feat(countsketch)!: ASAPv1 codec with counter type and mode; drop the protobuf, msgpack and delta formats

[GordonYuanyc]

Stacked on #85 (asapv1-only/infra). Part of the ASAPv1-only series: Go's Count Sketch gets an ASAPv1 codec and loses its protobuf, msgpack and delta formats.

Draft because of the open questions below.

Added

• CounterType (CounterFloat64 default, CounterInt32, CounterInt64) and Mode (ModeFast default, ModeRegular) on CountSketch, and MarshalASAPv1 / UnmarshalASAPv1 (kind 0x04 0x00). All three cs_* fixtures pass CheckGolden.
• Regular mode is real behaviour: every method that has the key bytes (Update*, UpdateString*, the sampled/at-rows/GOS variants, Estimate*, ProcessInput/ColForRow/SignForRow/UpdateCell with key bytes, and the top-k rebuilds) derives cells through one mode-aware path that matches Rust's regular path or Rust's fast path on every geometry (packed-64, packed-128, per-row). Confirmed byte-identical to Rust for both modes.
• The encoder refuses sketches whose bytes would lie (decided during review; easy to revisit): methods that only get a precomputed 64-bit hash can't reproduce regular mode or fast mode outside packed-64, so a write through them on such a sketch sets a sticky flag (kept by Merge, cleared by Reset) and MarshalASAPv1 refuses.
• Float storage with integer counter types: encode refuses non-integral or out-of-range cells (float64-counter sketches can't be encoded at all); decode refuses i64 values float64 can't hold exactly.

Changed

• In fast mode, key-based methods now place keys like Update and Rust on matrices outside the packed-64 layout (they used Go-only mix64 hashing there).

Removed

• SerializePortable, SerializeProtoBytes, DeserializeCountSketchFromProtoBytes, SerializeMsgpack / DeserializeMsgpack, SerializeDelta / DeserializeDelta; asapmsgpack CS full, cell-delta and sparse-delta codecs; proto CountSketchDelta, CountSketchCell, and count_sketch in both envelopes (reserved 11; reserved "count_sketch";). CountSketchState stays for UnivMon and Hydra (Hydra's cell now builds it locally).

Open questions

1. Deltas: ASAPCollector sends Count Sketch as proto full state plus CountSketchDelta; both are removed and ASAPv1 has no delta.
2. f64 counters: the collector's sketches are float64 (weighted / 1/p-sampled); they can't be encoded until they move to integer counters or the spec gains an "f64" Count Sketch counter type.
3. Power-of-two cols: Go decodes only power-of-two widths, so a valid Rust sketch with e.g. cols = 3 is rejected (documented on UnmarshalASAPv1, tested with real Rust bytes).
4. Gaps: cell-level writes (IncrCell, SetCell, MergeDelta) don't set the refusal flag; the OctoSketch per-cell path now hashes key bytes on every call (not benchmarked).

Verification (Go 1.24.9)

go build, go vet, gofmt, go test ./.... Known failures also on main: CocoSketch tests, TestKLL_Reset_SubsequentInserts (flaky), tests/cross_language without XTEST_DIR. Implemented and reviewed by separate agents; review findings are addressed in the second commit.

🤖 Generated with Claude Code