fix: preserve UTF-8 across shell output chunks

[yunyv]

Summary

Preserve UTF-8 characters when shell output splits them across reads. ShellManager._drain_stream() currently decodes each 4096-byte chunk independently with errors="replace". A valid multibyte character spanning two reads becomes replacement characters, so shell tool output can corrupt non-ASCII text and filenames.

Use a separate incremental UTF-8 decoder for each stdout/stderr reader, and flush it at EOF. Invalid bytes and incomplete trailing sequences still use replacement decoding. There are no new dependencies or changes to the shell API.

Reproduction

Run a shell command that writes b"a" * 4095 + "中".encode("utf-8") to stdout or stderr. On the base revision, the final character is returned as replacement characters. With this change, the original text is preserved.

Tests

• Added real subprocess regressions for two-, three-, and four-byte characters crossing the read boundary on both stdout and stderr: all six fail against the original implementation and pass with the fix.
• Added controls for empty output, invalid bytes, and an incomplete UTF-8 sequence at EOF.
• make check: passed, including lock validation, all pre-commit hooks, and mypy (57 source files).
• Full pytest suite with doctests: 594 passed, 20 skipped without optional tracing dependencies. A writable temporary HOME was used because this execution environment's default home is read-only.
• With the optional trace extra installed: HOME=/tmp/bub-test-home OTEL_TRACES_SAMPLER=always_on .venv/bin/python -m pytest --doctest-modules: 613 passed, 1 skipped (PowerShell unavailable). The sampler override prevents the host’s inherited sampling setting from disabling the in-memory tracing assertions.
• Python 3.12 tested locally; the repository CI covers Python 3.13 and 3.14.