Here's a counterexample from HashBackup (I'm the author):
A filesystem is being backed up. Some weird error occurs, let's say it tries to read an extended attribute and gets some goofy error that raises an exception and it's not in a try/except. That error bubbles up to the try/except for "save a file", prints the exception message, and continues by backing up the next file.
Is it more reliable to crash when reading an extended attribute, aborting the whole backup? Or is it better to just tell the user about the error, finish the backup, fail at the end of the backup with a non-zero exit code, and report it in an email?
At the time of failure, the only way I can see that the error could have been handled at the call site would be to return an empty extended attrbute. To me, that's more like the "silent but wrong" solution. And there are many of these situations for saving a single file: get the flags, read the extended attribute, read the permissions and times, figure out if it's sparse, etc. Instead of having error handling at each of these steps, they can have no error handling and just raise exceptions and the high-level "save a file" handler will do everything.
Of course, along the way there are some low-level try/except blocks around code that needs cleanups, like:
fd = None
try:
fd = os.open(...)
do some stuff
finally:
if fd != None:
os.close(fd)
I think because it sets up the right psychological incentives. Because almost any option other than those two ends up looking like, "Hmm, I don't know how to handle this, but maybe someone else will, so let's bubble it up and hope they do," in practice. It's just the way we humans are.
And that sets up an intractable moral hazard situation. It creates this middle ground where two different owners of two different parts of the code can finger-point and blame and try to pass the buck. And, knowing that (though probably only unconsciously), they won't just become sloppier about handling errors. They will become sloppier about allowing errors to happen in the first place, and doing things that make it harder to handle errors. And they won't even know they're doing it.
That's what I think of when I read your hypothetical, "Let’s say you want to read from a file during a HTTP request but the read fails n methods deep." That sounds to me like code that's been structured in a way that, on this team I was alluding to above, probably wouldn't have happened in the first place and definitely wouldn't have made it past code review. The pushback would be something along the lines of, "Why on earth are you doing both file and network I/O within a single business operation but at two very different depths in the call stack? Did you even notice what you were doing here?"