fix: keep the AI analysis from crashing on a non-UTF-8 stdout - #3096

Third of the same shape as #3050 (startup banners) and #3056 (result lines, and the spinner). The one terminal write in the AI module that was not covered is the analysis itself.

The defect

_stream_response() writes each token of the model's reply straight to stdout:

sys.stdout.write(content)
sys.stdout.flush()

The reply is arbitrary text. An OSINT run on an international username routinely comes back with Cyrillic or CJK. On Windows, Python takes stdout's encoding from the process ANSI codepage — cp1252 on a default install — so that write raises UnicodeEncodeError, the exception leaves the coroutine, and the analysis ends mid-sentence with the tokens already printed left on screen. PYTHONIOENCODING cannot rescue the PyInstaller build, which ignores PYTHON* variables.

surrogateescape is not a guard here, which is the part worth pinning. It is what sys.stdout actually carries, and it rescues lone surrogates only — an ordinary unencodable character still raises:

>>> "Привет".encode("cp1252", "surrogateescape")
UnicodeEncodeError: 'charmap' codec can't encode characters in position 0-5

test_raw_write_to_a_cp1252_stdout_is_what_fails asserts that, so the suite notices if the platform ever changes underneath the fix. The fix itself reuses the shape from the two merged PRs — notify._print_encodable — as a stream write rather than a print.

A correction to something I wrote

The _frames_for docstring, which I added in #3056, says the braille frames "raised UnicodeEncodeError inside the spinner's daemon thread: the thread died with a traceback over the output".

They did not. I only found this while probing upstream main for this PR. Unlike stdout, sys.stderr's error handler is backslashreplace, and PYTHONIOENCODING does not change it:

So the spinner never crashed — the animation just became a column of literal escapes. The fix was right; the reason I gave for it was not, and the difference between the two streams is exactly what this change turns on. The docstring now says what actually happens.

Left alone

print_streaming() writes to stdout the same way and has the same defect, but it has no callers anywhere in the tree. I would rather not change code there is no way to exercise; happy to include it if you would prefer.

Verification

6 tests added to tests/test_ai.py, driven by real cp1252 TextIOWrapper streams rather than mocks, so the codec decides. Checked by breaking it:

The last test exists because that third mutation survived my first draft: every other case here uses text cp1252 either carries unchanged or cannot carry at all, and both paths give the same bytes for those. A backslashreplace stream is the one that separates them.

Full suite on Windows: the set of failing test names is identical before and after (35 pre-existing, plus tests/test_maigret.py, which cannot be collected on Windows at all — os.geteuid does not exist there). flake8 clean.