feat(query-engine): add DDSketch aggregation type

[zzylol]

Summary

Adds a DDSketch quantile aggregation to the ASAPQuery data plane and query engine. This is the DDSketch slice of #787 (and the DDSketch aggregation type from #762); CountSketch and UnivMon are not included.

The accumulator wraps asap_sketchlib::DDSketch, the implementation sketch-bench measures, so the optimizer's DD cost rows describe what ASAPQuery runs. It is needed for the data-plane vs ClickHouse evaluation (#789), which compares DDSketch on both sides (quantilesDD(α) in ClickHouse).

• Type: AggregationType::DDSketch (aliases DDSketchAccumulator, ddsketch, dd). It is single-population like DatasketchesKLL, and capability matching serves Statistic::Quantile from it.
• Accumulator: DDSketchAccumulator handles update, quantile query, merge, and msgpack/JSON serialization. A mismatched alpha or a different accumulator type is rejected on merge. Negative values go to a mirrored store and zeros to a zero bucket, as in ClickHouse's quantilesDD; non-finite values are dropped. An empty sketch returns NaN.
• Config: a DDSketch aggregation requires parameters.alpha in (0, 1) and fails validation otherwise (InvalidAlpha). DDSketch::new panics outside that range, so the check runs before the accumulator is built.
• Dependency: asap_sketchlib moves from 94d76f6 to 5158bf2 for the zero bucket and negative store (fix(ddsketch): retain negative observations and zeros asap_sketchlib#141), which no release includes yet. No other sketch's tests change.
• Planner: builds {alpha} with a default of 0.01, overridable via sketch_parameters.DDSketch.alpha, and adds the sketch's properties (mergeable, not subtractable).

Not included

• The planner still maps quantiles to DatasketchesKLL (map_statistic_to_precompute_operator). Choosing DD is part of --planner milp (Optimizer: move sketch-bench #129 MILP into asap-planner-rs (--planner milp) #753).
• The MILP optimizer skips DDSketch (OPTIMIZER_SKIPPED_AGG_TYPES) until it has a parameter grid and atomic_costs mapping (Add optimizer aggregation families and capability/SLA metadata #762). Without the skip it would propose DD candidates with no alpha.
• A deterministic choice between quantile configs. If a KLL and a DDSketch config share a metric and window, capability matching picks whichever HashMap order puts first (already true of KLL vs HydraKLL). Breaking ties by aggregation_id is a follow-up.
• Subtraction for sliding windows. DD bucket counts could support it, but the accumulator only merges, matching KLL.

Tests

• Accumulator unit tests: quantiles stay within α on 1..=10,000; zero and negative inputs are counted; an empty sketch returns NaN; kwargs are validated; a merge equals a single sketch over all values (directly and through merge_with); mismatched α or type is rejected; serialization round-trips.
• Factory: the configured α reaches the accumulator; reset clears state; missing, out-of-range or string α is rejected; config_is_keyed agrees with the updater.
• Config: YAML loads with a valid alpha and is rejected without one.
• Capability matching: a quantile query matches a DDSketch config.
• Dependency: asap_sketchlib moves from 94d76f6 to 5158bf2 for the zero bucket and negative store (fix(ddsketch): retain negative observations and zeros asap_sketchlib#141), which no release includes yet. No other sketch's tests change.
• Planner: default α, and the override.
• End to end (e2e_grouped_quantile_over_ddsketch_is_within_alpha): Remote Write → precompute → one DDSketch per group → quantile by (job) (0.9, …). Each job has two instances and twenty zeros; each group's estimate is within α of the true p90, which the zeros move outside the bound if they were dropped.

cargo test for asap_types, promql_utilities, asap_planner and query_engine_rust passes; cargo test --workspace passes (1,187 tests). cargo clippy --workspace --all-targets --all-features --locked -- -D warnings and cargo fmt --check are clean.

Part of #787.

🤖 Generated with Claude Code