I'm mostly in the software space but in the past few years I've been doing a lot more embedded stuff, and the trend I notice is that companies are making great hardware, and then completely ruining its usefulness with bad software and firmware. It's kind of mind blowing to me because I always considered software to be the easy part of making a product, compared to, you know, etching microscopic patterns onto sand to make magical transistors appear in just the right way to do the task you want.
While I agree with your sentiment that there's a lot of poor software engineering in embedded space (especially in consumer-oriented novelty products, less so in established fields like industrial or telco), I can't but wonder what's wrong with Yocto? In my experience, it's quite the opposite: Yocto is the quickest path to get the firmware for a new device assembled, once you have climbed its pretty steep learning curve. I have built a few homebrew firmware build systems out of Debian and make/shell scripts (not my choice), you pretty quickly find yourself reinventing half of the stuff that Yocto does out of the box, but it's all bespoke, janky and hard to maintain. While with Yocto you just take the vendor's meta layer for BSP, put your application in another, and it bakes you a set of flashable images on the other end, complete with SDKs and other goodies for your dev workflow, reproducibly. It doesn't get significantly more sophisticated once you start to need kernel customizations, firmware updates with A/B partition layout, readonly rootfs, manage board- or customer-specific variants and other features that are very common in embedded systems but poorly or not at all supported in standard distros.
If you don’t already know anything about how to do these things in Linux Yocto is good for it. But if you try to swim against the current of the way the layer authors want you to do any of those things you will find it very challenging.
Take read-only root, for example. Usually it’s a one-line change in a config file plus an additional mount should you want an overlay in volatile memory. I don’t see how pulling in layers and config files does anything other than obscure how that works.
For a good example of what I hate about it, try building from Meta-Intel without graphics drivers.