The ST appnote is one of the more convincing / well written ones:
https://www.st.com/resource/en/application_note/an2867-oscillator-design-guide-for-stm8afals-stm32-mcus-and-mpus-stmicroelectronics.pdfmaybe ironically -- but perhaps because of -- their somewhat notoriously? weak oscillators. That is, they tend to use lower drive current, or less transconductance, than most, necessitating a higher ESR (lower CL) crystal than average.
There indeed exists a methodical design and testing procedure for crystal oscillators. The main problem is, whether required supporting information is available (e.g. gm of oscillator, at frequency, and at drive strength/mode(s); Cstray, Rfb, etc.), and whether min/max/median parts are available to actually test, plus env chamber to run at temp extremes, and supply voltage extremes; and, most of all, whether the budget and time is there to actually perform such testing, versus whether it's worthwhile at all.
Conversely, you can buy and measure random crystals: Fo, Cs, CL, ESR, ppm/°C, pulling, etc. But this is a bit of work, and moderately specialized hardware is required (I suppose 32kHz crystals you could do with a sound card, but mostly you'd want a VNA, of adequate spectral purity to resolve such narrow resonance), and if you're actually going to sort random parts, there's maybe no guarantee they'll be supplied consistently either, and now you've developed a receiving parts test program at whatever expense it is.
And as you can see, it can take specialized hardware. An AC current probe to resolve µA into a crystal, is not exactly trivial. Most people don't have one of those handy, and it's going to be special order. A regular 10x voltage probe is insufficient as well: Cload is significant even for MHz crystals, and both Rload and Cload is too much for most watch crystals (100s kohms). A FET probe is probably best; easy to make, at least, but also easy to fry, and if one doesn't want to make it, it's still another instrument to buy.
The principle of least surprise tends to apply, then, with no one bothering to test, and assuming the most common parts and most commonly repeated circuits are valid. And they do usually work, but no one knows what the production yield is, whether they conk out at low or high temperature, whether complaints / service calls / warranties can be traced to it or not, ..., and yes, it can be "blind leading the blind". Or perhaps, "blissful led by the ignorant", to merge with another phrase.
Regarding USB and Ethernet, at least, they're quite tolerant as far as I know -- but also, you'd have to know what the interfaces actually do, to know what clock tolerance is really acceptable; to know how they degrade if at all along the margins of that range; those sorts of things.
---
As for pulling -- note the C-crystal-C 'pi' structure acts to invert the parallel resonance to a series one, roughly speaking; C is inversely proportional to Zo of the filter. So you have an impedance there, and that impedance must be near the oscillator's required load impedance (or 1/gm or whatever, plus margin to ensure oscillation of course). You can vary Cload to vary Fo, but you vary Zo as well. You can also put C in series with the crystal, which reduces the crystal's effect (in the limiting case, there's lim-->0 capacitance in series, and all C in shunt -- no crystal in circuit at all!), i.e. the depth of the resonance, but keeps the impedance close to nominal while allowing you to reduce Cload. In total, you have strategies to pull it a little bit, up or down -- as for depth of pulling, it's usually some 100s Hz, maybe a couple kHz at most -- and if you have poor initial tolerance but acceptable stability, this might be a strategy to make use of it, assuming of course the resulting range is outside what's required for the hardware driven by it. But then you're also adding a calibration (trimming) step, maybe with a special probe to read the crystal (if a direct logic-level clock isn't available; usually Ethernet does have this, USB doesn't?), and the added test time starts to look like qualification testing, and...
As for that last part, hardware tolerance? If it's not in the datasheet -- or it is, but you want to know more -- you'll have to ask the manufacturer, of course.
![Smiley :)](https://www.eevblog.com/forum/Smileys/default/smiley.gif)
Tim