feat(metrics): Add native histograms and Prometheus protobuf exposition

[ethanolchik]

Applications currently cannot expose native histogram samples from this client. This adds opt-in native collection to Histogram and the Prometheus protobuf exposition needed to scrape it.

Fixes #576.

const registry = new client.Registry(
  client.Registry.PROMETHEUS_PROTOBUF_CONTENT_TYPE,
);
const histogram = new client.Histogram({
  name: 'request_duration_seconds',
  help: 'Request duration',
  nativeHistogramBucketFactor: 1.1,
  buckets: [],
  registers: [registry],
});

Existing histogram configurations retain classic behavior. Native histograms can retain explicit classic buckets for migration, or use buckets: [] for native-only protobuf output. Protobuf registry methods return a Buffer, represented as Uint8Array in the public TypeScript declarations.

The implementation includes:

• Standard exponential schemas -4 through 8, positive and negative sparse buckets, a configurable zero bucket, and native exemplars. The default budget of 160 populated buckets reduces resolution as needed and remains a soft limit at schema -4.
• Native JSON snapshots and sum, first, and omit aggregation through registries, clusters, and workers, including reconciliation of different schemas and zero thresholds.
• Protobuf encoding for native histograms and existing metric types. protobufjs/light uses the checked-in lib/metrics.json descriptor; lib/metrics.proto and npm run generate-protobuf make its source and regeneration available.
• Public types, regression tests, configuration and migration documentation, a runnable HTTP example, and changelog entries.

Validation performed locally:

• 761 tests in 34 suites passed on Node 22.23.2, 24.21.0, and 26.4.0. The configured Bun CI command also passed.
• The Linux check workflow passed through act on Node 24.21.0: ESLint, Prettier, and TypeScript.
• npm run benchmarks completed on macOS/Node 26.4.0. A focused registry comparison against upstream, with increased sampling, found no significant regression above a 5% threshold; default-label cases measured roughly 2–4% overhead.
• Manual interoperability checks with Prometheus 3.12.0 covered native counts, sums, quantiles, signed and empty histograms, worker aggregation, classic coexistence, and other metric types.
• Protobuf descriptor regeneration matched the checked-in file, and the package dry run included the runtime schema and implementation files.

The existing benchmark suite covers classic metrics. Dedicated native workload benchmarks and application-specific rollout validation remain follow-up work. Applications select the protobuf response format themselves; HTTP Accept negotiation is outside this change.

Developed with AI assistance (Codex), also disclosed in the commit trailer.