These are P0 judgment calls made within the settled architecture.
The neutral trace has one value per field, while Prometheus metrics can carry many label sets. Counts and counters are summed, capacity fractions use the maximum, and histogram p95 values are estimated from cumulative buckets after deterministic aggregation. This preserves a visible worst fractional pressure and avoids silently dropping label sets. It can still hide skew, so the output names the policy and keeps unknown metric names visible.
Public token counters become rates only when an API caller supplies a prior snapshot and increasing timestamps. The adapt command handles one exposition, so it omits those rates rather than assuming a scrape interval.
The command accepts one explicit HTTP(S) URL only. It rejects embedded credentials, follows no redirects, has a timeout and byte cap, uses no retry loop, and records only source_kind: "url" in the output. The endpoint itself is not persisted in trace metadata.
The comparison computes a per-metric mean over available samples and evaluates thresholds independently. A zero baseline has no relative delta, so only the absolute boundary applies. A missing metric is not_comparable, not a zero.
The PEP 517 backend is part of this repository. It creates sorted wheel and source members, normalized archive timestamps and ownership, fixed metadata, and package license records. Editable wheels may include a local source pointer by definition, but release wheels and source archives do not include local paths or generated caches.