fix: start self-learning worker from TUI

[EvanProgramming]

Fixes #164

What changed

• Start the existing detached, per-workspace self-learning worker after a successful TUI provider initialization.
• Preserve the existing in-process background-loop fallback when worker spawning fails.
• Add a TUI startup regression test proving the worker hook is invoked before the backend reports ready.
• Make the wheel smoke test inspect entry points inside its isolated installed environment, so the required installed-TUI gate is reliable.

Root cause

The legacy Rich CLI called _ensure_detached_learning_worker(), but the primary Bubble Tea JSONL backend did not. v2.0.3 was the latest release, so normal TUI launches could exit without any background worker.

Validation

• make check
• make lint
• make test (212 tests)
• python3.12 -m unittest tests.test_tui_protocol tests.test_learning_worker tests.test_distribution -v
• ./venv/bin/python scripts/wheel_smoke.py
• make docs-check
• make shell-check
• git diff --check

No credentials or user data are included.