fix: clear llmConfig when all LLM presets are deactivated

[AndrewDongminYoo]

Problem

When the user toggles off the last active LLM preset, toggleActive() sets activePresetId to null and calls persist(providerConfigs, null). Inside persist(), the if (newActive) block is skipped, so setLlmConfig() and saveLlmConfig() are never called.

The result: llmConfig in the store still holds the previous provider's config. Background queues (ingest-queue, dedup-queue, sweep-reviews, clip-watcher) all guard themselves with hasUsableLlm(llmConfig), which reads from the un-updated llmConfig and continues to return true — so processing carries on even though the user explicitly disabled all providers.

This is most visible with keyless providers (Claude Code CLI, Codex CLI, Ollama): turning them off has no observable effect on ingest.

Fix

Add an else branch to persist() that, when newActive === null, writes llmConfig to a state where hasUsableLlm() returns false:

} else {
  const cleared: LlmConfig = { ...llmConfig, provider: "openai", apiKey: "" }
  setLlmConfig(cleared)
  await saveLlmConfig(cleared)
}

Setting provider to "openai" (a keyed provider) and clearing apiKey makes hasUsableLlm() return false regardless of what the previous provider was — including keyless local CLI providers that hasUsableLlm would otherwise always pass.

resolveConfig() is safe: on re-enable it reads provider and apiKey from providerConfigs[newActive] (the user's saved per-preset settings), not from llmConfig, so the cleared values here do not affect the user's stored configuration.

Testing

1. Configure any LLM preset (including a keyless one like Claude Code CLI).
2. Start an ingest run.
3. Toggle the active preset off — the toggle turns grey and activePresetId becomes null.
4. Observe that in-progress tasks complete but no new tasks are picked up from the queue.
5. Toggle the same preset back on — processing resumes with the correct config.