fix(output): write --json as UTF-8 regardless of the console codepage (#546)

[padak]

Fixes #546.

The bug

kbagent --json flow list crashed with UnicodeEncodeError on a default
Czech/Polish/Hungarian Windows console (cp1250) whenever the output contained a
non-ASCII character. An arrow in a flow name was enough:

File "keboola_agent_cli/output.py", line 57, in output
    sys.stdout.write(response.model_dump_json(indent=2) + "\n")
UnicodeEncodeError: 'charmap' codec can't encode character '→'

--json exists to be piped into another program, so its bytes must not depend on the
terminal's active codepage.

Which writer actually crashes

Worth being precise, because it decides what the tests are worth:

Writer	Non-ASCII handling	Crashes on cp1250
response.model_dump_json() (pydantic) — output(), success()	emits raw UTF-8	yes
json.dumps(..., indent=2) — error()	escapes to \uXXXX (ensure_ascii default)	no

So only the two pydantic paths reproduce the report. The error envelope is routed
through the same helper anyway for consistency, and its test says so plainly rather
than posing as a regression test for a bug it cannot hit.

The fix

One write_machine_output helper behind all three call sites. It writes UTF-8 through
sys.stdout.buffer, bypassing the text layer's encoder entirely, so the codepage never
gets a say. The text layer is flushed first so anything already written through it keeps
its place in the stream; a stdout with no binary buffer (captured or replaced streams,
e.g. under test) has no encoder to bypass and takes the plain write.

Chosen over sys.stdout.reconfigure(encoding="utf-8") because reconfigure is not
available on every stream kind the CLI runs under, and over PYTHONUTF8=1 because the
reporter should not have to set an environment variable to get valid JSON.

Testing

Five tests in tests/test_output.py::TestMachineOutputIsAlwaysUtf8. They simulate the
cp1250 encoder with io.TextIOWrapper(io.BytesIO(), encoding="cp1250"), so they run on
every platform rather than needing a Windows box.

I verified they are load-bearing by reverting the fix: the two covering the crash
(output, success) fail with the reporter's exact UnicodeEncodeError; the other
three pass either way and are documented as invariant guards, not bug reproductions.

Also covered: the no-binary-buffer fallback, and that a human-mode write issued before a
machine-output write keeps its order (the reason for the flush).

make check equivalent is green — ruff, ty, and 4688 tests.

Reproduced end to end through the real CLI as well, not just the formatter. Forcing the
console encoding and asking for output that contains box-drawing characters (the v0.76.1
changelog entry):

PYTHONIOENCODING=cp1250 kbagent --json changelog --limit 8 --full

• before the fix: exit=1, UnicodeEncodeError: 'charmap' codec can't encode characters in position 4481-4482 — the reporter's failure;
• after: exit=0, output parses as UTF-8 JSON with ─ └ ├ intact.

(Worth noting for anyone repeating this: a small --limit proves nothing here. The most
recent changelog entries happen to be pure ASCII, so the command exits 0 either way.)

Notes for merge

• Side effect on Windows: JSON lines now end LF rather than CRLF, because the
  binary buffer does no newline translation. For machine-consumed output that is the
  better default and no parser cares, but it is a real behaviour change worth knowing.

• No version bump or changelog entry, matching how fix(update): make self-update safe on Windows #530 / fix(semantic-layer): support snapshot export on Windows #531 landed before the
  chore(release): 0.76.2 Windows update and export fixes #533 release PR bundled them. The line for whichever release picks this up:

  Fix (#546): kbagent --jsonno longer crashes withUnicodeEncodeErroron Windows consoles with a non-UTF-8 codepage (cp1250 on Czech/Polish/Hungarian Windows). Any non-ASCII character in the data — an arrow in a flow name was the report — made machine-readable output unusable, because pydantic'smodel_dump_jsonemits raw UTF-8 andsys.stdoutencoded it through the console codepage. JSON is now written as UTF-8 bytes throughsys.stdout.buffer, independent of the console. The PYTHONUTF8=1 workaround is no longer needed. Thanks to @MichalProchazka for the report.

• Independent of fix(update): defer the Windows self-update out of the running environment (#528) #543 (the Windows self-update fix); different subsystem, no overlapping
  files.

[devin-ai-integration]

Devin Review found 2 potential issues.