feat(control_plane): Phase 2 steps 3-9 -- merge L2/L3 onto asap-ir/asap-l2, relocate CSE, PromQL retarget, capability_for delegation

[zzylol]

Summary

This PR now bundles nine consecutive Phase 2 steps (all landed on this branch before review):

Step 3 -- query_expr.rs/relational.rs's L3 side onto asap-ir (first commit):

• QueryExpr / Predicate and their supporting types are no longer defined locally -- re-exported from asap_ir::intent_algebra (~22 variants vs. the pre-merge 16).
• Predicate is now L3Expr; BETWEEN desugars to Compare(Ge) AND Compare(Le).
• Partition doesn't exist in asap_ir -- folded into Aggregate.by: GroupKeys at construction time.
• ScalarSubquery is rejected at construction time (no equivalent in either this repo's parsers or ASAPController's own L2 lowering today).
• Fixes a real regression: Binder now also walks Filter/Aggregate.having/Join.pred/Project scalar trees for column names, required once Predicate became positional.
• Preserves avg_over_time as a p50 quantile-sketch approximation and (at that point) Rate/Increase/Delta collapsing onto AggIntent::Sum (disambiguated via outer_fn), per asap_tier_analysis's existing dispatch design -- both later corrected, see steps 6 and 6b.

Step 4 -- relational.rs's L2 side onto the new asap-l2 crate (second commit):

• QueryExpr/AggFunc/AggItem/SourceSpec/L2ProjectItem/L2SortKey, binder.rs, and column_resolution.rs are no longer defined locally -- re-exported from asap-l2 (same git dependency/pinned commit as asap-ir).
• Swaps promql-parser from crates.io 0.8 to the private ProjectASAP/promql-parser fork (pinned commit), matching what ASAPController's frontend-promql already depends on.
• lower.rs stays control_plane's own converter (not a re-export of asap_l2::lower), at that point to preserve the avg/rate/increase behaviors above.
• AggFunc::Frequency/Custom(String) (control_plane-only, no asap_l2 equivalent) preserved as behavior rather than vocabulary: promql.rs constructs plain AggFunc::Count, and lower.rs recovers Frequency-sketch routing via a grouped-or-windowed-Count trigger.
• Partition is gone at both L2 and L3 -- asap_l2's Aggregate carries without: bool directly.

Step 5 -- relocate cse.rs to optimizer (L4), adopt asap-plan's algorithm (third commit):

• dedupe_subtrees/CseWorkloadPlan move from intent_algebra::cse to optimizer::cse, matching ASAPController's layering (CSE is a cost-aware L4 optimizer decision, not an L3 intent-algebra concern).
• Re-implemented against asap_ir::intent_algebra::{BindingName, QueryId} directly (no types_v2 boundary conversion), using ASAPController's Vec+PartialEq structural-equality candidate scan instead of the old Debug-string-keyed HashMap.
• optimizer::cost::WorkloadCostPlan and workload_cost() switched to the same asap_ir types directly, removing an earlier boundary-conversion hack.
• Ported two ASAPController regression tests (quantiles_over_different_columns_do_not_dedupe, dedupe_subtrees_no_shared_subexpr_when_unique_keys_absent) alongside the pre-existing ones.
• R8 CommonSubexprElim (intra-tree, Merge-branch Scan dedup) and the relocated CSE pass (inter-tree, workload-level multi-root dedup) have disjoint scopes -- relocating CSE didn't require touching R8.

Step 6 -- PromQL topk/rate semantic retarget to match ASAPController (fourth commit):

• Adopts ASAPController's RankingMeasure/is_frequency_heavy_hitter gate for topk/bottomk instead of unconditionally forcing AggFunc::Count inside any topk(...): only descending topk ranking by count_over_time(...) now takes the heavy-hitter TopK path; everything else (bottomk, topk over a non-count measure) lowers to a generic Sort+Limit. Fixes a real correctness bug where e.g. topk(k, avg_over_time(...)) was silently treated as a Count/Frequency heavy-hitter aggregate.
• Rate/Increase/Changes/Resets/Delta/IDelta/Deriv/PredictLinear/DoubleExpSmoothing now map to their own dedicated AggIntents instead of collapsing onto Sum/Count. This activates capability_for()'s existing Rate/Increase -> ExactAgg(Increase) mapping and correctly routes the archive-only counter-derivative family (Delta/Changes/Resets/...) to None instead of falsely claiming an exact-agg capability.
• asap_tier_analysis.rs's outer_fn is computed independently of required_capability (straight off the raw PromQL AST function name), so this only needed test-expectation updates, not a dispatch-logic redesign.
• Adds structural regression tests asserting the raw QueryExpr shape (TopK vs Sort+Limit) for topk/bottomk, not just the flattened ParsedQuery view.

Step 6b -- avg_over_time routes exact, drops the p50 approximation (fifth commit):

• AggFunc::Avg now maps onto the literal AggIntent::Avg, matching asap_l2::lower's own mapping exactly, instead of the p50 quantile-sketch approximation carried since step 3.
• capability_for(&AggIntent::Avg) already correctly returns None (no ASAP-tier sketch substitute -- needs a cross-policy Sum+Count join), and ASAPController's own crates/plan/src/bind.rs treats AggIntent::Avg the same way (pass_through_intents_stay_logical keeps it a whole unsketched logical subtree) -- confirmed by reading the pinned-commit source directly rather than assuming.
• avg now routes through the exact/archive path like Sum/Count/every archive-only intent; QeCollector::collect_op's existing catch-all exact_required = true arm already handled this correctly, so the fix is purely the AggFunc::Avg mapping plus updated test expectations (aggregations goes from [Quantile]/quantiles: [0.5] to empty + exact_required: true).

