OTLP metrics completeness scope for aggregate mode

[aksOps]

Part of map #195. Resolves #194 finding 9.

Question

The receiver handles Gauge and Sum only; Histogram and ExponentialHistogram are silently ignored; AGGREGATE_METRIC_DIMS parses but point attributes never reach the series identity.

Decide the GA scope: native ExponentialHistogram merge into the scale-4 sketch (scale conversion rules), fixed-bucket Histogram handling without expanding observations (bucket-midpoint fold? reject with counter?), dims extraction from point attributes at the reduction site (interning path exists — InternDimValues), unsupported-type metrics, and whether OTLP partial-success reporting applies. Explicitly rule what stays out of GA.

[aksOps]

Resolution (frozen 2026-08-22)

Q1 — ExponentialHistogram scale conversion: adopt with one correction.
For input scale s > 4, downscale exactly to scale 4 by shifting indexes and merging counts. For s < currentScale, explicitly downscale the accumulated sketch to s — do not rely on incidental bin-collapse behavior. OpenTelemetry's perfect-subsetting property makes downscaling exact at the bucket-count level. Validate scale/index ranges and reject malformed points. Fold zero_count, count, sum, min, and max normally. Negative buckets: do not report whole-distribution percentiles when negative buckets contain observations — keep their count and scalar statistics, increment otelcontext_ingest_metrics_sketch_dropped_total{reason="negative_observations"}, and mark percentiles unavailable. Excluding negatives but publishing the positive-side p99 as the global p99 would still be a lie.

Q2 — Fixed-bucket Histogram: adopt A with stronger accuracy semantics.
Fold bucket counts directly as weighted synthetic observations; never expand count observations in a loop. Geometric midpoint only for finite positive buckets. Buckets that include negative values cannot enter the positive-only sketch unless min >= 0 proves the population is non-negative; otherwise preserve scalar count/sum/min/max and mark percentiles unavailable. Track the +Inf bucket separately: when a requested quantile lands there, return a lower-bound result (p99 >= last_finite_boundary), not an ordinary numeric estimate. Accuracy metadata must include source-histogram bucket error or unbounded_tail; degraded=true alone is not enough to describe an unbounded bucket. Explicit Histogram boundaries are part of the source approximation and can impose substantially more error than the scale-4 sketch itself.

Q3 — Cumulative histogram temporality: adopt A.
GA supports delta-temporality Histogram and ExponentialHistogram points. Reject cumulative histogram points completely from aggregate accounting, increment otelcontext_ingest_metrics_unsupported_total{type,reason="cumulative_temporality"}, and include them in OTLP partial-success rejected counts. Document collector-side conversion through cumulativetodeltaprocessor (supports cumulative Sum, Histogram, and ExponentialHistogram streams) — and document that the processor is stateful and needs stable routing to one collector instance for correct conversion.

Q4 — Dimensions from point attributes: adopt with a hot-path correction.
Extract the configured keys from each metric point and intern the canonical tuple into DimsID. Missing any configured key produces DimsID=0, preserving the existing all-or-nothing contract. Apply equally to NumberDataPoint, HistogramDataPoint, and ExponentialHistogramDataPoint. Do not retain the current per-point map[string]string allocation: configured dimensions are bounded, so scan the attributes into fixed/request-local slots or compile lookup indexes. Canonically support scalar string, integer, boolean, and double values; reject array and map values from identity with a reasoned counter. Cardinality continues to bind through the metric-series caps.

Q5 — Unsupported types and response honesty: adopt.
Summary remains wholly unsupported. Count every dropped Summary, cumulative Histogram, malformed ExponentialHistogram, or otherwise unsupported point and populate ExportMetricsPartialSuccess.rejected_data_points with the exact total, with a bounded message naming the rejected types and reasons. The Export still returns success for accepted points; clients must not retry a populated partial-success response. Zero rejected count is reserved for warning-only responses where all telemetry was accepted — this case genuinely rejects points, so the count must be nonzero.