test: port the legacy-search callers outside logical-optimizer to Stage 1

[zzylol]

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

Rebased: univmon_candidates.rs keeps #621's assertion that only L2 is certified (RelativeValue), on the Stage 1 candidates.

Problem

The legacy search (ASAPStrategies, search_workload*, explain_replacements, retain_exact) still had callers outside logical-optimizer: one devtool and a set of frontend, integration and executor tests. It cannot be deleted (Q68 (a), step 7c) until they move to Stage 1 or are deleted.

Changes

One decision per file. Shared helpers: stage1_candidates(root) in executor/tests/common/mod.rs and frontend-promql/tests/support.rs composes every Pass 1 choice for one query (up to 4096) and skips those that do not compose. stage1_plan(root) in the frontend-promql support composes the first summarized choice, or else the pass-through choice.

File	Decision
devtools/src/bin/sketch_coverage.rs	Rebuilt on Pass 1. Runs enumerate_local_logical_candidates per lowered query. It prints the sketch alternatives Pass 1 offers for each query (a Hydra alternative prints as Hydra(<algorithm>)), and per corpus the fraction of lowered queries with at least one. The CSE-shareable metric is dropped: Stage 1's identical-expression sharing is a workload variant, not a per-query replacement. Over every corpus: 135/2178 lowered queries get a sketch alternative at ε = 0.01.
frontend-promql/tests/observability/promql_corpus.rs, metrics_observability.rs	Ported to Stage 1 (stage1_plan). Pass 1 must accept every lowered query, and the query must compose to a candidate. Totals: docs 19 summarized / 28 unchanged / 0 errors; testdata 354 / 721 / 0; every metrics-observability corpus 0 errors. A choice can legitimately fail to compose, for example a Top-K heap over a closed schema or a summary over two sources. Selection skips such choices, so the probe does not count them as errors.
frontend-promql/tests/count_planning.rs	Ported to Stage 1 candidates. The Hydra test now checks that Pass 1 offers a Hydra alternative and that selection does not pick it (renamed grouped_count_offers_hydra_but_selection_keeps_per_group_state). Its input carries the series identity: Hydra hashes a non-null item, and PromQL labels are nullable.
frontend-promql/tests/univmon_candidates.rs	Ported to Stage 1 candidates. The synthetic AccuracyModel is gone. The shared count is lowered at an approximate target, because Pass 1 offers no UnivMon for an exact count (the legacy search did). Stage 1 nodes carry no guarantee, so the "no calibrated bound" check uses DefaultAccuracyModel::local_guarantee.
integration-tests/tests/promql_numeric_regressions.rs	Ported to Stage 1. plan takes each target's first alternative other than pass-through; the weighted-heap test takes the heap that absorbs the inner aggregate. Dropped: the node-guarantee assertions (Stage 1 sets none; the ratio-certification gap is listed in #623) and the Top-K evidence provider. The two-metric guarded expression became avg_over_time(a[5m]) + avg_over_time(a[5m] offset 5m): Pass 1 offers no summary over two sources. A signed sum's Count-Min heap is now checked to carry no non-negative proof, not to be absent; Stage 3 rejects it.
integration-tests/tests/precompute_raw_samples.rs	Ported: candidates are the Stage 1 compositions that Stage 3 could admit. Count-Min over unproven weights is skipped.
integration-tests/tests/cse.rs	Deleted. It pins the legacy search's memo groups. Its share-vs-recompute coverage is already listed in #623, and identical-sub-DAG merging is covered by pass2::identical_expressions's tests.
frontend-sql/tests/pearson_corr.rs	Ported: the exact plan is Stage 1's pass-through composition. Pass 1 offers only pass-through for corr. The exact-guarantee assertion is dropped.
executor/tests/weighted_topk_binding.rs	Ported where Stage 1 has the realization. Kept: grouped weighted top-k binding (see Findings, now #[ignore]), direct per-series rate heap with dynamic labels, maintained rate heap. These run on Count Sketch only, because Stage 1 proves no sign for rate values and the executor rejects a Count-Min heap without one. The evaluation-timestamp assertions are dropped, because Stage 1's heap evaluation outputs no timestamp column. Deleted: physical_binding_does_not_impose_an_accuracy_acceptance_policy (it only varied legacy evidence), spatial_topk_exposes_signed_heap_candidate_over_complete_snapshot (legacy current_series_topk_candidates only), and direct_rate_topk_exposes_heap_candidates_with_complete_series_identity (a closed catalog schema with no series identity has no Top-K item in Stage 1).
executor/tests/planspace_series_identity_heap.rs	Kept selection_never_commits_a_series_identity_heap (already on plan_stages). Deleted the four tests of the legacy search's whole-root proposals.
executor/tests/precompute_candidates.rs	Ported: the grouped-rate candidates come from Stage 1.
executor/tests/deployment_computation.rs	Ported: the Count-Min count(up) candidate comes from Stage 1.
executor/tests/current_series_heap.rs, deployment_computation.rs (population_dag), precompute_candidates.rs (population_topk_*), frontend-sql/tests/maintained_population.rs, frontend-promql/tests/maintained_population_horizon.rs	Unchanged. They call MaintainedPopulationStrategy::candidate directly, not the search. The plan for 7c is to keep candidate and drop only its ReplacementStrategy impl.

Findings: Stage 1 gaps these ports exposed (not fixed here)

1. Bug: the whole-expression heap (topk by(job)(k, sum by(service, job)(rate(...)))) keeps the outer top-k's partition keys. Those keys index the inner aggregate's output, but the heap is built over the raw rates. Here by(job) became key 0 = ts, and the output schema loses job. The legacy search remapped the keys. planner_weighted_topk_binds_at_either_deployment_phase is #[ignore]d with this reason.
2. Stage 1's heap evaluation outputs no evaluation-timestamp column; the legacy realization did.
3. No non-negativity proof for rate values, so no Count-Min heap over a rate.
4. No Top-K item identity for a closed schema without the series identity.
5. No summary over an input with two sources.
6. Hydra over PromQL needs the series identity column (labels are nullable).

Stacked on #631.

Test plan

• cargo fmt --all --check
• cargo clippy --workspace --all-targets --all-features --locked -- -D warnings
• cargo test --workspace --locked: 1508 passed, 0 failed, 15 ignored (after the rebase on main d4869a7; was 1484 passed, 23 ignored)
• cargo run -p asap-devtools --bin sketch_coverage -- --data-ingestion-interval-ms 1000 runs

🤖 Generated with Claude Code