Git(path=...).run(["diff", ...]) rewrites every byte that is not valid UTF-8 into the literal four-character text \xNN: captured output funnels through console_to_str, whose fallback is s.decode("utf_8", errors="backslashreplace"), and every public command surface is typed -> str. backslashreplace is not round-trippable, so the original bytes are unrecoverable, and no bytes-returning API exists. A captured diff of latin-1 content is no longer an applicable patch — git apply rejects it (patch does not apply); a blob read via show cannot round-trip at all.
Expected: captured output is byte-faithful to what git wrote — the same output-fidelity contract as #536, one layer down. (The #538 trim fix holds on 0.45.0: whitespace and the trailing newline survive; the decode is the remaining gap.)
Reproducible example: https://github.com/k0pernikus/libvcs-non-utf8-on-diff
uv run pytest (or mise run test) — one test asserts the subprocess git diff output round-trips the patch byte-exact (passes), one asserts the libvcs output against the same expected patch (fails: every 0xE9/0xE8 byte arrives as the literal text \xe9/\xe8).
run() already keeps the pipe in bytes mode internally (text=False, # Keep in bytes mode to preserve \r properly) — fidelity is lost only at the return boundary. A bytes-mode surface, or an errors=/encoding= knob honored at the final decode, would close this.
Environment: libvcs 0.45.0, Python 3.15, any recent git on a UTF-8 locale
Git(path=...).run(["diff", ...])rewrites every byte that is not valid UTF-8 into the literal four-character text\xNN: captured output funnels throughconsole_to_str, whose fallback iss.decode("utf_8", errors="backslashreplace"), and every public command surface is typed-> str.backslashreplaceis not round-trippable, so the original bytes are unrecoverable, and no bytes-returning API exists. A captured diff of latin-1 content is no longer an applicable patch —git applyrejects it (patch does not apply); a blob read viashowcannot round-trip at all.Expected: captured output is byte-faithful to what git wrote — the same output-fidelity contract as #536, one layer down. (The #538 trim fix holds on 0.45.0: whitespace and the trailing newline survive; the decode is the remaining gap.)
Reproducible example: https://github.com/k0pernikus/libvcs-non-utf8-on-diff
uv run pytest(ormise run test) — one test asserts thesubprocessgit diffoutput round-trips the patch byte-exact (passes), one asserts thelibvcsoutput against the same expected patch (fails: every0xE9/0xE8byte arrives as the literal text\xe9/\xe8).run()already keeps the pipe in bytes mode internally (text=False, # Keep in bytes mode to preserve \r properly) — fidelity is lost only at the return boundary. A bytes-mode surface, or anerrors=/encoding=knob honored at the final decode, would close this.Environment: libvcs 0.45.0, Python 3.15, any recent git on a UTF-8 locale