[GordonYuanyc]

Closes the sketch-bench side of the characterization gaps: G-6, G-3, G-7, the Hydra per-group statistics, and G-8 for every non-windowed SketchAnalytics type except SetAggregator.

What changed

• G-6 seeded KLL. kll and hydra-kll take an optional seed knob (--config 'k=200 seed=7'). With it the lib rows build with with_k_and_seed; without it they build as before, unseeded. The oxide KLL takes no seed, so it refuses one. KllDynamic, Coco, UniformSampling and Nitro take the same optional seed.
• G-3 quantile probes. --quantiles 0.5,0.99 asks a rank-error row at those fractions and reports rank_err_q* for each one. Without the flag the 101-point grid stays.
• G-3 stream statistics. The frequency, top-k and heavy-hitter comparators report stream_items, stream_distinct and stream_l2.
• G-7 top-k count error. The top-k comparator reports max_count_err and mean_count_err over the keys the sketch returned.
• Per-group statistics. --per-group-out writes group_key,n_q,error,d_q,l2_q. The flag is gated on a subpopulation capability instead of the hydra- name prefix.
• New membership comparator for Bloom. It probes up to 10k members and 100k keys that are not in the stream, then reports fpr, fnr (always 0 for a Bloom filter) and stream_distinct.
• G-8: 29 new lib rows, written out per type like the existing wrappers:

Type	Rows	Comparator
HyperLogLog<Ertl>	hll-ertl	cardinality
Kmv	kmv	cardinality
Bloom	bloom: cols must be a power of two, where the FPR bound holds	membership
SpaceSaving	space-saving, space-saving-topk	frequency, topk
CountMinSketchWithSpaceSaving	count-min-space-saving, count-min-space-saving-topk	frequency, topk
CountSketchWithSpaceSaving	count-sketch-space-saving, count-sketch-space-saving-topk	frequency, topk
CountMinSketchFold	count-min-fold, count-min-fold-topk	frequency, topk
CountSketchFold	count-sketch-fold, count-sketch-fold-topk	frequency, topk
CountSketchWithL2Norm	count-sketch-l2-norm	keyed-l2-norm
CountMinHll	count-min-hll: each record goes into every label subset	subpop-cardinality
ElasticSketch	elastic, elastic-topk: --dtype string only	frequency, topk
CocoSketch	coco: --dtype string only	frequency
KllDynamic	kll-dynamic (seeded)	rank-error
UniformSampling	uniform-sampling	rank-error
UnivMonTwoTier	univmon-two-tier-{cardinality,l1-norm,l2-norm,entropy}	keyed-*
UnivMonQ	univmon-q-{quantile,cardinality,frequency,topk}: all but quantile at --dtype f64 only, since UnivMonQ keeps values as f64	rank-error, cardinality, frequency, topk
NitroBatch	nitro-count-min, nitro-count-sketch: i64 only, the whole stream in one batch, no per-item latency	frequency

The top-k k is checked only when a top-k question is asked, so a frequency row builds at any capacity. topk_row takes k from the row's own wrapper.

Not included

• SetAggregator: it is exact, so its error is always 0.
• The windowed types: MicroCm, ExponentialHistogram, EhBucketSketch, ExponentialHistogramUnivMon and TumblingWindow.
• docs/sketch-bench.md is not updated with the new rows or flags.

Verification

• cargo fmt --check, cargo clippy --workspace --all-targets --all-features -D warnings and cargo test --workspace all pass.
• Each new row ran query and merge accuracy through the release CLI, and every error was in range.
• Bloom with r=4, c=65536 and D=27096 measured an FPR of 0.0136; the formula (1 − (1 − 1/c)^D)^r gives 0.0132.

🤖 Generated with Claude Code