Each of two 50p-frames becomes one field of a 50i-frame. But one field has got half of the resolution of one 50p-frame, so you can add every two rows of the 50p-frame to one. And that's vertical box-low-pass-filtering with 6db-gain. ;)

Yes, of course. Would be a nice idea to stack lines (which would also give 6dB gain) but the time difference between fields has to be maintained.

The time-difference is maintained, because fieldA comes from frame1 and fieldB from frame2.
Assume you are creating a top-field-first interlaced frame out of two consecutive frames frame1 and frame2, then:
The lines 0, 2, 4, 6,... (field-A/even) of the interlaced frame are the lines 0+1, 2+3, 4+5, 6+7,... from frame1 and the lines 1, 3, 5, 7,... (field-B/odd) of the interlaced frame are lines 1+2, 3+4, 5+6, 7+8,... from frame2.
If frame1=frame2 (nothing moved), then the lines 0, 1, 2, 3,... of the interlaced frame are the lines 0+1, 1+2, 2+3, 3+4,... of frame1(=frame2). That's frame1 filtered with the kernel (1, 1)T (T=transposed), also called box- or mean-filter of dimension 1x2 with an amplification of 2 (=6db).

That was understood. I was commenting on my proposal to stack corresponding lines from successive frames, which cannot work because the time difference between frames must be the time difference between fields. But thanks for explanation, anyway.