ollama_setup: "Type yes to proceed" prompt is dead — login runs unconditionally and can block ~2 min

Bug

ollama_setup/register_callbacks.py — lines 342–351 (handler), 219–251 (_run_ollama_login).

After an auth failure the handler prints "Type 'yes' to proceed, or anything else to skip and do it manually later" (lines 343–347) and then calls _run_ollama_login() unconditionally (line 351) — the user's input is never read. The comment on 349–350 admits the contradiction ("We can't do interactive input here (custom commands shouldn't block) So we just run it").

_run_ollama_login runs subprocess.run(["ollama", "login"], capture_output=False, timeout=120) synchronously from inside the custom-command handler. Consequences:

1. A browser OAuth flow launches even when the user would have typed "no".
2. In a captive/non-interactive context ollama login can neither be answered nor succeed, so /ollama-setup ... freezes the app for up to 120 s — the exact stall the printed instructions promised the user they could avoid.

Suggested fix

Read the user's answer first (via the proper async approval/input seam, or make login opt-in like /ollama-setup --login), or drop the misleading "type yes or anything else" text and make the message match the actual behavior.

Start in ollama_setup/register_callbacks.py at lines 342–351, then inspect _run_ollama_login at lines 219–251 and the custom-command input or approval path. Reproduce the auth-failure flow and verify that login is no longer launched unconditionally or allowed to block for 120 seconds when the user declines or input is unavailable.