[huangruiteng]

Odd-width packed RGB conversions can read a missing partner pixel past the logical source row. A tight BGR24 row ending at a protected page faults in the scalar reference, and changing row padding can change its U/V result.

The shared C shim now owns a reusable zero-padded input buffer for odd-width RGB24, BGR24, RGBA and BGRA. It copies logical bytes, including negative-stride inputs, and supplies the same zero-valued partner as the existing zero-padded public image buffers. Even-width inputs retain the direct conversion path.

Both work and judge image builds run the same guard-page and padding-independence check after compiling their reference. Task version becomes 1.1.0 because the reference binary and odd-width timing boundary change. Rebuild both task images and identify results by reference version when comparing historical scores.

Validation used real FFmpeg n7.1 static libraries in an isolated linux/amd64 task image. The original shim reproduced SIGSEGV for a tight width-3 BGR24 input. The fixed shim passed tight and colored-padding cases over four packed formats, multiple odd widths and heights, negative strides, and byte-for-byte parity against the original shim for even widths and zero-padded odd widths. All four test methods passed with the previous reference supplied. The normal image-build command executes the two safety methods and skips the two optional legacy-comparison methods.

Both shims compile with the task's build recipe plus -Wall -Wextra -Werror. A sampled process-only comparison under emulation measured fixed/old medians of 1.024 for 128×128 and 1.011 for 129×128. The samples were noisy and do not qualify native-host performance. The complete work and judge images have not been rebuilt in this validation.

The change is independent of the development-oracle work in #59 and the staged failure-feedback work in #58.