fix: annotate -d crashes on a file that is not .py - #2265

coverage annotate -d DIR crashes with a bare AssertionError when any measured
file has an extension other than .py.

A directory with run.py and a measured mod.pyw:

$ python -m coverage annotate -d ann
before:
  File ".../coverage/annotate.py", line 79, in annotate_file
    assert dest_file.endswith("_py")
AssertionError

after:
$ ls ann
mod.pyw,cover  run.py,cover

The same command without -d works, so this is specific to the directory path.

Cause

flat_rootname turns the dot of the extension into an underscore, so
mod.pyw becomes ..._mod_pyw, not ..._mod_py. annotate_file assumes
otherwise:

dest_file = os.path.join(self.directory, flat_rootname(fr.relative_filename()))
assert dest_file.endswith("_py")
dest_file = dest_file[:-3] + ".py"

That assert came from 84baf6e1, "refactor: remove an untaken branch in
annotate.py", which replaced an if dest_file.endswith("_py") guard with an
assertion. The branch was taken after all, just not by anything in the test
suite: a .pyw file, or a file measured by a plugin, reaches it.

The fix

Put the file's own extension back, whatever it is, instead of assuming .py.

Verification, in plain terms

I added test_annotate_dir_pyw, which measures a .pyw file and runs
annotate -d.

Then I put the buggy code back exactly as it reads on main today and re-ran
that test. It failed:

>           assert dest_file.endswith("_py")
E           AssertionError
coverage/annotate.py:79: AssertionError

I also tried a subtler wrong fix, hardcoding ".py" instead of reading the
file's real extension. That does not crash, it just writes the wrong filename,
and the test caught that too:

>       assert sorted(os.listdir("out_anno_dir")) == ["main.py,cover", "mod.pyw,cover"]
E       AssertionError: assert ['main.py,cover', 'mod_.py,cover'] == ['main.py,cover', 'mod.pyw,cover']

So the test fails both for the original bug and for a near-miss fix, and passes
only with the real one.

tests/test_annotate.py is 6 passed. ruff check and ruff format --check are
clean, no line over 100 characters.

On the rest of the suite: 194 failures before the change and the same 194 after,
with an empty diff between the two sorted lists. They are pre-existing in this
environment, which has no C extension built and no network.

The test runs on Linux: it exercises the reporting path for a .pyw file, which
is where the crash is, without needing Windows to execute one.

Disclosure: written with AI assistance (Claude Code). I reproduced the crash and the fix by running the real coverage annotate -d command against a measured .pyw file, and ran both checks above and the before/after failure-set comparison myself.