feat(control-plane): realize Planner window candidates and feed back costs #473

Goal

Make ASAPQuery-backend the owner of concrete realization and workload-level deployment optimization for the abstract summary, window-framework, and lifecycle candidates searched by ASAPPlanner.

ASAPPlanner owns abstract SummaryWindowFramework choices such as tumbling, sliding, and exponential histogram. This control plane owns the runtime/library and executable configuration that realizes each abstract choice, and feeds implementation feasibility and workload-specific resource evidence back into Planner ranking.

Required behavior

• Consume Planner's candidate identity without changing its statistic, grouping, accuracy, source coverage, lifecycle semantics, abstract window framework, or query-time semantics.
• Enumerate executor-feasible concrete implementations for every Planner candidate, including supported tumbling layouts, sliding/pane layouts, PromSketch exponential-histogram implementations, and future registered primitives.
• Keep logical query windows, abstract Planner window frameworks, and concrete physical panes/configuration as separate concepts.
• Filter implementations using live Collector/backend capabilities before reporting evidence; an unexecutable implementation must not make a candidate available.
• For every complete per-summary framework assignment, derive complete CPU, transient and retained memory, source/storage I/O, network, concurrency, and multiplicity evidence under the current DataWorkload, then bind it as a StreamingWindowFrameworkCandidate.
• When several concrete implementations realize the same abstract assignment, retain their identities downstream and report the cheapest feasible evidence for that assignment.
• Preserve a mapping from the winning Planner evidence to its concrete implementation so the selected abstract plan can be compiled deterministically.
• Reuse or adapt the existing ASAPQuery candidate-generation and facility-location/MIP formulation for workload-wide sharing, placement, and query-to-configuration assignment.
• Generate stable content-derived materialization identities and emit one cross-validated CompiledPlanBundle whose Collector, backend, and query projections agree.
• Keep concrete configuration and deployment identity in the backend bundle; do not add it to ASAPPlanner IR.

Acceptance tests

• Every Planner candidate is either backed by complete feasible implementation evidence or marked unavailable.
• Tumbling, sliding/pane, and PromSketch EH implementations are considered only for matching Planner framework candidates and advertised runtime capabilities.
• Multiple implementations of one abstract assignment are costed without leaking implementation IDs into Planner results.
• The concrete implementation corresponding to the winning evidence is retained and compiled.
• Shared workload queries can select one physical configuration with multiple assignments.
• Raw/exact fallback covers the same source snapshots and horizon as the Post-ASAP alternative.
• Unknown algorithms, window variants, stale statistics, and incomplete evidence fail closed.
• Deterministic input produces deterministic fingerprints and plan identities.

References

• feat(cost): model streaming summary resources ASAPPlanner#331
• feat(control-plane): migrate latest ASAPPlanner IR into the physical compiler boundary #448
• ProjectASAP/ASAPCollector#584
• ProjectASAP/ASAPCollector#586
• ASAPQuery configuration formulation: https://github.com/ProjectASAP/ASAPQuery/blob/main/.design_docs/sketch-config-optimization-formulation.md
• ASAPQuery optimizer formulation: https://github.com/ProjectASAP/ASAPQuery/blob/main/.design_docs/optimizer-mip-formulation.md