Fix repetition penalty forwarding in data generation

[Sehlani042]

Summary

• forward --repetition-penalty through SGLang's extra_body
• keep it distinct from OpenAI presence_penalty
• reject non-positive repetition penalty values
• add focused regression coverage

Root cause

generate_train_data.py mapped the CLI repetition penalty to presence_penalty. These are different sampling controls: SGLang exposes repetition_penalty as an extra request field, while presence_penalty changes token-presence behavior. The CLI therefore silently requested a different sampling policy than the user selected.

Validation

• deterministic regression tests failed before the fix and pass with this patch
• env PYTHONPATH=. uv run --with pytest --with openai pytest -q tests/test_generate_train_data.py (3 passed)
• uvx ruff check tests/test_generate_train_data.py
• uvx ruff check --ignore BLE001 scripts/data/generate_train_data.py tests/test_generate_train_data.py (the unignored source-file scan reports three pre-existing BLE001 findings outside this diff)
• python -m py_compile for both changed Python files
• git diff --check origin/main...HEAD

AI assistance

This change was developed with Codex assistance. I reviewed the diff and validation evidence before publication.