feat(sdk/metric): AggregationRawBuffer — raw-sample encoding slot

[zzylol]

Summary

First code PR for the three-axis framework (design doc: #N — stacked on docs/sdk-three-axis-framework). Adds AggregationRawBuffer — the SDK-side encoding where every Counter.Add / Gauge.Record call becomes its own NumberDataPoint on the wire instead of being reduced to a Sum / LastValue per attribute set per tick. This is the paper's "raw" baseline on §6.2c's encoding axis.

Changes

• sdk/metric/aggregation.go — new public AggregationRawBuffer{MaxEventsPerSeries} type, wired into the Aggregation interface + err() validation.
• sdk/metric/internal/aggregate/rawbuffer.go — aggregator impl (~140 LOC). Buffers (ts, value) per attribute.Distinct key on measure(); emits each buffered sample as its own metricdata.DataPoint[N] inside a Gauge[N] on collect(), then clears the buffer.
• sdk/metric/internal/aggregate/aggregate.go — Builder.RawBuffer method, mirroring KLLSketch / HLLSketch.
• sdk/metric/pipeline.go — dispatch into Builder.RawBuffer + isAggregatorCompatible.

Design notes

• Delta and cumulative call the same collect(). Raw-buffer has no meaningful cumulative semantics — re-emitting all history every tick would be useless — so the buffer always resets after collect regardless of requested temporality. (Documented in the collect() godoc.)
• Overflow is silent drop with a per-series drop counter. Not backpressure, because backpressure would conflate "SDK overload" with "app slow path" in experiment numbers. Exposing the drop count as a side-channel metric is a follow-up tracked in PROGRESS.md.
• Per-series cap, not global cap. The L-axis (label projection) sweep intentionally drives cardinality down; a global cap would couple the two axes.

Drive-by fix

hllsketch.go was broken against current sketchlib-go:

• HyperLogLog.Insert(float64) was renamed to InsertValue(float64)
• Estimate() (no args) was renamed to EstimateCardinality() int

Two mechanical one-liners, no behaviour change. Needed to pass go build locally.

Test plan

Verified locally (Ubuntu 22.04, Go 1.22 w/ GOTOOLCHAIN=auto picking 1.24 per go.mod):

```
$ cd opentelemetry-go/sdk/metric && go test -run RawBuffer -count=1 -v ./internal/aggregate
=== RUN TestRawBufferEmitsEverySample --- PASS
=== RUN TestRawBufferSecondCollectIsEmpty --- PASS
=== RUN TestRawBufferRespectsMaxEventsPerSeries --- PASS
=== RUN TestRawBufferHandlesManyAttributeSets --- PASS

$ go build ./... # sdk/metric full package — clean
$ go vet ./... # clean
```

Four unit tests cover:

1. Every Add/Record call in an interval produces one DataPoint — no SDK pre-aggregation
2. Second collect() with no new observations emits zero data points (cumulative doesn't accumulate)
3. Per-series cap honoured — extra measurements on a full buffer are dropped
4. Many distinct attribute sets each keep independent buffers

Still to do in follow-ups:

• fake-exporter wiring (EXPORTER_SDK_AGG=raw-buffer path)
• Producer-side columns in measure-baseline.py

🤖 Generated with Claude Code