meet_tvars: return bound tvar if both inputs are bound. Fixes #17648

[timholy]

Untested locally, let's see what breaks...

[vtjnash]

It passed, so does that mean it should be merged now?

[timholy]

Hmm, this doesn't go deep enough: I still can't turn on precompilation for ImagesAxes.jl.

[timholy]

OK, this version allows me to turn on precompilation for ImagesAxes. It passes tests locally, so assuming CI passes this version can be merged.

[vtjnash]

Should we run PkgEval on this before backporting?

[tkelman]

Already started what will hopefully become rc4 if it looks good. How critical is this?

[timholy]

The only known real-world bug this fixes is not being able to turn on precompilation in ImagesAxes.jl, so I'd rate it as non-critical. If this run works out, I'd recommend going with rc4 and waiting on this one until the next time you're backporting.

If rc4 needs more tweaks, then it would be nice to have.

[timholy]

Just thought of one more issue to consider, though: while I think it's unlikely, it's possible that there are packages relying on the current (broken) behavior. (This affects dispatch, so fixing the bug could conceivably change which functions get called.) If that proves to be true, there's always the possibility that fixing the bug in 0.5.1 would be breaking.

As long as you're aware of this risk and willing to accept the fact that we might need to encourage those packages to make fixes, I'd still say to go with rc4 as-is unless it requires additional PkgEval runs anyway.