feat(tui): the backend choice reads like a choice

[plombeer31]

Stacked on #223 → #222 → #220.

Why

The three options told the operator what each backend is, never what it costs them. "Local models (llama.cpp) — download and run locally" does not say the download is 2.7–22 GB; "Cloud models — configure API key and pick a model" does not say it is the fastest way to a working agent; nothing said the custom-endpoint path downloads nothing at all. Those trade-offs — privacy, money, time — are the thing being chosen between.

After

  atomic-agent can drive models three ways. Nothing here is permanent — you
  can add the others at any time from the menu.

  ›  Local models        llama.cpp on this machine. Private, free per token,
                         one download of 2.7–22 GB.

     Cloud models        OpenRouter, Anthropic, Gemini, Groq and 20 more.
                         Fastest to a working agent — needs an API key.

     Custom endpoint     An OpenAI-compatible or llama-server URL you already run.
                         Nothing is downloaded, nothing else is asked.

  ↑/↓ move   enter select   1–3 jump   esc skip   ctrl+c quit

Details are wrapped by hand into two lines rather than left to Ink, and the labels share a 20-column measure so the details line up as a column of their own — a ragged left edge there makes three comparable options read as three unrelated ones. The [1]/[2]/[3] prefixes are gone from the rows; the digits still work and the footer still says so.

The rowDetails tier from onboarding-fit already governs whether the details are drawn at all, so a narrow terminal gets the labels alone.

Tests

onboarding-screen.test.tsx asserts the cost phrases are present and that the screen no longer contains "not reachable" or "ECONNREFUSED" — the failure report this screen used to open with.

Captured at 100×30; matches the reviewed design exactly.