Fuzz harness: parquet writer-shape generation #27

Part of #22. Depends on the harness skeleton.

Generate parquet files with varied physical layout rather than querying one fixed file. Today the input space is examples/nano_hits.parquet plus a handful of targeted files added alongside recent fixes (batch_size_alignment.rs:37, page_index.rs) — all written by the same writer, and none produced systematically.

Properties to vary

• Row group size (including files with many small row groups, so a scan crosses boundaries mid-batch)
• Data page row count limit — a small limit gives the page index enough granularity to prune within a row group, which is what produces a row selection starting partway in
• Dictionary encoding on/off
• Per-column statistics level
• Writer version
• Compression codec
• Null density, including all-null and no-null columns
• Single file vs multiple files in a listing table

write_parquet in batch_size_alignment.rs:37-56 is the right starting shape — it already takes optional row group and page row limits.

Note on scope

This dimension is a weaker fit for #10 than it looks. That bug needed a file advertising a page index while one column chunk carries no offset index, which is a shape the Arrow writer does not readily produce. A corpus of files written by other tools is the more reliable answer to that class and is tracked separately.

Acceptance

• File properties are part of the seeded generation and appear in the reproduce output.
• The report counts iterations by row-group and page-limit configuration, so a run that only ever generated single-row-group files is visible rather than silent.