Improve wizard AI feedback and provider setup

[akirk]

Summary

Follow-up to #14.

• Providers: Ollama (http://localhost:11434) and LM Studio (http://127.0.0.1:1234) are separate options with their own server URL; Anthropic/OpenAI keep an API key field.
• Check connection: lists models from the provider (Anthropic/OpenAI /v1/models, Ollama /api/tags, LM Studio /v1/models) into a model dropdown, runs automatically when the key/URL changes, and reports failures with a hint.
• Busy state: spinner in the Generate button, a "Thinking… / Writing… / Applying changes…" line in the log, the form fields and Download/Playground buttons locked while generating, and an explicit "Finished" line afterwards.
• Files panel: new / modified / deleted badges, latest change flashes.
• Continue: when the model hits its output limit a Continue button sends the follow-up; the prompt field relabels to "Follow-up prompt" after the first run so it's clear where refinements go.

Test plan

• Headless Chromium with mocked provider endpoints: model listing for Anthropic and LM Studio, failed Ollama probe message, busy state mid-run, fields locked and unlocked, badges, follow-up label, then a real embedded Playground boot.
• Real provider run.