edit and apply_patch silently corrupt non-UTF-8 files (U+FFFD replacement)

Describe the bug

edit and apply_patch decode file bytes with a non-fatal UTF-8 decoder (new TextDecoder("utf-8") and new TextDecoder("utf-8", { ignoreBOM: true })). Invalid sequences are silently replaced with U+FFFD.

When oldString matches an ASCII region of a file that contains non-UTF-8 bytes elsewhere (GBK/Latin-1 comments, legacy strings, stray bytes), the tool reports success and writes the entire file back as U+FFFD, permanently destroying every non-UTF-8 byte.

The read tool already rejects this with MalformedUtf8Error (File is not valid UTF-8) via a fatal decoder, but edit/apply_patch don't — so a file that read refuses is silently corrupted by edit.

Steps to reproduce

1. Create a file with bytes const x = 1\n followed by GBK bytes d6 d0 ce c4 (中文).
2. edit oldString: "const x = 1" → newString: "const y = 1".
3. The file is written back with the GBK bytes replaced by U+FFFD, and the tool reports success.

Affected code

• packages/core/src/tool/edit.ts — decodeUtf8 uses new TextDecoder()
• packages/core/src/tool/apply-patch.ts — new TextDecoder("utf-8", { ignoreBOM: true })

Suggested fix

Decode with { fatal: true } and fail with an explicit "file is not valid UTF-8" error, leaving the file untouched (matching read).

Environment

• opencode version: latest dev