It can still crash

That "works" to some extent, but it is not guaranteed to work in all cases. Rust promises that a const block will be evaluated if it is reachable. It does not promise that an unreachable const block is never evaluated (in the general case, that would require solving the halting problem). If a Linear<T> instance actually exists over a large section of code, and especially if it is passed as an argument to some other crate, rustc may or may not figure out that it can elide the Drop impl, and so the assertion can fire even if a Linear<T> is never dropped.

There are different levels of this:

A weakly-undroppable type may not be dropped. It may be leaked, forgotten, put into an Arc/Rc cycle, or smuggled inside of ManuallyDrop. It may also be dropped on an unwinding panic (with the understanding that panics are bad, unwinding is worse, and some features just won't play nicely with them). It does not provide a safety guarantee, so unsafe code must assume that undroppable types may still get lost by other means.

A strongly-undroppable type may not be dropped, and additionally provides a safety guarantee that its drop glue is never called. It cannot be dropped on unwinding panic, so such panics are converted into aborts if they would drop the object (you can still use catch_unwind to manually handle the situation, if you really want to). Unsafe code may assume that the object is never dropped, but still may not make any other assumptions about the object's ultimate fate.

An unleakable type may not be leaked or forgotten. It must always be dropped or destructured. You may not put it into Rc, Arc, std::mem::forget, ManuallyDrop, MaybeUninit (which is just ManuallyDrop in a funny hat), Box::leak, or anything else that could cause it to become leaked (unsafe code may do some or all of these things, but must not actually leak the object). You also may not define a static of that type, because statics don't get dropped at exit and are functionally equivalent to leaking an instance on startup.

A strongly-linear type is both strongly-undroppable and unleakable. It cannot be dropped or leaked by any means, and is subject to all of the constraints listed above. It may only be destructured by code that has visibility into all of its fields.

Now my commentary:

* Weakly-undroppable types are already very useful as a form of linting. For example, the std::fs::File type currently does not have a close() method, because the type already closes itself on drop. But that means that it swallows errors when it is closed. The documentation recommends calling sync_all() (which is equivalent to fsync(2)) if you care about errors, but I imagine that some filesystem developers would have choice words about doing that in lieu of checking the error code from close(2). If File were weakly-undroppable, then it could provide a close() method that returns errors (e.g. as Result<()>) and fails the compilation if you forget to call it. This isn't a safety issue since the program won't perform UB if you forget to close a file, so we don't need a strong guarantee that it is impossible to do so. We just want a really strong lint to stop the user from making a silly mistake. It would also help with certain problems involving async destructors, but I don't pretend to understand async nearly well enough to explain that. On the downside, it would interact poorly with most generic code, and you'd probably end up copying the semantics of ?Sized to avoid massive backcompat headaches (i.e. every generic type would be droppable by default, and would need to be qualified as ?Drop to allow undroppable types).
* I'm not sure that strongly-undroppable types provide much of a useful improvement over weakly-undroppable types, and it would be absurd to provide both features at once. But it could be argued that having one exceptional case where drop glue is invoked is a recipe for bugs, so it might be a cleaner implementation. OTOH, if you have chosen to have unwinding panics, you probably don't want them to magically transform into aborts just because some library in your project decided to use an undroppable somewhere. I currently think that weakly-undroppable is the more pragmatic choice, but I think there are valid arguments to the contrary.
* Unleakable types would allow the original API for std::thread::scope to be sound, and would probably enable some other specialized typestates. They're also pretty invasive, although probably not quite as much as undroppable types. They would not solve the File::close() problem.
* Strongly-linear types are just the combination of two of the above features. If either feature is too invasive to be practical, then so are strongly-linear types. But they would provide a strong guarantee that every instance is cleaned up in one and only one way.

Of course, there is another problem: We cannot guarantee that an arbitrary Turing-complete program makes forward progress. If the program drops into an infinite loop, deadlock, etc., then no existing object will ever get cleaned up, meaning that everything is de facto leaked whether our types allow for it or not. To some extent, this is fine, because a program stuck in an infinite loop will never execute unsafe code that makes assumptions about how objects are cleaned up. To some extent, it is not fine, because we can write this function (assuming that ?Leak means "a type that can be unleakable"):

fn really_forget<T: Send + 'static + ?Leak>(t: T){
std::thread::spawn(move || _t = t; loop{std::thread::park();});
}

...or some variation thereof, and there is probably no general way to forbid such functions from existing. So any type that is Send + 'static (i.e. has no lifetime parameters and can be moved between threads) should be implicitly Leak.

The "obvious" approach is to make 'static imply Leak, and require all unleakable (and maybe also undroppable) types to have an associated lifetime parameter, which describes the lifetime in which they are required to be cleaned up. More pragmatically, you might instead say that 'static + !Leak is allowed as a matter of type coherence, but provides no useful guarantees beyond 'static alone, and unsafe code must have a lifetime bound if it wants to depend on something not leaking. I'm not entirely sure how feasible that is in practice, but it is probably more theoretically sound than just having !Leak imply no leaks by itself, and unsafe code probably does want to have a lifetime bound anyway (it provides a more concrete and specific guarantee than "no leaks," since it allows you to assert that object A is cleaned up no later than object B).