It probably took two years, because either the data sent out changed or you handling of the data (master or slave side) changed because of the actual values. Or maybe the processors didn't run with crystals (except for the RTC) and the baudrate drifted slightly.

The issue here is that the extra ms needed means your software wasn't rugged enough. For two years, it was running on the margin, but you had failed to identify this critical flaw. It's no different from having a design running at just below the temperature where it fails, or at just the limit of how low voltage the processor can work with.

All software development for embendded projects needs to set up a long list of critical limits, and then try to validate that there are ample safety margins. It's important to know that the slaves have whatever safety margin they need when picking up every single byte from the UART. Did you test with a higher baudrate or shorter delays than what you used for the final release?