[Bug] Per-repo "chat provider" picker also selects the writer for dashboard-launched generate jobs, overriding config.yaml

Describe the Bug

The dashboard's per-repo provider picker is documented and labelled as the chat provider (server/provider_config.py:set_active_provider docstring: "overrides only that repo's chat"), but server/job_executor.py (~line 403) resolves dashboard-launched index/generate jobs through the same get_chat_provider_instance(repo_path, repo_id=…). The per-repo repos[repo_id] entry is step 1 of the resolver, so it outranks the repo's .repowise/config.yaml provider/model set at repowise init.

This becomes a trap because claude_cli / codex_cli do not implement stream_chat, so dashboard chat fails with "selected model doesn't support streaming" for anyone whose writer is a CLI seat. The only way to make chat work is to set a cheap API model in the picker — which then silently becomes the docs writer for any job started from the UI.

Steps to Reproduce

1. .repowise/config.yaml: provider: claude_cli, model: claude_cli/claude-opus-5.
2. Dashboard → repo Settings → provider gemini / gemini-3.5-flash-lite (needed for chat to stream). PATCH /api/providers/active {provider, model, repo_id} persists it under repos[repo_id] in ~/.repowise/provider_config.json.
3. Start an index/generate job from the dashboard.

Expected Behavior

Concept pages written with claude_cli/claude-opus-5 (the repo's configured writer); chat uses gemini.

Actual Behavior

Pages are written with gemini-3.5-flash-lite. The job_executor.py comment says the picker is "the only one that carries a deliberate user decision", but config.yaml's provider/model is also deliberate and is the value every CLI path uses — so UI-launched and CLI-launched jobs write with different models.

Environment

• OS: macOS 26 (arm64)
• Python: 3.13
• Repowise: 0.50.0 (uv tool install)
• Claude Code CLI / Codex CLI 0.154.0 as subprocess providers

Additional Context

Suggested fixes (either): (a) separate chat_provider/chat_model from provider/model in repos[repo_id], with the picker writing the chat pair; or (b) in job_executor, prefer config.yaml when it names a provider and fall back to the picker only when it doesn't. Also worth a UI hint that CLI-backed providers can't be chat providers.

Start in server/provider_config.py at set_active_provider and trace the resolver used by server/job_executor.py around line 403, comparing the repos[repo_id] picker value with .repowise/config.yaml provider/model settings. Reproduce with a CLI writer and Gemini chat provider, then verify that dashboard chat still uses the picker while dashboard-launched index/generate jobs use the configured writer.