feat: use scenario-matched empirical costs in the optimizer - #725

Goal

Turn benchmarked sketch costs into a dataset-conditioned cost-based planning workflow for ASAPQuery, starting with KLL quantile_over_time on the Google task-usage trace.

What has been done

• Load versioned, workload-profiled atomic cost documents from sketch-bench.
• Enforce accuracy and atomic CPU latency SLAs in the optimizer, with structured infeasibility errors.
• Make query cost fanout depend on the cardinality of grouping-label tuples (N_G).
• Add the Dataset Wrangler: it materializes a trace scenario, inventory, and reproducible manifest.
• Require an atomic-cost profile selector containing both workload and scenario identity; mismatched benchmark data is rejected.
• Record and preserve the first provenance-gated Google KLL planning run and its intermediate artifacts.

Current scope is deliberately research-first: CPU-only optimization, no filter selectivity, KLL as the first workload, and E2E retained as selected-plan validation rather than the optimizer sweep mechanism.