[kvlonge]

Why

Command-output capture split any line over 64 KiB that contained non-ASCII text one character at a time in Python. Large --json compile reports with non-ASCII SQL literals spent about 0.3–0.9 s there, on full and reused compiles alike (CHI-504).

Changes

• Encode non-ASCII text to UTF-8 once and cut chunks on character boundaries by stepping back over continuation bytes; the ASCII fast path and the fallback for lone surrogates are unchanged.
• Chunk sizes, boundaries, rejoined text and raised exceptions are identical to before.

Verification

• Chunking 10 MB of non-ASCII text: 875 → 22 ms. A 120-model DuckDB compile with a 9.8 MB report: 2.67 → 1.74 s wall (full) and 2.42 → 1.46 s (reused).
• Randomised equality test against a copy of the previous implementation, boundary cases and a speed guard; the reviewer also fuzzed 200,000 strings against main with zero mismatches.
• One independent review; make test, make check-ci.