point processes at 0 or 1 of a section do not deliver anymore current if extracellular is inserted.
That is true, and it seems very strange. And it has an unfortunate side-effect. Suppose you are studying a real cell that has a synapse attached to a particular point along the length of a particular dendrite. Let's say your model represents that dendrite with a single section, and that section needs a particular value of nseg for good spatial accuracy. Then the locations to which you can attach a synapse are 0, L * (0.5 + i)/nseg for i=0..nseg-1, and L microns from the section's 0 end. The likelihood that any one of these locations corresponds exactly to the placement of the synapse in the biological neuron is nil. You could get very close by making nseg very large, but that's going to increase the computational burden of executing simulations. And if you change nseg after the model's geometry has been set up, you're likely to move the synapse to an unexpected location (unless the change is multiplying or dividing nseg by an odd factor).
Or you could represent the biological neuron's dendrite with two sections that are exactly the right lengths, connected end-to-end. For example, suppose the dendrite is 200 um long and the synapse is 37 um from its proximal end. Just make section A that is 37 um long and section B that is 163 um long, and connect B's 0 end to A's 1 end. Attach your synapse to the 1 end of A, and you can do whatever you like to A.nseg and B.nseg without affecting the position of the synapse. And this will work just fine, as long as you don't insert extracellular into A.