Allow native review targets to accept supplemental custom instructions

I can reproduce this as a CLI parser/help consistency bug on current CLI 0.133.0 (Darwin 25.4.0 arm64 arm).

codex review --help advertises a positional prompt:

Usage: codex review [OPTIONS] [PROMPT]

Arguments:
  [PROMPT]
          Custom review instructions. If `-` is used, read from stdin

But passing a prompt with a native target is rejected before review starts:

$ codex review --uncommitted "focus on panic paths"
error: the argument '--uncommitted' cannot be used with '[PROMPT]'

Usage: codex review --uncommitted [PROMPT]

For more information, try '--help'.
exit=2

I also added a focused local parser test to capture the expected behavior implied by the help text:

#[test]
fn review_uncommitted_accepts_supplemental_prompt() {
    let cli = MultitoolCli::try_parse_from([
        "codex",
        "review",
        "--uncommitted",
        "focus on panic paths",
    ])
    .expect("review help advertises [PROMPT] with --uncommitted");

    let Some(Subcommand::Review(ReviewCommand { args, .. })) = cli.subcommand else {
        panic!("expected review subcommand");
    };

    assert!(args.uncommitted);
    assert_eq!(args.prompt.as_deref(), Some("focus on panic paths"));
}

Running it fails with the same root parser issue:

$ just test -p codex-cli review_uncommitted_accepts_supplemental_prompt
TRY 1 FAIL codex-cli::bin/codex tests::review_uncommitted_accepts_supplemental_prompt
...
review help advertises [PROMPT] with --uncommitted: ErrorInner { kind: ArgumentConflict, context: ... InvalidArg "--uncommitted", PriorArg "[PROMPT]", Usage "codex review --uncommitted [PROMPT]" ... }
Summary: 0 passed, 1 failed

Likely source: codex-rs/exec/src/cli.rs defines ReviewArgs with conflicts_with_all = ["base", "commit", "prompt"] on --uncommitted and similar prompt conflicts on --base / --commit. That makes the prompt impossible to use with native targets even though generated help/error usage still exposes [PROMPT].

If supplemental prompts are intended, the parser conflict should be removed and the prompt should be threaded as additional review context while preserving the native target. If they are not intended, the help/error usage should stop advertising [PROMPT] alongside native review targets.

I can still reproduce this on current CLI codex-cli 0.142.5 on macOS/Darwin arm64.

The current help/usage text still appears self-contradictory: it advertises [PROMPT] together with native review targets, but the parser rejects that combination before any model call is made.

Minimal repro in a temporary git repo:

codex review --base main 'custom review instructions'
codex exec review --base main 'custom review instructions'
codex review --uncommitted 'custom review instructions'
codex exec review --uncommitted 'custom review instructions'
codex review --commit HEAD 'custom review instructions'

Observed examples:

$ codex review --base main 'custom review instructions'
rc=2
error: the argument '--base <BRANCH>' cannot be used with '[PROMPT]'

Usage: codex review --base <BRANCH> [PROMPT]

$ codex exec review --base main 'custom review instructions'
rc=2
error: the argument '--base <BRANCH>' cannot be used with '[PROMPT]'

Usage: codex exec review --base <BRANCH> [PROMPT]

$ codex review --uncommitted 'custom review instructions'
rc=2
error: the argument '--uncommitted' cannot be used with '[PROMPT]'

Usage: codex review --uncommitted [PROMPT]

$ codex review --commit HEAD 'custom review instructions'
rc=2
error: the argument '--commit <SHA>' cannot be used with '[PROMPT]'

Usage: codex review --commit <SHA> [PROMPT]

Why this matters for automation: an orchestrator currently has to choose between native Codex review targeting (--base, --commit, --uncommitted) and per-run supplemental instructions such as acceptance criteria, non-goals, round context, or a machine-readable verdict contract. It cannot use both in one native review invocation.

Any of these outcomes would resolve the ambiguity:

1. allow supplemental instructions with native review targets, matching the current help text;
2. add an explicit option such as --instructions / --instructions-file; or
3. update the help/usage text to clearly state that prompts are not supported with native target flags.