Hosts are expected to be able to handle 576 byte IP packets = be able to reconstruct packets of that size from fragments if they are behind a < 576 MTU. For a long time, the TCP MSS would default to 536 (576 - IP - TCP) or 512. However, there is no real requirement about the minimum link MTU apart from the 68 bytes in RFC 791. People often attribute a 296 byte MTU recommendation for slow links with header compression to RFC 1144. This isn't entirely unreasonable, but it's not explicitly in the text.
In any case, doing PMTUD for UDP in IPv4 doesn't make much sense. And having the kernel set the DF bit and then expecting the application is completely wrong, as many UDP applications can't reduce their packet size, or can't do this without serious loss of functionality, and fragmentation at the source host also can't be controlled by applications.