[tryu-ai]

read_images_binary() decodes image names one byte at a time, so valid UTF-8 names such as café.png or 场景.png raise UnicodeDecodeError. Accumulate the NUL-terminated name bytes and decode the complete name once, preserving strict UTF-8 error handling.

COLMAP writes the complete name followed by NUL (official writer). Regression tests cover ASCII, two-/three-/four-byte UTF-8, mixed Chinese/emoji names, four consecutive records with varying point counts and aligned coordinates/IDs, and invalid UTF-8 rejection.

Validation on macOS arm64 / Python 3.9.6:

• New regression tests: 5 failed / 4 passed before the fix; 9 passed after.
• ruff check and ruff format --check: passed.
• CPU test subset including the new tests: 496 passed.
• Full pytest tests is blocked by missing Playwright. The non-E2E suite is blocked by missing client build/nodeenv; the complete suite was not run successfully.

No GPU tests or real reconstruction datasets were used.