feat: expose kernel codec accessors and edge sampling

[zzylol]

Stacked on #488.

Why

ASAPQuery-backend #805 stores Planner kernel states directly and keeps only storage codecs. Five gaps in the kernels forced workarounds there: a copy of the UnivMon accumulator, a msgpack round trip to reach the weighted-frequency kernel, a hand-written mirror of the exact state's serde shape, a clone of each cached sketch per ingest frame, and rejecting every edge-sampled frame (0 < sample_p < 1) because kernels had nowhere to keep p.

What

Each new public API exists for one backend need:

API	Backend need
UnivMonAccumulator::from_sketch(UnivMon), sketch(), AggregateCore::estimate for UnivMon	Delete the asap_summary_state::univmon shim; encode/restore with sketchlib's codec
WeightedFrequency::algorithm(), shape(), to_bytes(), from_bytes()	Drop the msgpack round trip in the codec
ExactAccumulator deserialization validates against its family (serde(try_from))	Drop decode_exact's shape mirror; no new method
AggregateCore::as_any_mut	Update cached sketches in place on ingest deltas and read-path merges
from_sketch(sketch, sample_p), sample_p(), merge_sample_p(p) on DDSketch, HLL, Count-Min, Count Sketch	Accept sampled edge frames again, persist p, record p of deltas applied in place

How: sample_p

p is kept in a private field, so it is always in (0, 1]. from_sketch rejects 0, values above 1, and NaN. The backend should map the proto3 default 0 to 1 before calling it.

Kernel	Scaled by 1/p	Unscaled
DDSketch	bare PointCount (total count)	quantiles
HLL	Cardinality, bare PointCount	none
Count-Min, Count Sketch	query_key	none

These rules match the kernels at 76fbbf1 for DDSketch, HLL and Count-Min. Count Sketch had no sample_p there; it gets the same 1/p rule because it is linear.

Merge rules:

• Equal p values merge and keep that p.
• 1 merged with a sampled p gives the sampled p. This is the old "prefer sampled" rule, which exists because window-reset and Planner-created bases are unsampled.
• Two different sampled p values are rejected. The old code silently kept the left operand's p here.

Before this PR

The backend decodes SketchEnvelope{CountMinState, sample_p: 0.25} and fails with "frame is edge-sampled … not supported by Planner kernels".

After this PR

let cms = CountMinSketchAccumulator::from_sketch(decoded, 0.25)?;
cms.query_key(&key)          // stored 10 reads as 40
cms.sample_p()               // 0.25, persisted next to the sketch bytes

Validation

• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings, and cargo test --workspace --no-fail-fast all pass.
• Written first, these tests failed on the base and pass now: exact decode rejecting mismatched or unsupported families, and the UnivMon readouts.
• Ported from 76fbbf1: DDSketch raw 10 at p=0.1 reads 100 with quantiles unchanged; HLL reads exactly 4× raw at p=0.25; Count-Min reads 4× raw at p=0.25. Each also has a merge-policy test.
• New tests cover round trips for the UnivMon adoption and the weighted-frequency bytes, rejection of terminal-mode and invalid UnivMon sketches, and in-place as_any_mut updates.
• An independent reviewer agent reviewed the change. Its finding that deltas applied in place could not record p led to merge_sample_p.

Backend follow-ups

• sample_p is private, so the backend's struct literals must switch to ::new or from_sketch.
• NativeSummaryOutput must implement as_any_mut.
• Stored bytes must carry p next to inner.to_msgpack(). Until they do, keep rejecting sampled frames.
• Count reads made directly on inner (e.g. inner.estimate(&str) or row totals) are unscaled. Use query_key(&KeyByLabelValues::new_with_labels(vec![item])) or divide by sample_p().
• The merge rule assumes p is fixed per series. An adaptive p that changes mid-series gets an error for two different sampled values.

🤖 Generated with Claude Code