feat: self-contrast for rows without augmentation

[EYH0602]

Summary

• Fix: Instead of dropping rows that have no code transformation (transformed is None), use self-contrast — copy the original code into the transformed field so the model sees two views of the same code with different MLM masks. This creates "easy" contrastive pairs that act as implicit curriculum learning alongside the harder invariant-contrast pairs.
• Config flag: Add self_contrast (default true) to PretrainConfig and both CLI subcommands (run --no-self-contrast, pretrain --no-self-contrast) to allow disabling the behavior.
• Experiment configs: All supcon/ and grouped/ YAML configs updated to use the full dataset (data/csn.jsonl) with self_contrast: true.
• Documentation: Added doc/self-contrast.md with analysis of why self-contrast is feasible and beneficial (MLM masking as augmentation, implicit curriculum, data efficiency, SupCon compatibility).

Motivation

Transformation operators can fail on certain code samples (e.g., no for-loops to convert, no local variables to rename). Previously these rows were silently dropped, wasting a large fraction of the dataset. Self-contrast retains every row by leveraging independent MLM masking as a natural data augmentation — analogous to SimCSE's dropout-as-augmentation strategy.

Changes

File	Change
modeling/pretrain.py	Replace dataset.filter with dataset.map that fills None → original code (when self_contrast=True)
modeling/config.py	Add self_contrast: bool = True to PretrainConfig
modeling/cli.py	Add --self-contrast / --no-self-contrast to both run and pretrain subcommands
experiments/**/*.yaml	Switch dataset to data/csn.jsonl, add self_contrast: true
doc/self-contrast.md	Design doc explaining the approach

Test plan

• Run pre-training with self_contrast: true (default) and verify no rows are dropped
• Run with --no-self-contrast and verify rows without augmentation are filtered out
• Compare training loss curves between self-contrast and no-self-contrast
• Evaluate downstream task performance (clone detection, defect detection) for both settings