So here's my question: how big does a speck have to be before it is noticeable on a thermal image?
Assuming the window is about a mm or so above the FPA and the rays from the lens are converging, any speck is going to appear on the FPA slightly smaller than it really is. The actual geometry depends on the lens characteristics and front of window to FPA distance but let's say for the sake of argument that a 10um speck casts a 9um shadow on the FPA.
Now we have to consider the microbolometer pixel pitch. Let's start with 17um. Assuming a 100% fill factor (which is essentially unachievable but makes the assumptions easier), one would need a carefully positioned speck to shade just one pixel. Based on the earlier assumptions, a speck of about 18 or 19um would be big enough to obscure a 17um pitch pixel. But in the 'best' position, equally shading four adjacent pixels, the effect on each pixel would be to shade just one quarter of each. Would that be enough to show up? Or are we now in the position that we need a (square) speck 35um on a side before there are discernible effects?
If the rays converge quickly from the lens and/or the front of window to FPA is larger than my guesses, a much bigger speck is needed to achieve the same masking effect. On an older 50um sensor you'd need an easily visible speck around 0.1mm square to have a major effect.
The sharpness of the lens will have a big impact too. LWIR lenses are generally made from a number of compromises in a row and are nowhere near as sharp as a DSLR lens (they don't have to be) but this might increase the speck tolerance by say 50%. So now we're looking at perhaps a 0.15 x 0.15mm speck - that's well into blob territory!
I am quite happy for anyone to correct my assumptions and arithmetic but this little thought experiment does begin to indicate why really small, shallow scratches are nothing to worry about.