You can put the lambda into the remove_if invocation:

auto tail = std::remove_if(list.begin(), list.end(),
                           [&dir](const QString& file) {
                               return file.startsWith(dir);
                           });

tail instead of it ?

You are missing the m_pendingFiles.remove(...) and m_recentlyEmitted.remove(...) calls here.

Variant A:
You can use std::partition instead of std::remove_if here (which keeps the elements between tail and .end() intact), and iterate over the tail to remove the elements from m_pendingFiles and m_recentlyEmitted. Last, do the erase(tail, m_cache.end()).

Variant B:
Use std::remove_if, but call m_pendingFiles.remove() (dito m_recentlyEmitted) inside the lambda.

Variant A is definitely the much cleaner approach.

I suppose developers on my level will appreciate code to be most self-explanatory. I'm not objecting though, we might al well go "all the way":

list.erase(std::remove_if(list.begin(), list.end(), [&dir](const QString& file) {
        return file.startsWith(dir);
    }), list.end());

Ooops.

Following Qt coding style, the [] () { should go on a separate line, as written above

I think having list.erase as a separate statement makes reading simpler.

indent by 4, please. And align }); with [&dir]

According to std::partition
what matches isDescendant should be from m_cache.begin() up to startRemoving, unless you want to remove what does not match isDescendant, isn't it?

++it is faster.

GCC generates identical code for both ...

Did you check? This is certainly correct for i++ if i is an int. But for iterators, which are classes, this is often not the case, but I may be wrong here. In any case, this is a minor nitpick you can simply ignore, so please go ahead :-)

Yes, objdump output is identical.
It is likely true for any decent compiler when the iterator implementation is fully contained in the headers. The compiler is aware the iterator copy is unused and can drop the respective code.