feat(plan): L3→L4 binding pass + sketch-vs-exact boundary decision (#98)

[zzylol]

Closes #98.

Implements the L3→L4 binding and the sketch-vs-exact boundary in asap-plan (which now depends on asap-sketch — L3 in, L4 out, still no front-end dependency).

boundary.rs — the per-intent accuracy decision

realize(&AggIntent) -> Realization is one exhaustive match over the intent vocabulary, so a new AggIntent variant cannot compile without an explicit decision — no silent fall-through, per the issue's coverage requirement. Three outcomes:

• Sketch { kind, params } — for the approximate-capable intents under an Epsilon/EpsilonDelta target, per the issue's map: Quantile→KLL (DDSketch as the listed alternate), Cardinality→HLL (Theta/KMV alternates), TopK→CMS-with-heap, approximate Count→CMS. sketch_candidates() keeps the full per-intent candidate list in one place for the future cost model (Wire workload-level CSE into a cost model #6/Add logic to detect which optimizations are applicable to a query workload #33); realize binds the head.
• ExactAccumulator { kind, params } — the mergeable exact summaries: Sum/Count(Exact)/Min+Max→MinMax/Rate/Increase. Gated on agg_is_mergeable (the previously caller-less helper — Avg/StdDev/Variance need richer partial state and pass through).
• PassThrough — everything with no summary form, explicitly: exact quantile/top-k/cardinality, classic-bucket HistogramQuantile (PromQL: histogram_quantile classic-bucket vs sketch-able quantile is a structural heuristic — drive it from sample type/metadata #79 — pre-aggregated le buckets are not re-sketchable; the generic Quantile path is), Group/CountValues, and the per-series transform families (PromQL: native-histogram accessor functions (histogram_*) unsupported #43–PromQL: presence functions (absent/absent_over_time/present_over_time) unsupported #47, PromQL: extended aggregation operators (limitk/limit_ratio/count_values/group) unsupported #49, PromQL: long-tail functions (extra *_over_time reducers, sort family, min_of/max_of) unsupported #51).

Parameters are sized by inverting each family's error bound against the AccuracyTarget: ε=0.01 → KLL k=200 (matching the design-doc worked example), default_cardinality inverts back to HLL precision 14 exactly, δ→CMS depth ⌈ln 1/δ⌉ (0.01 → 5), with a DEFAULT_DELTA for ε-only targets and saturation for degenerate ε.

bind.rs — the binding pass

bind(&QueryExpr) -> Rc<L4Node> (plus bind_in with an explicit BindingScope for CSE'd roots) walks the L3 tree and fires the boundary decision per node over nested aggregate spines:

• a sketch realization becomes SummaryAgg (kind+params committed, col resolved from input_col() or SampleValue, by carried) wrapped in the SummaryEstimate readout — the Sketch(kind, params) column type appears on the agg edge and stops at the estimate, per the L4 type invariant;
• an exact accumulator is a bare SummaryAgg (its state ≡ the value; no readout query exists or is needed);
• everything else wraps as Logical(...) with the schema lifted to all-Primitive.

Edge schemas reuse the canonical L3 output_schema derivation with only the summary-state column retyped, so per-series label preservation (rate), output_names overrides, and time-index tracking carry over for free.

Conservative fallbacks (documented in the module docs): SummaryExpr::Logical boxes a whole L3 subtree with no L4 children, so a logical operator above a bindable aggregate (Filter/BinaryOp over a quantile) subsumes it unbound; multi-intent Aggregates and HAVING also stay logical. Rewriting through logical parents is the L4 rule engine's job (#6/#33), not this pass's.

Tests

• Coverage matrix (boundary.rs): all ~44 intent variants pinned to their category, plus assertions that every accumulator pick is agg_is_mergeable and every sketch pick is on a genuinely approximate target.
• Boundary per accuracy target: same intent, Exact vs ε=0.01 vs ε=0.05 → three decisions; HLL p=14 inversion; (ε,δ) CMS sizing; TopK heap=k; degenerate-ε saturation.
• Bind pass: KLL+estimate shape and both edge schemas, exact-sum accumulator, per-series rate label preservation, nested per-node binding, logical-parent/HAVING/multi-intent fallbacks, SQL named-column resolution.
• End-to-end pin (asap-e2e/tests/l4_binding.rs): quantile(0.99, rate(http_requests_total[5m])) from query string → Estimate ∘ SummaryAgg(Kll{200}) ∘ SummaryAgg(Rate) ∘ Logical(TimeRange→Scan), node by node with edge types.

Docs: reconciled the four stale "#98 is tracked/stub" references in design.md and the asap-plan crate docs.

Verified: cargo test --workspace (326 tests, 0 failures), cargo clippy --all-targets clean.

🤖 Generated with Claude Code