Does copy_on_write prevent mutation across DataFrame column assignment syntaxes?

The pandas copy-on-write experimental flag aims to isolate in-place DataFrame modifications from shared views, but its behavioral scope across common column assignment patterns remains ambiguous. This question investigates whether enabling the flag consistently prevents data mutation side-effects when assigning new columns or modifying existing ones via different syntaxes, and what performance trade-offs accompany the isolation guarantee.

Key uncertainties include the flag's interaction with existing versus new columns, its applicability to direct versus label-based assignment, and whether the experimental status in pandas 2.2 implies version‑specific quirks that could affect production reliability. Empirical verification across pandas 1.5 through 2.2 releases would be needed to map the full scope of behavior.

Does copy_on_write prevent mutation when assigning to an existing column via direct column assignment? How does the flag affect assignments through label-based indexing when the column already exists? What are the realistic memory and latency impacts for a DataFrame with 1M rows when the flag is enabled?