@nightowlsdev/metrics

The metric-series store (FR-059): a small, library-shaped observation plane for numbers that move over time, exactly the shape the knowledge and graph planes silently corrupt. A point's identity is `(orgId, orgScope, metric, at, dims)`; `record()` overwrites rather than duplicating and `recordBatch()` is one all-or-nothing transaction. `compare()` is the pinned period-over-period question, half-open windows, a `dataLag`-derived cutoff that truncates only the current window (never shifts the previous one), calendar-aligned buckets flagged `partial` at every edge, totals computed over RAW observations not buckets-of-buckets, and a signed `pct` with an absolute-value denominator. Float reduction is ordered and identical in both stores (Postgres never runs a native `float8` sum, because doubles are not associative and PostgreSQL raises on an overflow JS turns into `Infinity`). Library-shaped like `@nightowlsdev/auth-sessions`, you own the `pg` Pool (a peer dependency), the package owns its `metrics_000N` migration series and never runs DDL, and the engine never reads or writes these rows, the consumers are agents and hosts. Security posture is server-only: RLS on with no policy and an explicit `REVOKE`, so the Pool must own the table or hold BYPASSRLS, and a browser gets these numbers through your server, aggregated, never as a raw row feed. `orgScope` on reads is an IDENTITY SELECTOR, not FR-055 visibility, `undefined` selects the org-wide series only (there is no unrestricted all-scopes read), because an aggregation that silently mixed departments into one number is the exact corruption a metrics store exists to prevent. NOT for this plane: engine-emitted metrics, dashboards, alerting, downsampling, cross-scope aggregate reads.