Fix non-atomic file writes that can corrupt cached data

[h0bnobs]

Description

Fixes #127.

Writes to the parquet price cache, stock_info.json and pickled model files were made directly to the final path. If the process crashed mid-write (OOM, SIGKILL, disk full), the file was left in a corrupted partial state, and the next run would either fail to read it or silently load malformed data.

This PR adds an atomic_write context manager in src/util/atomic_io.py that writes to a temporary file in the destination directory and atomically renames it into place with os.replace. On failure the temp file is removed, so an existing file is never clobbered by a partial write. Readers therefore only ever observe either the old complete file or the new complete file.

The issue referenced a df.to_csv() call that has since become df.to_parquet(), but the underlying non-atomic write remained and is fixed here.

Changes in this pull request

• Add atomic_write context manager in src/util/atomic_io.py, exported from src.util
• Use atomic writes for the parquet price cache in StockDataDownloader._download_single
• Use atomic writes for stock_info.json in StockDataDownloader.save_stock_info
• Use atomic writes for pickled models in Model.save_model and StackedStockPredictor.save_model
• Add unit tests for atomic_write covering success, replacement, failure rollback and temp-file cleanup
• Add a regression test that a failed save_stock_info preserves the existing cache file

🤖 Generated with Claude Code