Problem

In ObjectDetectUDPStage::Process() the name field of each detection was serialised with a hard-coded memcpy of 253 bytes:

std::array<uint8_t, name_length> name;
memcpy(name.data(), detection.name.c_str(), name_length - 2);  // always 253 bytes
name[name_length - 1] = '\0';                                   // only byte 254

Three bugs:

1. Buffer overread: detection.name is typically a short label (e.g. "person", 6 bytes). memcpy always reads 253 bytes, reading well past the end of the string — undefined behaviour.
2. Uninitialised bytes: the array is not zero-initialised, so bytes between name.size() and byte 253 contain stack garbage sent over the UDP socket.
3. GCC 12 false positive: append<std::array<uint8_t,255>> triggers a spurious -Wstringop-overflow / -Warray-bounds diagnostic from vector::resize() inlined deep in stl_algobase.h. The access is entirely within bounds; GCC's alias analysis loses track of the allocation origin.

Fix

• Zero-initialise the array (name{}).
• Copy only min(detection.name.size(), name_length - 1) bytes.
• The NUL terminator is already present from zero-initialisation; no explicit name[254] = '\0' needed.
• Suppress the GCC 12 false positive with a narrow #pragma GCC diagnostic pair around the append template definition, with a comment explaining the root cause.

The on-wire format is unchanged — the receiver still reads exactly name_length (255) bytes for the name field.

Testing

Compiled with GCC 12.2.0, -O3 -Wall -Wextra -Warray-bounds=2 (OpenCV enabled) — no warnings.