You're adjusting Rmatch and Cmatch with the objective of getting a good S21 "shape", while ignoring S11, S22,
Yes, but I am also looking at insertion loss via the normalised trace in the simulation. However, in a typical setup, this wouldn't be possible so the aim is to get the best S21 shape across the passband with minimum passband ripple.
If this is done correctly, it should also coincide with the best overall s11 shape across the passband even though it can't (yet) be measured.
This does not tell you anything about the "input impedance" of the DUT. But it tells you the "optimal" terminating impedances of the DUT (call it Zopt, in your case around 361 - j487 ohms).
Yes, it lets me know Zopt. If you compare 361 - j487 and 1010R || 23pF at 9MHz they are essentially equivalent.
You would then devise impedance transformers to transform whatever impedance(s) are facing the DUT (50ohms or whatever) into Zopt. If those those impedance transformers are lossless you will get an overall insertion loss (shown in your video as S56) much better than what was seen with Rmatch/Cmatch (S21 in your video). But the return loss could also be very poor on both sides.
I'm not sure what you mean here. I quickly added an L match to the input and output as in the image below and you can see that the passband is still essentially identical with 2dB loss. The traces overlay almost perfectly. I've also added s55 which is the input of the L match. This is as good as it gets with this crystal filter. It isn't possible to get better than 11dB return loss across the whole passband and this is because this filter was designed for this return loss. In other words, no matter what I do with the L match values, I can't get a nicer return loss plot that what you see in the plot below.
On the datasheet, you would then say that the "input impedance" of the DUT is the conjugate of Zopt, assuming that this will lead designers to make matching networks presenting Zopt to the DUT.
Yes, that's it. Obviously, this technique isn't as elegant or as powerful as using a VNA, but it does usually give accurate results for Zopt