You are applying synthetic signals with varying bandwidth. Although the gaussian noise has constant energy, that energy is spread over a wider bandwidth as you increase the number of samples. A purely mathematical/simulation effect that is not matched with real systems as are being discussed here. So it is not the FFT length that is making that difference, but the different noise you added to the signal.
Sure, what happens at the end is: When the analog wideband noise is sampled, then all noise frequencies beyond Nyquist are folded into the 0...Nyquist range. And Nyquist is of course higher when the sample rate is higher, therefore a higher sample rate distributes the sampled noise over a larger bandwidth, resulting in a lower power density per Hz. Consequently a bandpass filter with a given ENBW extracts less noise power from the high-sample-rate signal than from the low-sample-rate signal. Still the higher sample rate helps to push down the random noise floor w/o increasing the measurement time.
I don't agree that random noise is of no practical importance. Of course it won't be perfectly white, but It is certainly a nasty issue in practice, too, which needs to be addressed if we don't want that a measured high dynamic range transfer function of a DUT gets buried in the noise floor. OTOH, a small amount of dithering noise is even useful if we want to sqeeze out more than 50dB dynamic range from an 8-bit ADC (improves SFDR, INL,...).
Now repeat the same with some aggressor signal that is contained in a small bandwidth (extreme case an adjacent sine), adding more sample points does not improve the attenuation of signals the same distance apart, that attenuation is set by the FFT + window function.
If you could make a steeper filter by interpolating more points in-between you could make an infinitely steep FIR in a finite length (obviously impossible).
When you initially wrote "with zero noise suppression", I did intpret it it as random (wideband) noise, and not as arbitrary "agressor signal".
I fully agree that a narrower filter bandwidth requires a longer measurement interval. And if a filter with a better selectivity than the (rather lousy) sinc response of a rectangular window is desired, then a longer measurement interval is required as well.
[ However, if only harmonics of the stimulus need to be eliminated, and if a frequency plan can be arranged, such that the window size is an exact integral multiple of the stimulus period, then a rectangular window is still fine (transfer function has zeros at all harmonics), and it has the lowest ENBW among all window functions for a given window size (measurement interval). ]