This is something that I have always found quite counter-intuitive too. The way I think of it is that the end point adjustments are a bit like moving the reference point so that the height of each rectangle is equal to the value of the function at its centre rather than the value of the function at its left-hand edge.

Also it is worth bearing in mind that with 10,000 points the rectangle approximation will already be quite good so it will be quite close to the trapezium approximation anyway. You wouldn't expect a huge difference.