fix(kll,cms): use proto decoder on ingest to match emit wire format

[zzylol]

Summary

• The kllprocessor and countminsketchprocessor emit proto-encoded sketches (SerializePortable() + proto.Marshal for KLL, SerializeProtoBytesFO() for CMS) but their inbound paths were calling the gob-format decoders (DeserializeKLLSketchFromBytes, DeserializeCountMinSketchFromBytes).
• All production traffic that emits typed KLLSketch / CountMinSketch data points uses the proto envelope. The wire-format mismatch made every inbound pre-aggregated sketch fail to deserialize and silently drop on merge.
• Switched both ingest sites to the matching proto decoders: kll.DeserializeKLLSketchFromProtoBytes and cms.DeserializeCountMinSketchFromProtoBytes — symbols that already exist in sketchlib-go (verified at /home/.../sketchlib-go/sketches/KLL/portable.go:58 and .../CountMinSketch/portable.go:111).

Why this surfaced now

PR #220 (3b05c93) audited the processor sink/wire-format expectations end-to-end as part of unbreaking the post-DataDog→sketchlib-go fixtures. While reviewing the existing TestBatchModeTransmitSketch (which already correctly decodes the emit payload via DeserializeKLLSketchFromProtoBytes), the asymmetry on the ingest path stood out: the emit and ingest sides of the same processor disagreed on the wire format. TestProcessor_TumblingWindow_Correctness in CMS shows the same pattern — emit decoded with DeserializeCountMinSketchFromProtoBytes, but deserializeCMS (the function used on ingest) calls the gob path.

What the round-trip test proves

Added TestRoundTripIngestProtoSketch in both packages. Each test:

1. Builds a source sketch with a known multiset (KLL: 10 floats; CMS: 7 inserts at known hashes).
2. Serializes it via the processor's own emit-side helper (serializeKLLSketch / serializeCMS) — i.e. exactly the bytes a peer collector would put on the wire.
3. Feeds those bytes back into the same processor as a typed KLLSketch / CountMinSketch input data point.
4. Verifies the processor's emitted output decodes to the same Count() (KLL) / per-hash frequency estimates (CMS) as the source.

Pre-fix, both new tests fail (KLL: no output emitted because deserialize errored before merge; CMS: gob: duplicate type received while trying to gob-decode proto bytes). Post-fix both pass, exercising the full ingest → merge → emit → decode round-trip on the production wire format.

No emit-side behavior change

The diff is one-line per processor on the ingest path only:

• kllprocessor/processor.go lines 208 and 462: DeserializeKLLSketchFromBytes → DeserializeKLLSketchFromProtoBytes.
• countminsketchprocessor/processor.go line 693 (deserializeCMS): same swap.

serializeKLLSketch, serializeCMS, appendTypedKLLSketchDataPoint, buildWindowMetricsAndReset, cloneCMS, the delta-encoding ingest branch, and every other emit-related callsite are untouched. Wire format on the emit side is unchanged; downstream consumers see byte-identical output.

Test plan

• cd opentelemetry-collector-contrib-patch/processor/kllprocessor && go test ./... — passes (includes new TestRoundTripIngestProtoSketch).
• cd opentelemetry-collector-contrib-patch/processor/countminsketchprocessor && go test ./... — passes (includes new TestRoundTripIngestProtoSketch).
• Confirmed both new tests fail when only the ingest deserializer is reverted to the gob path (catches the bug they document).

🤖 Generated with Claude Code