fix: skip codex config.toml rewrite when there's nothing to merge

[chelsealong]

Fixes #4563

Problem

When Codex has no Specify-managed event hooks configured (the exact
scenario in #4563: a fresh specify integration install codex with no
extensions/overrides declaring any handlers), resolve_events() returns
{}, and install_integration_events() routes through its
empty-resolved-map branch into _remove_native_event_hooks() ->
_remove_toml_entries() — not through _merge_toml_fragment().

_remove_toml_entries() always rewrote the destination file, even when
stripping Specify-owned hook blocks was a no-op (no such blocks were
present) — i.e. a true no-op install/teardown against a config.toml with
no Specify content at all.

That unconditional rewrite went through Python's text-mode read_text()/
write_text(), which perform newline translation. On Windows this
silently turned an LF-terminated pre-existing .codex/config.toml into
CRLF, producing a git-visible diff with no semantic content change.
Because the file is outside the Codex/Spec Kit managed manifests in this
scenario, specify integration status --json reports a clean/healthy
state while git diff shows the file as modified — exactly as described
in #4563.

An earlier version of this PR patched _merge_toml_fragment() instead.
That function does have the identical unconditional-rewrite shape, but it
is only reached when there is at least one supported, non-empty event to
merge — the empty-map/no-op case never calls it, so that fix had no effect
on the actual bug. This revision fixes _remove_toml_entries(), the
function that is actually executed on the issue's repro path (confirmed
by tracing every caller of install_integration_events/resolve_events
and by instrumented runs of the CLI's specify init / specify integration install codex against a real .codex/config.toml).

Fix

_remove_toml_entries() now skips the write (return False, leaving the
file completely untouched — byte-for-byte, including its original line
endings) when stripping Specify-owned blocks left the content unchanged,
mirroring the existing "unreadable file" skip path already in this
function and the analogous guard already present in
_merge_toml_fragment().

The previous, harmless-but-insufficient _merge_toml_fragment() guard is
left in place (it doesn't hurt, and does fix a separate reachable no-op
case where a fragment is being merged but the file already contains it
identically).

Test

Replaced TestTomlNoOpMerge with
test_no_events_leaves_existing_config_untouched, which drives the real
production shape: install_integration_events(integration, tmp_path, manifest, {}) — the empty resolved-events map resolve_events() actually
returns when no hooks are configured — against a pre-existing
.codex/config.toml with no Specify-owned content. It asserts both that
the file's bytes are unchanged and that the file's mtime is unchanged
(byte-equality alone doesn't catch an unconditional rewrite on Linux,
where the platform line separator is already \n; the mtime check
verifies no write occurs at all, which is what actually mangles line
endings on Windows).

Confirmed the test fails without this revision's fix (git checkout HEAD -- src/specify_cli/events.py to restore the pre-this-fix version, which
still has the previous PR's _merge_toml_fragment-only patch):

FAILED tests/integrations/test_events.py::TestTomlNoOpMerge::test_no_events_leaves_existing_config_untouched
AssertionError: assert 1789266231860510616 == 1789266231859813750

With the fix applied:

tests/integrations/test_events.py::TestTomlNoOpMerge::test_no_events_leaves_existing_config_untouched PASSED
129 passed in 1.09s

Full suite (python3 -m pytest tests -q): 8054 passed, 12 skipped.

ruff check on the changed files shows the same pre-existing findings as
before this change; none are on the added/modified lines.

AI assistance disclosure

This PR was authored by an autonomous AI coding agent (Claude Code). An
independent review (also AI-assisted) found that the original version of
this PR patched the wrong function; this revision traces the issue's
actual repro path, fixes the function that is really executed
(_remove_toml_entries), and rewrites the regression test to exercise
that real path instead of a synthetic input no production caller
produces.