Manuals are for slow cars and drivers that can't left foot brake.  Which is why NASCAR still uses them.  The real racing world like F1, Porsche and Ferrari have moved on toward better performance than 'manuals' are capable of delivering.  Modern physics has exceeded their potential.  I wouldn't say Sebastian Vettel or Fernando Alonso have no business in a race car if they didn't know how to drive a manual.  Point being, sometimes there are more important things for people to focus on that are more relevant for what they are trying to accomplish.  That doesn't excuse ignorance of very significant and personally relevant knowledge however.     

Uh, windowing sinc with sinc gives you the shorter sinc. That's why nothing happened.

Ahm, both are windowed sinc filters. The result is by definition longer (the order of both filters summed plus one).
It's not that nothing happened, but that the filter cannot remove what was already removed so no extra visible ringing is added. There is however ringing added way way down - after the end of data1 I looked at the samples of data3, there's ringing about -200 dB down. After quantization all of this is gone.

Will look into other filters you suggested.

The thing is, as soon as you have a properly bandlimited signal, let's say a low pass starting at 20 kHz reaching stopband close to 22.05 kHz, you can add an indefinitely steep linear phase reconstruction filter just below 22.05 kHz and the signal will be reconstructed perfectly down to quantization limits and up to just below 22.05 kHz.

Any linear phase filter with slow roll-off, such as an apodizing filter, will just "add" aliasing that rolls off as slow as the filter dictates.
Any steep min phase filter will cause crazy phase shift, cheby type 2 obviously more than butterworth.
Any linear phase filter with less attenuation will also "add" aliasing, but at a more or less constant level comparable to raising the noise floor.

Simple interpolation will not only cause roll off in the potentially audible range but also do a bad job at removing aliasing. In order to get good SNR with something as a simple cubic hermite spline you need quite a high oversampling ratio to begin with.