Complete Part 2: audit and ingest xRouteBench

[MaharshPatelX]

Summary

• pin ulab-ai/xRouteBench at immutable revision ea4b6e1b29d9a734f55f0a637baf326bad6aa681
• add a 41-file integrity manifest and exact 17-config schema/missingness snapshot
• add a checksum-, schema-, dtype-, and row-count-validating Parquet adapter
• add a safe-by-default verified downloader and synthetic test fixture
• publish the Part 2 audit and update data policy, repository docs, and project tracker

Audit findings

• 17 configs and 41 Parquet files
• 245,903 physical rows across all configs
• 47,382,959 bytes (47.38 MB / 45.19 MiB)
• all 41 local SHA-256 hashes and row counts match the pinned upstream source
• source is public and ungated, but the pinned card has no declared dataset license and the repository has no LICENSE file
• no upstream rows are committed; the tracked fixture is authored synthetic data
• the dataset-card prose says 4,487 generic training queries, while the pinned file and Hub APIs contain 4,489; the manifest records the file-backed value

Validation

• ruff format --check .
• ruff check .
• mypy src tests
• pytest — 11 passed
• python -m build
• both manifest JSON files parsed successfully
• the downloader verified the pinned 8,450-byte sample
• the real 18-row candidate-model Parquet sample loaded identically with two batch sizes

Scope boundary

This PR completes Part 2 only. It does not implement canonical Part 3 tables, deduplication, leakage-safe splits, baselines, model training, or benchmark claims.