That makes it sound like you're thinking that the chances for a conflict (same hash value) depend just on the size of the hash value. But it's worse than that, thanks to algorithmic insights, inputs that produce the same hash value can be found more cheaply than just by chance. Also, they can be found so cheaply that it's computationally trivial to do even for an individual [1]. Thus anyone who can change what data you receive, can trivially also make it hash to the same MD5. (I haven't checked whether publicly available software exists to generate conflicting files that are also, for example, valid gzip files, but I surely expect that an entity like the NSA (and perhaps blackhats, too) can create such software on their own.)
[1] "produce a collision for two inputs with specified prefixes within hours, using off-the-shelf computing hardware": https://en.wikipedia.org/wiki/Md5#Security
Yes, there are many sites that still tell users to check MD5 hashes. I think it's evident that those are totally worthless for some years now. Worse, it may be giving some people a false sense of security. OTOH those places might give those who do know a hint that perhaps the authors or distributors don't know about security. (Sorry for the harsh words...)
That is indeed a limitation. I'd suggest you explain what md5sum is, how to use it, and what its limitations are -- specifically that collisions are now trivial to instrument.
Developers / vendors still relying on md5sum really should be named and shamed these days.
Although there are some systems where I'm familiar with the code - such as Bitmessage - in most cases I havn't personally checked the code myself. Ideally Freedombone would be a pure blend (I think that's also the aim of FreedomBox), but for now it does involve downloading some non-packaged systems.