Add optimizer aggregation families and capability/SLA metadata #762

This was generated by AI during triage.

Change

Expand promql_utilities::AggregationType and establish one aggregation-capability descriptor for optimizer-facing behavior. The descriptor must be the single source for: query statistic support, exact versus approximate accuracy, benchmark accuracy metric, and comparison direction. Eligibility consumes this descriptor instead of maintaining a separate type match.

Target capability mapping:

• sum and count: MultipleSum
• top-k: existing CountMinSketchWithHeap and sketchlib CSHeap (CountSketch with heap)
• increase: MultipleIncrease
• min and max: MultipleMinMax
• quantile: KLL, DDSketch, and HydraKLL

Add DDSketch, CountSketch, CountSketchWithHeap/CSHeap, and UnivMon aggregation types where appropriate. CSHeap is the concrete name used by asap_sketchlib; choose an ASAPQuery enum spelling that is clear and map it explicitly to that implementation.

SLA metadata

For every approximate family that is eligible under accuracy_sla, define the benchmark metric and whether lower or higher is better. Error metrics use actual_error <= 1 - accuracy_sla; recall metrics use actual_recall >= accuracy_sla. Exact families bypass the accuracy measurement. Missing or malformed measurements make an approximate candidate ineligible.

In particular, do not use CMS relative_error_mean as an accuracy SLA metric unless its semantics and scale are validated. The benchmark output has values far above [0, 1] because rare-key long-tail errors dominate. If Count-Min Sketch is limited to heavy-hitter/top-k serving, use an appropriate bounded top-k quality metric such as recall instead.

Delivery boundary

This issue may introduce types and optimizer metadata before runtime support exists. The engine must return an explicit unsupported-aggregation-type error for any type it cannot execute. Such candidates must be excluded by default and considered only with --milp-allow-unsupported-sketches.

Part of #753.