Keep calibration and held-out samples disjoint

[SloptimistPrime]

The held-out split shuffled repeated rows, allowing bundled seeds to appear in both files.

Deduplicate content across domains before splitting, while preserving the original formatting. Report the smaller unique sample count instead of treating repeated text as new data. Reject splits with too little unique data and output paths that refer to the same file. Calibration-only runs keep their previous sample counts.

21 tests pass for repeated text, Unicode and whitespace variants, cross-domain overlap, determinism, invalid arguments and file aliases.

This prevents exact normalized sample overlap, not all near-duplicates or overlap with the model's pretraining data.

I used AI assistance and reviewed and tested the changes.