v0.12.1

Measuring admission capacity

When the portal refuses a stream because it is full, the metrics now say which capacity ran out rather than collapsing every refusal into a single overload signal. Clients see the same response; operators see the difference, which is the only place the distinction can be acted on.

• Refusals counted by cause. Stream slots, download headroom, worker backoff and worker rate limits are separate series, including the two refusals that previously reached no metric at all — one decided before the stream exists, one raised after the response status is already committed.
• Occupancy integrated, not sampled. Time spent at the admission cap accumulates in-process, so saturation shorter than the scrape interval still leaves a trace and the mean is exact at any scrape cadence.
• The cap is published. Alert expressions no longer have to restate the configured limit.

The served API reference also gains an error-handling page for the taxonomy introduced in v0.12.0: which field to branch on, which to match, which never to parse, and what to do with a code you do not recognise.

See spec/12-observability.md for the metric inventory and docs/openapi/04-errors.md for the error contract. (#144)

Full Changelog: v0.12.0...v0.12.1