datagen: child_of columns for hierarchical labels

[zzylol]

What changed

• aqpbm-datagen: one new column kind. child_of: <index> with fan_out: f nests a string column under an earlier string column: value = parent row's value + . + child index in [0, f). The index is the column's own distribution draw (0-based rank in its domain), so the existing skew/seed options pick it, independently of the parent.
• Refused before drawing: child_of without fan_out or the reverse; child_of naming itself, a later column or no column; a non-string parent or child; fan_out: 0; fan_out disagreeing with the distribution's domain size (or an unbounded distribution); cardinality or a string: block on a child.
• configs/datagen/hydra_hier.yaml: region 4 → service fan-out 25 → endpoint fan-out 25, an independent status label (4, Zipf 2.0), then an i64 value column (user_id, 1e6 Zipf 0.8). Same layout the hydra rows read: label columns (joined with ;) then the value column.
• docs/aqpbm-datagen.md: a child_of section. ColumnSpec gains two optional fields, so the struct literals elsewhere get child_of: None, fan_out: None and the checked-in record schema is regenerated.

Why

Hydra study/eval work (§6.3) needs hierarchical label schemas (schema width d, nested groupings).

Tests

aqpbm-datagen (45 pass): determinism; every child value is <parent>.<i> with i < fan_out; distinct children per parent ≤ fan_out; the child index equals the column's own ranked draw, and Zipf(1.2) head share lands in its expected range; hydra_hier.yaml loads and generates 4 regions / 100 services as parents; one table-driven test for every validation error. Each was checked by mutation (index /2, dropped domain check). cargo fmt --check, cargo clippy --workspace --all-targets -D warnings, cargo test --workspace clean.

Evidence

approxbench sketchbench --variant hydra-cms --library lib --dtype i64 \
  --spec configs/datagen/hydra_hier.yaml \
  --config "rows=3 cols=1024 cell_rows=3 cell_cols=1024" \
  --operations query --metrics accuracy

row	insert throughput (records/s)	query accuracy
hydra-cms / lib	194,462	are_all 35.47, are_top10 0.117, 145,365 probes over 4 regions
hydra-cms / polars	17.8M	are_all 0.0
hydra-hll / lib (rows=3 cols=1024)	212,090	relative_error_mean 0.0177
hydra-hll / polars	15.7M	0.0

(insert numbers from the same command with --operations insert --metrics throughput; release build, 200k records, 15 cell insertions each.)

🤖 Generated with Claude Code

https://claude.ai/code/session_01FhcWJExZcmVqS6r6rEtjis