L4: implement the L3→L4 binding (AggIntent→SummaryKind) and the sketch-vs-exact boundary decision

[zzylol]

Context

L1–L3 are built and tested end-to-end (lower_promql/lower_sql → canonical L3 QueryExpr, demo: cargo run -p asap-lower --example topk_ir). Above L3, asap-sketch defines the L4 vocabulary (SummaryExpr/L4Node/SummaryKind/SummaryParams/L4Schema) but nothing constructs it — there is no L3→L4 lowering pass, so the entire sketch layer is defined-but-unwired.

Work

1. L3→L4 binding pass — walk the L3 tree and emit SummaryExpr:
   • the AggIntent → SummaryKind map: Quantile→KLL/DDSketch, Cardinality→HLL/Theta/KMV, TopK→CMS-with-heap, Count(approx)→CMS, exact accumulators for Sum/Count/Min/Max/Rate/Increase;
   • a pass-through/exact policy for the ~30 intents with no sketch realization (per-series transforms Math/TimeFn, presence, counter-derivatives, histogram accessors, Group/CountValues, the *OverTime reducers) and the non-aggregate nodes (Relabel, Sample, InfoJoin, VectorFromScalar/ScalarFromVector, EvalTime, …).
2. Sketch-vs-exact boundary decision (asap-plan/src/boundary.rs) — the per-node choice between an exact operator and a sketch, driven by:
   • the AccuracyTarget threaded onto approximate intents;
   • the currently caller-less helpers agg_is_exact / agg_is_mergeable (asap-ir/src/intent_algebra/agg_intent.rs);
   • HistogramKind sketchability (PromQL: histogram_quantile classic-bucket vs sketch-able quantile is a structural heuristic — drive it from sample type/metadata #79): AggIntent::HistogramQuantile (classic buckets) is not re-sketchable; the generic Quantile path is.
3. Tests: an AggIntent→SummaryKind coverage matrix (every intent maps or is explicitly exact/pass-through — no silent fall-through), boundary decisions per accuracy target, and a first end-to-end query-string→L4 pin.

Notes

• Replaces the dead TODO(#34) references in asap-plan (that issue covered L3 canonicalization and is closed — the pass landed in asap-l2::canonicalize).
• Cost-model wiring and optimization-applicability detection stay tracked by Wire workload-level CSE into a cost model #6 / Add logic to detect which optimizations are applicable to a query workload #33; this issue is the binding + boundary only.

[zzylol]

@milindsrivastava1997 FYI