Metering: current UTC-day bucket ({window, as_of, currency, total, by_model, by_key}), raw token split + derived spend_micros

Per-key aggregation (same raw-split shape). CAPPED at the top 1000 rows by spend (the FinOps-relevant ordering); by_key_truncated says the cap fired, never a silent cut.

Per-(model, provider) aggregation: cost attribution by model (the FinOps unit).

The raw consumption counts + the derived spend estimate: the one shape shared by total, by_model rows, and by_key rows, so a consumer writes ONE aggregation reader.

The raw consumption counts + the derived spend estimate: the one shape shared by total, by_model rows, and by_key rows, so a consumer writes ONE aggregation reader.

A metering window: [start, end) epoch seconds.