The sensor is a single "chip".
The sensor is designed as a portrait sensor, the number of "active pixels" is 4992 x 6668.
There are some additional pixels which are of no concern for this matter.
When you hold the sensor into the light and tilt it, you will see that the light is reflected unevenly. You will be able to make out 6 areas. 2 rows of 3 columns (portrait mode).
This has to do with the manufacturing process. Whatever the reason for having those areas is doesn't matter, important is that they exist. On page 9 of the data sheet the stitching effect is defined to be typical 1% and maximum 3%.
The sensor has 4 (not 2) readouts which are located in the corners (not sides).
When an image is taken, all pixels are captured simultaneously. So the data in the sensor (so to speak) is identical regardless of the number of readouts used.
Once captured the data can be read out using 1, 2 or 4 readouts.
If you use 1 readout, the first sensor line is copied into an output register and the pixels are shifted out one after the other, as soon as the register is empty (all pixels of the line are shifted out) the next line is loaded into the register and shifted out and so on.
If you use 2 readouts you can use the same register, but the register will shift out 2 pixels per clock, one on each "end". The left half of pixels from one readout, the other half from the second pin. Note that the right hand pixels are shifted in reverse order. Outside pixels first so to say.
Alternatively you can use 2 registers. The top half of lines (one after the other of course) will be loaded into the top register and the second half into the bottom register. Both registers will be shifted out at the same time.
The top register gets the top half of rows the bottom register the bottom half. Again lines are shifted out in the outside first order. So the top register will shift out lines 0, 1, ... the bottom will shift out lines 6667, 6666 ... (talking visible lines only).
With 4 readouts both variants are combined.
Please remember that the sensor data is identical regardless of the readout mode used.
So if there is a stitching error of 3% it will be there no matter how many readouts you use, but ...
If you use more than one readout, you will add an additional error for the separate signal paths.
I haven't found a specification how much the amplifiers of the readouts differ and of course I can't say how much the external amplifiers differ.
This might add up to the problem, but it's difficult to guess how much. An educated guess would be easier, if I'd know which readout variant is used (1 readout or 2 readouts / 1 register or 2 readouts/ 2 registers).
But what can be said is, that if the problem is related or at least increased by the readout mode, it would have to be the 2 readout / 2 registers mode (or 4 readout mode which I doubt is used).
It has two be 2 readouts because 1 readout doesn't add different errors to top and bottom halves and it would have to be 2 registers because the error shows usually as a difference in top and bottom half of the image (remember it's a portrait sensor) and to get different paths for top and bottom half you have to use separate readouts for the lines.
So the sensor is not stitched in the classical two parts sense but it is stitched in the sense of areas that differ.
Two sensor stitching
Any sensor has a frame of some sort, because the sensor can not be manufactured in a way that the sensor-pixels are placed directly up to any edge of the chip.
So you will always have a gap. The gap is very large compared to the size of a single pixel.
Assuming that nobody wants to live with a multi-pixel gap in the center of his images, he optical path will have to be split, so that one part of the light will hit one sensor the other part the other sensor.
Which sounds rather difficult especially if the solution has to work for wide angle and tele lenses alike. It doesn't make the design simpler that the "splitter" should not reduce the amount of light hitting the sensor.
Not to speak of the problems to make the sensors match in color and brightness rendition.