[zzylol]

Part of #147 (Model side).

• MetricFacts gets value_range: Option<(f64, f64)>: smallest and largest positive sample value. validate_facts rejects anything but 0 < lo <= hi < inf.
• For dd rows on a metric with a range, memory per instance is (floor(ln(hi/lo) / ln((1+α)/(1−α))) + 1) × 8 B, with α from sketch_config.params.alpha. This matches dd_footprint in sketch-bench/src/wrappers/dd/mod.rs for positive values (one store, floor + 1 buckets). Without a range, or for other families, the measured mem_bytes_per_instance is kept.
• One helper, instance_memory_bytes(deployment, facts), replaces the direct reads in ingest, merge and storage memory, and in candidate dominance pruning. Pruning compares KLL and DD candidates for the same quantile RAQE, so it has to use the same size. prune_dominated_candidates now takes facts.
• small_problem gives the duration metric a 1 ms–60 s range. DD is undeployable by default, so this only matters with --allow-undeployable-families. Docs updated.

Not done: buckets are also capped by values per instance (dd_footprint's min(count)), but facts don't carry that yet. That's the items-per-instance work in #147. The AutoSketch baseline's own resource ranking still uses the table value. Plans it picks are still scored with the range.

Tests: bucket formula, range vs. fallback (no range / non-DD), and validation. cargo test --workspace, cargo clippy --workspace --all-targets -- -D warnings and cargo fmt --check pass.

🤖 Generated with Claude Code

[zzylol]

Grilling outcome (agreed with @zzylol):

• Q5: the metric-wide range is a conservative upper bound per group; documented.
• Q6: buckets are now capped by values per instance, λ·x/card(G), as dd_footprint caps by count (new test).
• Q7: merge CPU stays measured; the size sweep will cover bucket-dependent merge cost.
• Q8: pruning using the same size, and AutoSketch unchanged, accepted.

🤖 Generated with Claude Code

[zzylol]

Review fix before merge (e0edca1): the merge accumulator holds the whole lookback, so DDSketch's bucket cap there is λ·L/card(G), not one window's λ·x/card(G). instance_memory_bytes takes the span; merge passes L, ingest/storage/pruning pass x. Test: 5 s windows, 10 s query → 10 values → 10 buckets (was 5).

🤖 Generated with Claude Code