fix(pass1): offer no per-group sketch for SQL COUNT(*)

[zzylol]

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

Rebased on main d4869a7 (DF 54). Added fix: integrate with #614: DF 54 sorts ORDER BY COUNT(*) by the aggregate's own column, so canonicalize promotes the ranked test query to TopK over the Count (as it already did ORDER BY c); the executor has no native TopK, so sql_count_star_candidates_compose_and_execute now executes every choice only for the grouped and ungrouped queries. priced_sql_count_star_candidates_bind still covers all three.

Problem

For SQL COUNT(*) (grouped, ungrouped, or ORDER BY COUNT(*) DESC LIMIT k), Pass 1 offered per-group Count-Min, Count Sketch and UnivMon realizations. Each one hashes the grouping column with a unit weight (sql_row_count_update). All of them lower to physical plans, but none binds in the executor:

• Count-Min and UnivMon fail with keyed summary weight must be a finalized value column.
• Count Sketch fails with summary family has no native DAG state implementation.

Stage 3's capabilities rejected Count-Min and Count Sketch, but it priced the UnivMon candidate as valid, nearly tied with the selected exact plan. So Stage 3 could select a plan that cannot run.

These sketches also do nothing useful here. Every row in a group hashes that group's key, so the sketch is just a counter that takes more memory. HydraCms is the only grouped-count sketch that helps.

Changes

• enumerate_local_logical_candidates: when the intent is Count and every child is SQL rows (closed, no PromQL series identity), drop the Realization::Sketch alternatives. HydraCms is added afterwards and stays. PromQL count and count_over_time keep Count-Min, Count Sketch and UnivMon.
• Tests:
  • pass1_sql_coverage::sql_count_star_candidates_compose_and_execute (replaces the grouped-only test): for the grouped, ungrouped and ORDER BY COUNT(*) DESC LIMIT 2 forms, every Pass 1 choice composes, compiles, binds and executes to the exact counts. HydraCms is included and is exact on this tiny data.
  • Invariant: priced_sql_count_star_candidates_bind and priced_example2_candidates_bind check that every physical candidate plan_stages prices with executor_models() binds in the executor. Without the fix, both new SQL count tests fail on the UnivMon candidate.
  • Unit test sql_count_star_offers_no_per_group_sketch: SQL COUNT(*) gets exactly pass-through, exact Count and HydraCms. PromQL count by and count_over_time still get Count-Min, Count Sketch and UnivMon.
  • Updated the alternative counts in filtered_aggregates::pass1_offers_filtered_count_alternatives (now every alternative executes), sql_hydra_count_dp_equals_exhaustive and sql_filtered_aggregates_dp_equals_exhaustive (18 → 9 combinations).

Notes:

• The Example 2 invariant uses Q3's floating product (as in planner_layering_example2's Q3_FLOAT). With the integer Q3, Stage 3 prices an exact Sum over the Int64 c * c that does not bind, because the executor's summary_build only takes Float64 updates. That is a separate gap and is not fixed here.
• Example 2's selected plan is unchanged ("Q1 exact · Q2 exact (Count acc) · Q3 exact (Count acc) · shared summary"). The committed tools/dag-viewer/examples fixture is byte-identical.
• Related: SQL top-k ordered by an alias (ORDER BY c DESC LIMIT k) does not bind #619 (SQL top-k ordered by an alias).

Test plan

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace: 1671 passed, 0 failed, 21 ignored (re-run after the rebase on main d4869a7)

🤖 Generated with Claude Code