executor: HydraCms kernel for count and point queries

[zzylol]

Rebased on main d4869a7 (DF 54).

Why

#580 item F / W7: Hydra for count and point queries first. The types already model GroupingStrategy::SharedMultiSubpopulation { kind: HydraCms, .. }, but the executor rejected every shared grouping. Its only Hydra kernel was HydraKLL, which automatic search never emits. Part of #509.

What

• Before this PR: a SummaryAgg with HydraCms grouping failed at binding (invalid Count-Min dimensions or grouping strategy, or keyed summary weight must be a finalized value column). family_merges ignored grouping, so HydraKll merges type-checked even though they have no accuracy model.

• After this PR: count by (job) and count of service="checkout" by (job) run against one shared grid, not one CMS per job:

  job	count (Int64)	checkout (Float64)
  api	4	2
  batch	2	1
  db	3	1

  Each estimate is at least the exact value and within e·N·(1/shared_columns + 1/width) of it. Two builds over a split stream, merged with SummaryMerge, return the same answers as one build over the whole stream.

How

• summary_kernels/hydra_cms.rs reuses asap_sketchlib::Hydra, whose cells are HydraCounter::CM, with a one-column schema: the whole group tuple is one subpopulation, so rows do not fan out to sub-groupings. Readouts:
  • Item point count: Hydra's median estimator.
  • Group count: the minimum over rows of the cell's Count-Min row mass. With non-negative weights every row overestimates, so the minimum is sound, at most the median, and an integer. sketchlib has no group-mass query, so this mirrors its private subkey encoding. The kernel tests fail if the two encodings drift apart.
• HydraCmsGroup is the row state: an Arc to the shared grid, plus the group it reads.
• Factory: create_hydra_cms(family). create_planner_accumulator still rejects shared grouping, because it builds one state per group.
• Capability checks for the update, the native family, the evaluation and the state payload.
• New operator SharedSummaryBuild, plus the planner branch in bind_operation.
• Merge soundness: family_merges is now grouping-aware.
  • HydraCms and HydraCountSketch → true. Same-shape grids add cell-wise under the same hashes, so the result is the grid of the union, and Theorem 2 still holds.
  • HydraKll → false, since it has no accuracy model.
  • The kernel rejects shape mismatches, because sketchlib's CMS merge panics on them. It also caps total weight at i32::MAX, the limit of sketchlib's i32 counters.
• Tests first: before the wiring, tests/hydra_cms_execution.rs failed 2 of 3 (the binding errors above). After it, they pass.

Planner contract (for the later Pass 1/Stage 1 PR; not emitted yet)

SummaryAgg {
  family:   Sketch(SketchKind(Cms, Cms { width, depth }), G),
  grouping: G = SharedMultiSubpopulation { kind: HydraCms,
              params: HydraCms { width, depth, shared_rows, shared_columns } },
              // width/depth must equal the family's; all four > 0
  input: SummaryUpdate {
    item:   Some(Column(<one item column: Utf8 | Int64 | Bool, non-null>)),
    weight: Constant(1.0)                      // count
          | Column(<Float64, integral, ≥ 0>),  // weighted count
    weight_domain: NonNegative { .. },         // required
  },
  reduction: Reduce(GroupKeys::by(<group columns>)),
  filter: None,
}
SummaryEstimate { query: PointCount { key: SampleValue, value: None } }      // group count → Int64
SummaryEstimate { query: PointCount { key: Named(item), value: Some(v) } }   // item freq in group → Float64
SummaryMerge                                                                  // allowed (family_merges)

Output rows: [group columns…, state], one per group. default_hydra_params already gives this shape from a per-group SketchParams::Cms.

Not in this PR

• Top-k (W7 defers it).
• HydraCountSketch and HydraKll kernels.
• The precompute (raw-sample) path, which still rejects shared grouping.
• Pass 1 / Stage 1 offering Hydra.
• Memory accounting: each group row reports the whole grid's size, so reservations are conservative. A SummaryMerge also merges the grid once per group.

Gates

• cargo fmt --all --check: passes.
• cargo clippy --workspace --all-targets --all-features -- -D warnings: passes.
• cargo test --workspace: 1,590 passed, 7 ignored. feat(executor): execute the exact SQL entropy fallback; Example 2 status #599 had 1,580 / 7; this PR adds 10 tests.

🤖 Generated with Claude Code