> Additionally, I stumbled across some probably older code (maybe from when `error` didn't throw an exception immediately). I moved the clean-up code to before throwing the exception here:
> https://hg.savannah.gnu.org/hgweb/octave/rev/2a12350f9410

This actually won't have any effect.  I've never been happy with this, but the octave_stream class has it's own error() function which is not the global error function that Octave uses elsewhere.  I made a comment about this at the top of oct-stream.cc a long time ago because I too was very confused about this.  The comment is:

// Programming Note: There are two very different error functions used
// in the stream code.  When invoked with "error (...)" the member
// function from stream or base_stream is called.  This
// function sets the error state on the stream AND returns control to
// the caller.  The caller must then return a value at the end of the
// function.  When invoked with "::error (...)" the exception-based
// error function from error.h is used.  This function will throw an
// exception and not return control to the caller.  BE CAREFUL and
// invoke the correct error function!

The change you made

+                        is.clear (is.rdstate () & (~std::ios::failbit));
                         error (who, "format failed to match");
-                        is.clear (is.rdstate () & (~std::ios::failbit));

is still probably a good idea because anyone who is familiar with Octave is likely to have the same confusion.  Putting the is.clear call first works whichever version of error() is called.

Sadly, I am re-opening this bug report on the development branch (Octave 9).  I had to make extensive changes to sscanf in order to make it more Matlab compatible and to resolve issues with bug #63383.  However, this now causes a failure for this test:

[line,count,errmsg,nextindex] = sscanf(' ,1 ',' %s ',1)

I verified the fix, and will close this bug report.

The only other thing is that it might be nice to have a BIST regression test for this using the compact test case

[val, count, errmsg, nextpos] = sscanf('p','%c')

Here is another example which has the same issue:

>> [val, count, errmsg, nextpos] = sscanf('p','%c')

val = p
count = 1
errmsg = sscanf: format failed to match
nextpos = 2

Following bug report #56396 (no error when there is one) but in the counter case, I noticed an error returned in version 7.1.0 on MacOS for the following minimum working example:

clear all
[line,count,errmsg,nextindex] = sscanf(' ,1 ',' %s ',1)
[line,count,errmsg,nextindex] = sscanf(' ,1 aa',' %s ',1)
[line,count,errmsg,nextindex] = sscanf(" ,1 "," %s ",1)
[line,count,errmsg,nextindex] = sscanf(" ,1 aa"," %s ",1)

the output is

line = ,1
count = 1
errmsg = sscanf: format failed to match
nextindex = 5
line = ,1
count = 1
errmsg =
nextindex = 5
line = ,1
count = 1
errmsg = sscanf: format failed to match
nextindex = 5
line = ,1
count = 1
errmsg =
nextindex = 5

Note the two error messages "sscanf: format failed to match" that seem strange.
For information the same example works correctly (void error message) on version 5.2.0 on a windows machine, and also with matlab R2022a.

Thanks in advance.