[tisonkun]

Summary

This draft implements the rough consensus from the TCK discussion for cross-project review.

• Add Count-Min serial version 2. Derive row seeds as XXH64(row_index.to_le_bytes(), base_seed) and use XXH64 for item hashes.
• Map version 2 item hashes to buckets with (hash >> 1) % num_buckets.
• Preserve the legacy Rust version 1 hash path when reading, updating, and reserializing old images.
• Reject merges between different serial versions and document the remaining cross-language version 1 limitations.
• Add independently computed version 2 known-answer vectors and legacy Rust version 1 regression coverage.

Compatibility

New sketches serialize as version 2. Legacy Rust version 1 images remain readable and retain version 1 when reserialized, so their counters are not mislabeled as using the new hash scheme.

Version 1 images from other implementations remain unreliable for queries, updates, and merges because the image does not identify their legacy row-seed derivation. Existing C++ version 1 snapshots continue to be covered for deserialization and byte round-tripping.

Cross-language version 2 TCK snapshots and byte comparisons still need to be added after the format details are agreed across implementations.

Refs: apache/datasketches-tck#18
Refs: #297

Validation

• cargo x check
• cargo x prepare-testdata
• cargo x test
• cargo x lint