Tracking: RQE frontend for the paper eval #886

Goal

For the paper eval, run ASAPQuery's side of an experiment without PromQL/SQL parsing: a query is an RQE expression tree (self-contained RQE leaves, combined by aggregations and binary ops), POSTed to a new engine route that looks up each leaf's sketch directly by RQE id. Baselines (Prometheus, ClickHouse, …) keep PromQL/SQL. Covers sketch-bench eval tables and the random workloads, PromQL- and SQL-style queries.

sketch-bench side: ProjectASAP/sketch-bench#224

Steps (one PR each, stacked PRs, worktrees via gwt.sh)

• A1 feat(asap-common): RQE expression tree type and validation #880 (PR feat(asap-common): RQE expression tree type and validation #887)
• S1 sketch-bench#222 (PR sketch-bench#225)
• A2 feat(asap-tools): RQE trees and feeder-format data facts in random workloads #881
• S2 sketch-bench#223 (covered by sketch-bench#226 + feat(asap-tools): add accelerated random E2E workload artifacts #879; trees: sketch-bench#227, PR sketch-bench#228)
• A3 feat(planner): --planner saved reads RQE trees and writes rqe_configs #882 (PR feat(planner): --planner saved reads RQE trees and writes rqe_configs #888)
• A4 feat(query-engine): RQE tree query route #883 (PR feat(query-engine): answer RQE expression trees on POST /api/v1/rqe_query #889)
• A5 feat(asap-tools): rqe query protocol and frontend switch for e2e runs #884 (PR feat(asap-tools): rqe query protocol and frontend switch for e2e runs #890, code; e2e check pending)

Decisions

• Use: paper eval only; the experiment scripts call the RQE route for ASAPQuery. The PromQL/SQL paths are unchanged.
• Tree: leaf rqe is self-contained (id, metric/table, capability, lookback_ms, grouping, phi, k, value). agg = avg/min/max/sum with by only, by ⊆ child labels. binary = + - * / (scalar operands allowed), equal label sets. Top-k only at the root. Leaf capabilities: sum, quantile, cardinality, topk_by_value (rate and topk_by_count out of scope; queries using them get no expr).
• Where the tree comes from: written by the generators next to promql/sql (expr), for existing query shapes only; new templates (e.g. avg by (G)(quantile_over_time), binary in random workloads) are a follow-up.
• Sketch lookup: direct by RQE id via a new rqe_configs section of inference_config.yaml, written by --planner saved (roll-up deployments included); query_configs kept.
• Random workloads: planned like eval tables: sketch-bench converts a manifest to a table (S2), solved with --target p95 (any accuracy_sla ≠ 0.95 fails), converted by --planner saved. data_facts use the paced feeder's metrics format so the feeder and the planner agree: data Pareto α 2.0, Zipf θ 1.1 over 10,000 series, 10 samples/s; events uniform labels, user_id Zipf θ 1.1 over 10,000 keys, 10 records/s per leaf.
• Route: POST /api/v1/rqe_query {expr, time} → Prometheus vector JSON with exactly Prometheus' output labels. No fallback. No data → empty result; unknown RQE id / invalid tree → 400.
• Client: rqe protocol; configs keep PromQL/SQL text, a side file maps it to expr, so results and calculate_fidelity are unchanged. query_engine.frontend: promql|rqe (default promql).
• Out of scope: SQL accuracy scoring (asap-tools: calculate_fidelity never scores SQL queries #885), filtered RQEs, Prometheus range queries, new templates.

Done when

One classic sketch-bench table and one random workload run end to end through the RQE route with the paced feeder, matching the PromQL route for the same plans wherever the PromQL route works.