Your missing the noise generated by the active devices. For communication equipment it is rare to set impedance by dead weight resistors. Sometimes a resistor may there for biasing, maybe even stability, but is it usually several times higher then the device input impedance to keep the resistor's noise temp from degrading overall noise figure.
Noise voltage generated across a resistor is square root of (4kTR) in V/root Hz.
The total noise power must take into account the bandwidth of interest. noise power is kTB.
A good way to work with series of devices is to model each stage as an input summation of all the input noise power sources followed by an idea noiseless amp.
For example, if the noise figure of an amplifier is 2 db, it will have an input noise temp equivalent noise contribution of ([(NF/10)^10]-1)* 290. For 2 db noise figure, the noise factor of the amp will have the equivalent input noise temp of (2-1)*290 degs K or 170 degs K.
If this amp is fed from a satellite dish looking into the sky then the antenna noise temp is a function of the blackbody radiation of what it is looking at. Sky pointed antennas may have a low noise temp because the sky can be quite low depending on what it is pointing at. Lets say it has 50 deg K equivalent noise temp. Then the input to the amp model would be 50 deg K + 170 degs K or 220 degs K. The total noise power within the bandwidth of interest is KTB or 1.38x10^-23 * 220 degs K * BW3db. This input noise power would be gained up by the power gain of the amplifier. This becomes one noise power input to the next stage which is done as a summation just like the first stage.
You can see the NF of the first amp should be lower since its contribution is greater then the input noise temp from the antenna. You can also see that as the noise power are gained up by successive stages the noise contribution from downstream devices is masked by higher and higher noise power from earlier stages that are gained up through earlier devices.