In the frequency domain a signal with a carrier offset looks like the following-
This is usually modeled as the desired baseband signal convolved (in the frequency domain) with a complex tone with frequency equal to the carrier offset.
As you are probably already aware, convolution in the frequency domain is equivalent to multiplication in the time domain. Thus, the baseband signal is multiplied, sample by sample, by the complex tone in the time domain. As the complex tone is rotating around the complex plane, at a speed that is proportional to its frequency, this causes the baseband signal to rotate as well. We can see that when we recognize that multiplication of polar values causes the phases to add, so the tone's phase is added to the baseband signal's, which causes the rotation.
So, given that, does the output of the symbol matched filter also rotate? Yes.
Let's look at a simple example: BPSK with the symbol being a simple +1 and an initial phase offset of 0.
Although the signal starts at 0 phase, each sample rotates it farther and farther around the complex plane. Your main question seems to be about what happens when the symbol matched filter is applied, and you bring up the raised root cosine (RRC) filter in particular. The RRC filter is just a weighted boxcar filter. It sums up sequential samples after a weight has been applied to each. For ease of drawing, let's assume an RRC filter that only has three taps. The middle tap will be the biggest, with the edges being smaller.
Hopefully it is obvious that the angle of the sum of these weighted samples/vectors will not be 0. Now picture the next three samples. They will be rotated even more than the first three, and the weighted sum of them will also be rotated more. This does, of course, continue.
At this point it is hopefully clear that the matched filter does not prevent rotation, it just averages it.