I investigated this issue. I discovered a memory corruption which has the potential to corrupt WSPR tones in the second half of the transmission only. The first half is fine (first 84 symbols of the 162, to be precise). WSPR includes a lot of redundancy and forward error correction techniques which is why it can withstand QRM. This also means that decodes are possible even if some symbols have been corrupted. Which is why this issue escaped detection until now.?
I have corrected this issue and run some tests here. I already emailed Hidehiko-san a test .hex file. If anyone else would like to test this please could you email me privately, and I will send you the .hex file for testing.