Do not let the console encoding decide the exit code

[sotashimozono]

The bug

The final print(report) writes to stdout under the platform default encoding, while the report contains em dashes and status emoji. On a console that cannot encode them the print raises UnicodeEncodeError, the script dies, and the gate exits non-zero even when every reference resolved — a clean bibliography is reported as a failed check.

Found on a Japanese Windows shell (cp932) while evaluating this action for a LaTeX paper repository. PYTHONIOENCODING=ascii reproduces it on any platform:

$ PYTHONIOENCODING=ascii python verify_references_gate.py --allow none.allow --report r.md < clean.jsonl
Traceback (most recent call last):
  File "verify_references_gate.py", line 216, in main
    print(report)
UnicodeEncodeError: 'ascii' codec can't encode character '—' in position 52
$ echo $?
1        # every reference was valid

CI is unaffected (Linux, UTF-8), so this only bites people running the gate locally — exactly when they are debugging a bibliography.

The fix

--report and --github-output already open with encoding="utf-8"; stdout now gets the same treatment: reconfigure to UTF-8 with errors="replace", plus a lossy fallback if stdout cannot be reconfigured (replaced by a StringIO in tests, for instance). The exit code then always reflects the reference check rather than the terminal. Report content is unchanged, and under ascii stdout it is still printed in full UTF-8.

Test

The existing smoke job is Linux/UTF-8 only and could not have caught this, so the PR adds an encoding job that runs the gate under ascii, cp932 and utf-8 stdout and asserts both directions — a clean record exits 0 with no traceback, a broken record still exits 1. Verified locally before pushing:

ascii    clean_exit=0 (want 0)  broken_exit=1 (want 1)  traceback=0
cp932    clean_exit=0 (want 0)  broken_exit=1 (want 1)  traceback=0
utf-8    clean_exit=0 (want 0)  broken_exit=1 (want 1)  traceback=0

🤖 Generated with Claude Code