feat(planner): optimizer accepts label-set facts and costs by cardinality

[milindsrivastava1997]

Closes #756. Part of #753.

Summary

• Label-set facts replace --dataset / --rho. asap-optimizer-cli --label-set-facts <yaml> takes series_count per (metric, spatial filter) and cardinality per (metric, spatial filter, grouping labels). Arrival rate is derived as series_count / scrape interval. The label schema now comes from the workload's metrics: hints, which are required. Missing facts produce an error listing every expected key; unused facts only warn. dataset.rs and SUBPOPULATION_COUNT are removed.
• Cardinality-aware costing. Memory and merge/subtract work scale with output_group_count for keyed maps (Multiple*) and with instance_count otherwise. Query reads scale with output groups, except top-k (one heap read per bucket). Trivial accumulators get an analytical per-key memory estimate (dictionary-encoded label codes + value, hash-table slack). Single-group Sum/MinMax/Increase are no longer proposed, since they tie with their Multiple* forms.
• Optimizer configs now match what the legacy planner deploys (pre-existing bug):
  • build_config now uses set_subpopulation_labels, so keyed sketches are one instance rather than one per group (or one top-k heap per series).
  • CMS/HydraKLL now deploy their paired DeltaSet key aggregation, referenced alongside the value in the query config.
• does_precompute_operator_support_subpopulations handles Set/DeltaSet and non-top-k CMS+heap instead of panicking. Legacy planner calls are unchanged.
• Docs: .design_docs/optimizer-v1-implementation-plan.md updated, plus runnable example workload and facts files.

Notes

• Filed Capability-matching fallback never matches keyed aggregations (grouping_labels = []) #779: the engine's capability-matching fallback never matches keyed configs (pre-existing; also affects the legacy planner). The primary query-config path is unaffected.
• Key-tracker retention is ⌈range / slide⌉ panes. The legacy planner reuses the value's count, which falls short for sliding windows.

Test plan

• cargo test -p asap_planner -p promql_utilities, clippy, cargo check --workspace
• New tests: facts parsing/validation/missing-key reporting (incl. topk by), metric-hint errors, label split per sketch type, cost scaling per class, top-k reads, end-to-end CMS + key tracker deployment and query-config references
• asap-optimizer-cli on the example files: HydraKLL + DeltaSet for the quantile query, MultipleSum for sum by; without --atomic-costs the run reports the top-k query as unservable (EXACT was removed in feat(planner): model assignments as SLA-aware items #776)

🤖 Generated with Claude Code