perf: use Parquet metadata for row counts in validate command

[shaypal5]

Context

From Copilot review on #16 (COPILOT-5): the validate command currently loads every table/task Parquet fully into memory via pd.read_parquet() even when only row counts or column names are needed.

Problem

For larger bundles this could be slow and memory-intensive.

Proposed solution

• Use Parquet metadata (pyarrow.parquet.read_metadata()) for row counts instead of loading full DataFrames
• For FK checks, read only the required columns via columns=[fk.child_column]
• For leakage checks, read only schema/column names without loading data

Priority

Low — v1 bundles are small (~5K leads), so this is not a blocker. Worth doing before scaling to larger datasets.

🤖 Generated with Claude Code