence" becomes "the selected plan contains no Hydra state" and "no UnivMon state", the families whose candidates lacked evidence.
• operator_sharing::each_side_of_union_all_holds_a_summary_estimate: now #[ignore]d with the same reason as the other Pass 2 sharing and Stage 1 coverage parity with the retired MajorPass #580 ignores. Stage 3 selects the raw plan, because query-time summaries never cost less until Stage 2 plans materialization.
• planner/tests/summary_sharing::certified_frequency_evaluations_share_one_univmon_state: deleted. It ran the legacy search and selection with a UnivMon-preferring cost model, and its stage-pipeline counterpart frequency_moments_share_one_univmon_in_the_stage_pipeline already covers the behaviour. PREFER_UNIVMON goes with it.

Stacked on #624.

Test plan

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace: 1610 passed, 0 failed, 14 ignored (after the rebase on main d4869a7; was 1607 passed, 14 ignored)

[zzylol]

Closing as part of the upper-stack consolidation (G1): its ports of the legacy callers onto the planner pipeline are in #674. Closed with the author's approval.