About the changes

Adds a new Histogram metric type to the impact-metrics system, following the same patterns as existing Counter and Gauge metrics.

Key Features

• observe(value, labels?) method - Records observations with optional labels
• Prometheus-compatible bucket system - Configurable buckets with automatic Infinity bucket (in prometheus users don't have to specify Infinity bucket themselves and we do the same)
• Memory-efficient storage - Uses bucket counts instead of storing raw values to handle millions of observations (check Discussion points for the impact of this decision)
• Full restoration support - Can restore exact histogram state after collection failures (more complex than gauge/counter due to the nature of histograms)
• Type-safe discriminated unions - NumericMetricSample | BucketMetricSample prevents invalid states (histograms are very different from numeric counters/gauges)

Important files

Start review by reading tests to understand how histograms work in prometheus.

Discussion points

The backend will need to support batch histogram restoration from collected bucket data, which prom-client doesn't natively support, so we'll likely need to build custom logic to reconstruct prom-client histograms from our bucket counts. I still think it's better than recording millions of samples and then calling observe million times on the server side.