Fix UnicodeEncodeError on startup when stdio uses a non-UTF-8 encoding

[LHMQ878]

Fixes #1436

Problem

On Windows, launching kimi from Git Bash dies immediately:

'gbk' codec can't encode character '▐' in position 0: illegal multibyte sequence

▐ is ▐, from the welcome banner logo in ui/shell/__init__.py:

▐█▛█▛█▌  Welcome to Kimi Code CLI!
▐█████▌  Send /help for help information.

The banner is rendered through the shared console singleton in src/kimi_cli/ui/shell/console.py, which inherits whatever encoding the standard streams have. On Windows those default to the system locale encoding (GBK/cp936 on Chinese installs), which cannot represent the box-drawing glyphs, so the very first console write raises UnicodeEncodeError and the CLI exits before it starts.

The same singleton also prints ✓ (U+2713) in ui/shell/oauth.py and ui/shell/setup.py, so logout and setup were exposed to the identical failure.

Fix

Reconfigure sys.stdout / sys.stderr to UTF-8 when the shell console module is initialized.

I first tried leaving the locale codec in place and only adding errors="replace". That stops the crash but renders the banner as partial-byte garbage (?��?��?����), because GBK is multibyte and replacement happens per unencodable character. Forcing UTF-8 keeps the glyphs intact, and I verified on a GBK-default Git Bash that a UTF-8 byte stream displays ▐█▛█▛█▌ correctly. errors="replace" is kept as a last resort for anything UTF-8 still can't carry (e.g. lone surrogates).

Streams that don't expose reconfigure(), or that reject it (detached/closed), are skipped rather than failing startup.

Verification

Environment: Windows 11, Git Bash, sys.stdout.encoding == 'gbk', locale.getpreferredencoding() == 'cp936' — i.e. the configuration from the report.

• Reproduced the crash first: rendering the banner through a GBK-backed Rich console raises UnicodeEncodeError: 'gbk' codec can't encode character '▐'.
• With the fix, stdout.encoding becomes utf-8, and both the banner panel and the ✓ message render correctly, in a terminal and with stdout redirected to a file.
• Confirmed the crash returns if the new call is removed, so the test is exercising the fix rather than an unrelated change.
• Added tests/ui/test_console_encoding.py (6 tests, all passing): reconfiguration of both streams, tolerance of streams without reconfigure(), tolerance of OSError/ValueError from it, plus a guard asserting the banner glyphs genuinely aren't GBK-encodable so the premise can't silently rot.
• ruff check and ruff format --check pass on both changed files.

Note on overlap with #2560

I have a separate PR (#2560) fixing Fixes #2532, the same class of bug in utils/server.py's print_banner, which covers the kimi web and kimi vis startup paths. That one uses plain print(), so it needed a different fix. This PR covers the interactive shell path and is independent — neither depends on the other, and they touch different files.

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.