I guess my problem with that is, you're always going to have issues like this. Like...literally always, unless NPCs never use resources that aren't specifically things PCs could use already. Which puts us right back into the nightmare of 3e DMing, where every single fight is effectively creating multiple new PCs to throw at the actual PCs.
Something has to give. There has to be some allowance for abstraction and convenience here, simply to spare the DM's sanity. Just as HP are an abstraction over actual injuries, an abstraction that usually doesn't conform at all to how real injury works; just as 6-second rounds of combat are an abstraction that rarely conforms well to how an actual fight would work (just watch a fencing match!); just as initiative, and discrete chunky levels (though I know you choose to embrace that abstraction and say "no, this bizarre behavior really is how things work and folks know that"), and discrete CRs, and martial characters with resources rather than universally at-will abilities, etc.
"This NPC uses a magical action that isn't a spell accessible to you" or its non-magical equivalents, as long as such things are used relatively unobtrusively, still caring about diegetic stuff but not being totally constrained by "only and exclusively what PCs could use," is fundamentally necessary for both DM ease-of-use and DM creativity. If that path is absolutely verboten, you've just made DMing both significantly harder and far less diverse and extemporaneous.
I don't run D&D, so the comparison is far from perfect, but if I had to run things this way in my Dungeon World game, a good half of all fights the players have had not only would have been forbidden, but would have been genuinely impossible to implement. Solely to acquire just the tiniest bit of extra diegetic effect, practically a grace note, it would have sacrificed a huge swathe of the most enjoyable combats my players have faced.
Of course, DW actually does have generalized structures for how to construct moves, e.g. the "roll+MOD, on 10+ choose three, on 7-9 choose two" model or the "roll+MOD, on a 10+ gain 3 hold (or similar), 7-9 gain 2 hold, spend hold 1 for 1 to do any of the following" model, alongside a couple other basic formats, so drafting a new move that covers something relevant is literally as easy as picking the format you like best for this application and filling in options. But, again, imperfect analogy is imperfect; the point was simply that harping so hard on diegetic requirements to the exclusion of all else can have far more serious costs than folks seem willing to recognize.
Or, if you prefer? The reason we even speak of "diegetic" elements at all is because there is the question of whether the music in our audiovisual entertainment is diegetic or not. If film was not allowed to use non-diegetic music, it would be massively impoverished; consider the many beautiful scores of John Williams (Indiana Jones, Star Wars, Jaws, Superman, ET, Harry Potter, Close Encounters of the Third Kind, Schindler's List, Jurassic Park). An absolute requirement of perfect diegetic fidelity is just as dangerous as a total slap-dash "who the frick cares about being diegetic???" would be--perhaps moreso, because at least the latter can end up accidentally diegetic, while the former is not allowed to be even accidentally non-diegetic, even if doing so would be of benefit.
