feat(hints): route hints + brand header to stderr; drop TTY suppression

[christophergeyer]

Summary

The advisory `hint:` lines and the `🦖 roar <sub>` brand header used to self-suppress in non-TTY contexts (CI, redirected output, agent capture). Stdout stayed clean for pipelines, but agents like Claude Code that capture roar output programmatically saw zero of the nudges — exactly the audience that benefits most.</sub>

This swaps the suppression for the stream choice:

• Stream: hints + brand header go to stderr (`click.echo(..., err=True)`). Stdout stays machine-clean for pipes/redirects without the gate having to do it.
• Gate: `hints_should_print()` drops the `sys.stdout.isatty()` check. Fires whenever `hints.enabled = true` (default) and `output.verbosity != "quiet"`.
• Color/emoji: ANSI styling now keys off `sys.stderr.isatty()`, so captured logs stay plain.
• `run_report.next_steps_hint` drops its parallel `pipe_mode` check for the same reason.

User-visible change

CI users who captured roar output without explicitly setting `hints.enabled = false` will start seeing hints on stderr. The escape valve hasn't changed:

```
roar config set hints.enabled false

or

roar config set output.verbosity quiet
```

Worth a release-note line.

Why this shape

• Trading the gate for a stream change preserves pipeline cleanliness — `roar dag | jq …` and `roar show > out.txt` still work, by virtue of where the line goes, not by suppressing it conditionally.
• Aligns with the git/npm/cargo convention of advisory output on stderr.
• Agents and CI logs capture both streams; both audiences now see the nudges without needing config tweaks.

Test plan

• `tests/unit/test_cli_format.py` (new) — direct stream-invariant tests for `hints_should_print`, `make_hint_printer`, and `print_brand_header`.
• `test_init_hints_appear_in_non_tty_on_stderr` + `test_init_brand_header_appears_in_non_tty_on_stderr` — agent-visibility regression guards (assert presence on stderr, absence from stdout).
• `test_next_steps_hint_appears_in_pipe_mode_for_agents_and_ci` (run_report).
• Existing config-disable + quiet-suppression tests kept and passing.
• 942 unit tests passing, 1 pre-existing skipped.
• `ruff check .` + `ruff format --check .` + `mypy roar` all clean.

🤖 Generated with Claude Code