v0.2.1

fix: harden AI cleanup prompt against reply-style output and tone drift

- dictated text wrapped in delimiters, explicit never-respond instruction
- default to MESSAGE on ambiguity; temperature lowered to 0.2
- grammar fixes now strictly minimal-edit, preserving the speaker's tone
- strip wrapping quotes from model output
- bump version to 0.2.1