Do not report a fix that --ignore-regex made impossible

[Eljees]

Fixes #2056.

What happens

$ printf '1nd\n1nd_2nd\n' > test.txt
$ codespell test.txt --ignore-regex _ --write-changes
FIXED: test.txt
  test.txt:1: 1nd ==> 1st
  test.txt:2: 1nd ==> 1st
$ echo $?
0
$ cat test.txt
1st
1nd_2nd

Line 2 is reported as fixed, is not fixed, and does not affect the exit code — so a CI run goes green with the misspelling still in the file, and running codespell again prints the same false FIXED: forever.

Why

Detection and writing look at two different texts. Detection walks the line with --ignore-regex substituted out (_ignore_word_sub replaces the match with a space), so it sees 1nd 2nd and finds 1nd. --write-changes then substitutes on the original line:

if options.write_changes and fix:
    changed = True
    lines[i] = re.sub(rf"\b{word}\b", fixword, lines[i])
    fixed_words.add(word)
    changes_made.append((line_number + 1, word, fixword))
    continue

In 1nd_2nd there is no word boundary between 1nd and _, so re.sub matches nothing. But changed is set before the result is inspected, changes_made records a change that never happened, and continue skips the warning branch that would have set the exit code.

The change

Keep the result of the substitution, and only claim the fix when the line really changed; otherwise fall through to the normal warning. --ignore-regex semantics are untouched — the word is still detected, it is just reported instead of silently dropped. This is the second of the two outcomes the reporter said would be acceptable ("either fix it, or tell me it can't").

After:

FIXED: test.txt
  test.txt:1: 1nd ==> 1st
test.txt:2: 1nd ==> 1st
$ echo $?
65

Tests

test_ignore_regex_with_write_changes in codespell_lib/tests/test_basic.py, next to the existing --ignore-regex tests. It asserts all three symptoms: line 1 is rewritten, line 2 is not listed as fixed, and the exit code is non-zero. On main it fails on the middle assertion.

pytest codespell_lib/tests/test_basic.py: 85 passed on main, 86 passed with this change, no failures either way. ruff check, ruff format --check and mypy are clean (mypy reports the same two pre-existing chardet stub errors on an untouched tree).