[zzylol]

PR 5 of the Hydra implementation spec, stacked on #199 (hydra/target-grouping).

What changed

• hydra-kll-cdf (lib, polars): new variant, registered the same way the hydra-univmon-* rows register one row per statistic. It asks the KLL grid HydraQuery::Cdf(x) per subpopulation. The thresholds x are the group's own values at φ = 0.01..0.99 (99 points, index floor(φ·n)). It is scored by absolute CDF error |F̂_q(x) − F_q(x)| through a new comparator, SubpopCdfErrorGT (subpop-cdf-error, capability subpop-cdf), which reuses the rank-error truth (sorted values per group). The polars baseline takes the exact share at or below x over the same sorted groups. Per-group output from hydra: score any target grouping; per-group error output #199 (err_*, groups_scored, --per-group-out) applies.
• hydra-univmon-sum (lib, polars): each record is inserted with count = Some(value) (new hydra_shared::update_counted), so the UnivMon grid's L1 is the group's sum. It is scored by relative error against exact per-group sums: new SubpopSumGT (subpop-sum), and the polars baseline is group_by(subset).agg(v.sum()). It builds at i64/u64 only. --dtype f64 or str is refused by name, and so is a value outside [0, i32::MAX] (the library's weight is an i32). The univmon drivers now take a feed fn so counted and weighted inserts share one code path.
• KLL footprint: memory_hydra_kll now uses kll_lib_bytes::<f64>(cell_k) per cell, so each cell's merge buffer of cell_k items is counted, as it is for the kll-* lib rows. At rows=3 cols=1024 cell_k=200 this moves the footprint from 26,984,704 to 31,899,904 bytes (+1,600 B per cell).
• Docs: README (Hydra flags section), docs/concepts.md capabilities, and the hydra datagen yaml headers.

Not done: hydra-hll cell_lg_k

asap_sketchlib 0.3.0 defines HydraCounter::HLL(HyperLogLog<ErtlMLE>), and HyperLogLog is an alias for HyperLogLogP14. P12 and P16 exist as standalone types but can't go in a Hydra cell (this is true of every sketchlib checkout I have too). A cell lg_k sweep needs an upstream change (a generic HLL cell), so it is out of scope here. hydra-hll stays at P14.

Tests (each checked by mutation)

• aqpbm-core: sum truth adds the values (40/90 on the fixture, n_q = records), refuses an f64 value column by type; the CDF truth counts ties at or below (F_a(10) = 2/3), the exact answer scores 0, a zero answer misses by the share.
• sketch-bench: KLL footprint is hand-derived as (1004 + 200)·8 + 62·8 per cell (fails without the buffer). The lib Cdf below k is exact inside the group (fails if asked as Quantile). The polars CDF baseline gives the share at or below (fails with <). Weighted L1 equals the exact sums (45, 7, 42) on a non-colliding 3×1024 grid (fails if inserted unweighted). Values i32::MAX+1 and -1 are refused by name. The polars sum baseline answers 40/30.
• aqpbm-cli/tests/hydra_grouping.rs: hydra-kll-cdf and hydra-univmon-sum were added to the any-subset test (polars err_max = 0, lib ≤ 0.01 on a big grid, for groupings 1 and 0,1). A new test checks that --dtype f64 is refused by name at lib and polars.
• cargo fmt --check, cargo clippy --workspace --all-targets -D warnings, cargo test --workspace: clean.

Evidence (release build, --operations query --metrics accuracy,memory --runs 1)

hydra-kll-cdf, configs/datagen/hydra_columns_f64.yaml (200k rows), rows=3 cols=1024 cell_k=200:

group cols	lib err_mean / p50 / p90 / max	polars err_max	groups	lib bytes
0	0.0040 / 0.0036 / 0.0064 / 0.0163	0	200	31,899,904
0,1	0.130 / 0.096 / 0.261 / 0.82	0	9,950	31,899,904

hydra-univmon-sum, configs/datagen/hydra_columns.yaml (200k rows), rows=3 cols=1024 cell_heap_size=64 cell_sketch_row=3 cell_sketch_col=256 cell_layer_size=4 (relative error):

group cols	lib err_mean / p50 / p90 / max	polars err_max	groups	lib bytes
0	0.332 / 0.212 / 0.895 / 1.91	0	200	114,327,808
0,1	500.9 / 32.5 / 635 / 84,343	0	9,950	114,327,808

For comparison, the unweighted hydra-univmon-l1-norm at the same config with 0 gives err_mean 0.330 and err_max 1.96. At 0,1, about 10.2k subkeys per row land in 1,024 columns, so small groups take collision mass from heavy ones. That is the saturation the study (PR 6) maps; the exactness test above shows the rule holds on a grid without collisions.

--dtype f64 → hydra-univmon-sum/lib counts each value by itself, an integer weight, so builds at i64 or u64; --dtype f64 is not an integer.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FhcWJExZcmVqS6r6rEtjis