As for the MIME type, for image types I'd say it's more than stable enough. Certainly much, much more stable than the 6.7% error rate mentioned in the article here, I'd be surprised if it was even 1%. If you double click on an image on your desktop for example, you can in almost all cases expect that it will be opened correctly. It ceases being a heuristic entirely if you tell the webserver that *.png is image/png, and only put PNGs with names ending in ".png".
Guess those are the reasons why I got out of web development in 10 years ago, everything's held together by scaffolding and needlessly wasteful and inefficient there.
They’re not used for rendering, they’re used for figuring out what to fetch. A HEAD request would be far less efficient than knowing ahead of time what to fetch: 1 request versus 2N+1 requests.
What you suggest sounds all fine but the entire web is user input for a browser, so no matter what, you need to define how to fail. If you can fail gracefully, you might as well do so, because a failure might not even be triggered by bad code/configuration on your side but simply by flaky network issues.
I do wish that content negotiation (e.g. Accept headers) worked properly. In the end though, those hints implement a subset of content negotiation in a reasonable way, given the state of affairs.
