When MxFp4Codec::QuantizeBlock encounters an all-zero input block, the early-out path zeroes kBlockSize (32) bytes instead of the encoded block size sizeof(MxFp4Stream) (17 bytes: 1 scale byte plus 16 nibble bytes). Writing 15 bytes past the end of the block corrupts the adjacent block when blocks are encoded out of order (such as across parallel batch boundaries in Compress or via packed_ofs sub-ranges) and writes past PackedEnd on the final block of a tensor.

[copybara-service]

When MxFp4Codec::QuantizeBlock encounters an all-zero input block, the early-out path zeroes kBlockSize (32) bytes instead of the encoded block size sizeof(MxFp4Stream) (17 bytes: 1 scale byte plus 16 nibble bytes). Writing 15 bytes past the end of the block corrupts the adjacent block when blocks are encoded out of order (such as across parallel batch boundaries in Compress or via packed_ofs sub-ranges) and writes past PackedEnd on the final block of a tensor.

Zero sizeof(MxFp4Stream) bytes in the all-zero early-out path so block quantization stays within the bounds of a single encoded MxFp4Stream block.

Based on
#1044