You could switch it to unsigned 32-bit time_t, and most things would still work right up to 2106.
With any luck, civilization will collapse by 2038 and it won't matter anyway. But, struggling back up from collapse on scavenged 32-bit microcontrollers powered by scavenged solar panels, we won't need the extra burden of post-y2038. Switch to unsigned 32-bit time_t for the sake of the (surviving) grandkids.
You keep saying this, but you're too late. The 64-bit signed time_t path was already chosen long ago and is being implemented.
Not on those.
The kernel has already been fixed, but userspace (glibc) is AFAICS still in progress. Glibc is going to do it in a similar fashion to large file support (LFS), where if you define some particular macros you get the 64-bit time_t interfaces, otherwise you'll continue with the old 32-bit time_t. Similarly, existing binaries will continue to work, so in order to take advantage of 64-bit time_t you need to recompile (with those macros set, and fixing any resulting problems of course).
You can either do the ABI break and lose most existing users when the main set of apps stop working, or have the main set of apps stop working anyway in 2038.