query-engine: semicolon-joined group keys collide #912

Problem

Group keys are turned into strings by joining label values with ; and turned back by splitting on ;. The encoding is not injective, so distinct groups can collide:

A label value containing ; therefore merges two groups in a sketch, and decoding splits it into the wrong label list.

Where

• Encode:
  • KeyByLabelValues::to_semicolon_str (data_model/key_by_label_values.rs)
  • extract_group_key (precompute_engine/ingest_source.rs)
  • CMS, HydraKLL and CMS-with-heap accumulators (sketch keys)
  • accumulator_factory.rs (CMS / CMS-with-heap updaters)
  • DeltaSetAggregatorAccumulator::serialize_to_bytes
• Decode:
  • KeyByLabelValues::from_semicolon_str
  • build_group_key_label_values (precompute_engine/worker.rs)
  • CountMinSketchWithHeapAccumulator::get_topk_keys

extract_group_key maps a missing label to "". That matches Prometheus semantics (missing == empty) and is not part of this issue.

Fix

One lossless encoding, defined once on KeyByLabelValues and used by every site above in place of ; join/split. Proposed: a JSON array of the label values (["a;b"] vs ["a","b"]), via serde_json, which is already a dependency.

Tests

• Regression: ["a;b"] / ["a","b"] and [] / [""] encode differently and round-trip exactly.
• Sketch accumulators: two groups that collide today get separate estimates (CMS, HydraKLL, CMS-with-heap top-k keys).
• Precompute engine end to end: a label value containing ; comes back as one label, not two.

Ordering

After #911. #911 keeps the ; encoding so its baselines isolate the sketchlib upgrade; this change moves sketch hashes on its own.