Stephen Boyd [off-list ref] writes:

quoted

As can be seen in this change the only place where we actually want to
do something clever is with the to/cc/bcc variables, where setting
them on the command-line (or using --no-{to,cc,bcc}) should clear out
values we grab from the config.

All the rest are things where the command-line should simply override
the config values, and by reading the config first the config code
doesn't need all this "let's not set it was on the command-line"
special-casing, as [1] shows we'd otherwise need to care about the
difference between whether something was a default or present in
config to fix the bug in e67a228cd8.

This broke my workflow.

I specify --identity=<account> on the commandline and I want that to
pick out my send-email config from my global .gitconfig file that
corresponds to that identity. With this change, the config is parsed
before the getopt part so --identity on the commandline is a nop and
never looks into the config file to figure this out. So at least
--identity is special in addition to --to,cc,bcc.

Ah, sorry that nobody noticed that case, but you are right.  Because
the ident is used as a part of the key to find identity-specific
configuration values, if the command line gives one, we must have an
access to it before we start reading the configuration.  In that sense,
it is more fundamental to special-case the option.

We are past -rc0, so I am inclined to revert the change (and perhaps
replace it with the other "fix" that did not break the parsing order
like these patches did), with an expectation that a clever fix will
be found later, *unless* a simple and correct fix is found quickly.