Step 6c -- stddev_over_time/stdvar_over_time route exact, drop the IQR-proxy approximation (sixth commit):

• Same fix class as step 6b: AggFunc::StdDev/Variance now map onto their literal AggIntent::StdDev/Variance (matching asap_l2::lower exactly) instead of the two-quantile [q(0.25), q(0.75)] "IQR proxy" approximation carried since step 3.
• capability_for already declared both archive-only (Avg | StdDev | Variance => None), so the proxy was silently claiming a QuantileApprox ASAP-tier capability neither this repo's own capability table nor ASAPController's asap-plan actually backs with a real bind rule -- same bug class, found by auditing every remaining agg_func_to_intents arm against asap_l2::lower's literal mapping after fixing Avg.
• The Merge-of-siblings fan-out this used to trigger for StdDev/Variance collapses back to a plain single Aggregate.

Step 7 -- resolve capability_for/rules::dispatch divergences on Count and Min/Max (seventh commit):

• Count{non-Exact} was claiming CardinalityApprox/HLL ("distinct count" semantics) -- ASAPController's own crates/plan/src/bind.rs::readout maps Count to SketchQuery::PointCount (a frequency point-query, CMS), matching this repo's own BindCmsOnCount rule. distinct_over_time/COUNT(DISTINCT) always lower to AggIntent::Cardinality, never Count, so the CardinalityApprox premise never had a real caller -- AggIntent::Count{non-Exact} is unreachable via this repo's own PromQL frontend today regardless (lower.rs only ever constructs Count{Exact} or the Extension-based Frequency intent), so this is a consistency fix, not a live routing change.
• Min/Max were claiming QuantileApprox (min = quantile(0), max = quantile(1)), but no rule in sketch_algebra::rules actually implemented that -- bind_kll_quantile/bind_ddsketch_quantile only ever match AggIntent::Quantile, never Min/Max, so the promised coverage didn't exist and Min/Max silently fell through to archive regardless of what capability_for claimed. ASAPController's own crates/plan/src/boundary.rs treats Min/Max as an exact mergeable accumulator (SummaryKind::MinMax), same tier as Sum/Rate/Increase -- correct, since comparing two partial extrema needs no approximation at all. bind_exact_agg.rs now binds Min/Max -> AggregationType::MinMax (keyed -> MultipleMinMax), using the data plane's already-fully-wired MinMaxAccumulator; capability_for now returns ExactAgg(MinMax) to match.

Step 8 -- capability_for delegates to asap_plan::boundary::implementation_for (eighth commit):

• Adds asap-plan/asap-sketch as new git dependencies (same repo/rev as asap-ir/asap-l2; asap-plan depends only on asap-ir, so this pulls in no datafusion/front-end weight) and bumps the pin to pick up ProjectASAP/ASAPController#138-139 (a new CostModel trait for deployment-pluggable sketch ranking, and the realize/Realization -> implementation_for/Implementation Cascades-terminology rename plus CountSketch/CountSketchWithHeap SummaryKind variants).
• capability_for no longer maintains its own hand-written AggIntent -> capability match -- that was a second, parallel judgment kept in sync with asap-plan's own decision by hand, and had already drifted twice (the Count/Min-Max divergences step 7 just fixed). It now keeps only the Extension/Frequency special case (deployment-specific; asap-plan has no opinion on an intent shape it can't see into, by design) and delegates everything else to asap_plan::boundary::implementation_for, translating the returned Implementation into this repo's coarser Capability vocabulary via a new implementation_to_capability helper.
• One deliberate override survives the delegation: Count{Exact} is forced to None (archive) rather than trusting implementation_for's ExactAccumulator claim, because the data plane has no working count accumulator (SumAccumulator conflates Sum and Count).
• All existing tests pass unchanged -- confirms the delegation is behaviorally identical to the hand-written match it replaces.

Step 9 -- bump the ASAPController pin to pick up Implementation::is_satisfied_by (ninth commit):

• Picks up ProjectASAP/ASAPController#140: asap-plan's Implementation::is_satisfied_by (materialized-view-style matching over the SummaryKind vocabulary -- "does an already-available Implementation satisfy a required one", the query-optimization "answering queries using views" question, narrowed to that crate's own vocabulary).
• Purely a pin bump, no source changes -- is_satisfied_by is not consumed here. It's a distinct, narrower-scope question from this repo's own Capability::is_satisfied_by (deployment-specific index matching against an actual inventory of already-materialized sketch instances, shared wire format with data_plane's sketch_index) -- the two are separate types serving separate layers, not one delegating to the other. Wiring control_plane's Capability::is_satisfied_by to build on top of asap-plan's is tracked as a possible follow-up, not done here.

Phase 2's SQL frontend merge (frontend-sql) is out of scope for this PR per plan -- PromQL only.

Test plan

• cargo build -p control_plane -- clean after all nine commits
• cargo test -p control_plane -- 772 passed, 1 pre-existing failure (invalid_sketch_type_override_falls_back_to_default, confirmed identical via git stash against the base commit, unrelated to this change)
• cargo fmt -p control_plane applied each step; incidental reformatting of untouched files reverted to keep the diff scoped

🤖 Generated with Claude Code