Non-ECC RAM errors are rare but real.

Very low. For a multi-day compile, the chance of a single-bit flip is less than 1%. Not zero, but not worth losing sleep over.

The rate of soft errors (cosmic rays, etc.) is roughly 1 bit flip per 1-2 GB of RAM per year in the real world. So if you have 32 GB and are compiling for 3 days straight, you’re looking at something like a 0.3% chance of hitting an error. Heavy load doesn’t change the cosmic ray rate, but thermal stress from sustained high temperature can increase the risk of hard errors (bad memory cells). That’s more of a QC issue.

If you’re building a production CI server or compiling a kernel that goes into medical devices, get ECC. For a hobby project or home build? Stop worrying and hit “make”. The compiler will catch most corrupted data anyway via checksums in intermediate files.

The only time it bit me was a bad stick, not a transient. That’s a different problem — memtest86+ before a long compile is smarter than buying ECC for a desktop.