Pre-baking would be more of a trick pipeline whereas using physically based equations and simulating how light actually interacts with the environment would be the later.
That's fair - but I don't think there's a really hard line that we can draw. Eg. I can do just in time evaluation and generate probes in real-time, on demand(and thus running path-tracing at runtime, just not quite 'real'time), giving me the results of baked solution without the drawbacks. Not that this would be a platform-specific solution, just since we were on topic of simulation vs tricks.
More specifically, find several examples of features being run on the console that just wouldn't be possible on PC
Ok that helps. First let me make a point here that context is rather important for these, so what works in one title may not work that well everywhere else. And also, the only thing we can really talk about is 'prohibitively expensive' at any given time, nothing is ever truly impossible.
That said - one of the most obvious examples of past few years is VR and various lens-shape optimizations for rendering. These work exceptionally well on fixed platforms (PSVR and Quest most notably) and are incredibly cumbersome to use/implement on PC - no standard features in hardware(even within single vendor's solutions), no API-access to certain acceleration structures (like H-Tile buffers), largest percentage of the market actually missing useful hw or API support alltogether, and unpredictable results where gains on one GPU can actually be non-existant or even detrimental on another.
On top of that - solving for CPU overhead is even harder - while on a box that gives you access to push-buffers you can literally draw multiple-views N-times faster(CPU overhead) even on GPUs that offer no hw-tricks for variable resolution render targets.
The use-case is highly specific, because most of these types of optimizations are. And obviously mileage will vary and results can be much less dramatic than this case.
Another thing worth calling out (less for performance and more flexibility) is eg. mesh-shader style pipeline setup, which has been in use on current-gen consoles for quite some time now, thanks to async-compute and some lower-level API access - ie. it isn't even a new pipeline-feature to consoles, even though it is to PC platform.
Older analogue would be things like PS2 hw-accelerated single-pass shadow-volumes with NO-cpu intervention needed (where PC had to do multiple-pass, CPU extrusion, CPU vertex generation), but that involved specific hardware features as well as low-level access.
Definitely clumsy to use and I'd much rather have a closed system like the Quest.