Optimizer: infer N_g from an external dataset #693

Goal

Provide the optimizer with real label-group cardinality N_g from an explicit external dataset, without contacting Prometheus.

Scope

• Require a dataset input for the greedy optimizer CLI.
• Define a canonical series-inventory format: one row per metric series, with metric and label columns.
• For each profile key (metric, spatial_filter, grouping_labels), count distinct grouping-label tuples after applying the filter.
• Treat an empty grouping-label set as N_g = 1.
• Deduplicate raw samples by series identity if sample-level data is supported.
• Pass the resulting count into CandidateConfig and use it for non-subpopulation-aware ingest/query costs.
• Keep subpopulation-aware candidates at N_g = 1.
• Fail clearly when the dataset does not cover a required profile; never interpret missing data as N_g = 0.
• Keep the dataset loader separate from the optimizer core so an equivalent Prometheus-backed provider can be added later.

Tests

• Distinct counts for empty, single-label, and multi-label groupings.
• Spatial-filter handling.
• Repeated samples do not inflate cardinality.
• Missing profile coverage produces a clear error.
• Offline CLI and in-memory test fixtures work without Prometheus.