Metrics: explicit-bucket Histogram (fixed buckets → GCounter)

[keddie]

🤖 This comment was generated by Claude on behalf of @keddie.

Part of #1201.

Explicit-bucket Histogram — fixed buckets → GCounter

The classic Prometheus-style histogram: the consumer predefines bucket boundaries, and each bucket counts how many observations fell in it. Mergeable form is trivial and reuses machinery we have.

Shape

Histogram ≈ (bucket_boundary → GCounter)   // fixed key set, chosen up front
          + GCounter                        // total count
          + GCounter-of-sum (optional)       // running sum for the mean

• Buckets are a fixed, config-given set of boundaries — the key space never changes.
• Each bucket is a GCounter so counting is idempotent under anti-entropy re-delivery (the per-source-keyed counting trick; naive count-add double-counts).
• Merge = per-bucket GCounter merge — associative, commutative, idempotent.
• Serializes to OTLP HistogramDataPoint (explicit bounds + bucket_counts).

Why this exists alongside DDSketch (#1201's sub-issue)

Explicit-bucket vs. DDSketch is a real choice, not redundancy:

• Explicit — you pick boundaries; great when you know the range and want fixed, cheap, predictable buckets. Bad if you guess wrong (p99 lands in one giant low-res bucket).
• DDSketch (exponential) — log buckets auto-cover whatever range appears at uniform relative resolution; better for latency spanning orders of magnitude.

Ship both; they map to two distinct OTel point types (Histogram vs. ExponentialHistogram).

Notes / design surface

• Bucket boundaries are a cluster-wide constant (like α for DDSketch and p for HLL) — two histograms merge only if boundaries match. Fixed config, not per-call.
• Decide whether to carry sum and min/max (OTLP allows optional min/max) — min/max are not cleanly mergeable as CRDTs unless done as separate min/max-registers; call it out.

Acceptance

• bucket → GCounter histogram with fixed boundaries + total count (+ optional sum).
• OTLP HistogramDataPoint serialization.
• Tests: record/merge, per-bucket idempotency under re-merge, boundary-mismatch rejection.

Placement (zoo vs. exporter-local) is resolved once for the epic — see #1201.