feat: all-five-sketch query_statistic + cold-store wiring (P1)

[zzylol]

Summary

Backend half of the runtime e2e: each sketch type now has a verified PromQL → backend path through the modified-OTLP wire format (typed Metric.data = {DDSketch | KLLSketch | HLLSketch | CountSketch | CountMinSketch}). Companion PR in ASAPCollector (#204) lands the agent + harness.

query_statistic implementations

Accumulator	Statistics	Notes
DDSketchAccumulator	Quantile / Sum / Count / Min / Max	—
HllSketchAccumulator	Cardinality (Count accepted as alias)	lets the existing PromQL count(...) path land on HLL
CountSketchAccumulator	Topk / Count / Sum	no-key fallback returns row-mean total
CountMinSketchAccumulator	Count / Sum	no-key fallback returns min-row sum — canonical CMS total-event estimator, exact when each insert increments one cell per row

Factory wiring

accumulator_factory.rs: DDSketchAccumulatorUpdater wired in alongside CMS / CountSketch / KLL / HLL updaters so the precompute_engine recognises AggregationType::DDSketch from streaming.yaml.

Modified-OTLP envelope decoders go through from_sketchlib_proto_bytes / from_msgpack_bytes against SketchEnvelope { sketch_state: Some(SketchState::*(state)) }.

P1 — cold-store flag

asap-query-engine/src/main.rs:

• --cold-store-root with env = "ASAP_COLD_STORE_ROOT" plumbed through a build_adapter_config helper that selects prometheus_promql_with_cold when set.
• Combine with --forward-unsupported-queries to keep Prom as the tail of the chain; without it, unsupported shapes return empty.
• Four unit tests pin the wiring matrix (cold × forward).

This unblocks the three-way harness (#66) reporting the cold-tier accuracy figure end-to-end, and the cold path is what carries the e2e flow today (warm-tier sketch ingest is dropped at the OTel gateway — see ASAPCollector#204 follow-ups).

Known cleanup (not a blocker)

compatible_agg_types in capability_matching.rs doesn't list CountMinSketch under Statistic::Sum, but query_logics treats CMS as the canonical approximator for both Sum and Count. The exact-match find_query_config path bypasses capability matching, so the runtime e2e succeeds today; reconciling the two tables is a separate change.

Test plan

• cargo build / cargo check clean
• All five sketches round-trip through agent (sketchlib-go) → wire → backend accumulator
• DDSketch: histogram_quantile(0.5/0.9/0.99,…) → 21.12 / 47.95 / 104.60 within ε=0.01 (relative_quantile)
• KLLSketch: histogram_quantile(0.5,…) → 18.26 within ε=0.16 (rank_quantile)
• HLLSketch: count(http_requests_total) → 149.68 within ε=0.008 (relative_cardinality)
• CountSketch: sum_over_time(http_requests_total[1m]) → 24266 within ε=0.03 (additive_frequency)
• CountMinSketch: sum_over_time(http_requests_total[1m]) → 145735 within ε≈0.0027, δ=0.03125
• P1 unit tests: cold-store-root × forward-unsupported (4 cases)

🤖 Generated with Claude Code