refactor: split dictation cleanup from agent invocations + inference scope & provider registries

[gabrielste1n]

Summary

Three layered abstractions that together let the wake-word agent path use its own model and prompt, separate from text cleanup.

1. Prompt registry (`src/config/prompts/`)

• New `PROMPT_KINDS` map (cleanup, dictationAgent, chatAgent) with i18n key + fallback per kind.
• Single `customPrompts: Record<PromptKind, string>` store field, store-backed and reactive.
• One read API: `resolvePrompt(kind, opts)`.
• Migration runs only if legacy `customUnifiedPrompt` / `agentSystemPrompt` keys exist; fresh installs are no-ops, no empty values written.

2. Inference scope resolver (`src/config/inferenceScopes.ts`)

• `INFERENCE_SCOPES` map with `fallbackScope` baked in (noteFormatting → dictationCleanup).
• `selectResolvedLLMConfig(state, scope)` + `setResolvedLLMConfig(scope, patch)`.
• `selectResolvedMeetingReasoning` rewritten as a thin wrapper.

3. Dictation-agent split (user-facing feature)

• New `dictationAgent` scope with its own model, provider, mode, custom API key, etc.
• Wake-word detection extracted to `config/agentDetection.ts`.
• `audioManager.processTranscription` routes to dictationAgent model + prompt when wake word fires AND a model is configured. Falls back to existing single-model behavior otherwise.
• New 4th LLM settings tab "Dictation Agent" with `` panel + `<PromptStudio kind="dictationAgent">`.
• i18n keys added across all 10 locales (English authoritative; others fall back to English per project i18n rule).

4. Provider registry (`src/services/ai/inferenceProviders/`)

• 7 of 8 providers extracted into single-purpose modules with shared `InferenceProvider` interface.
• `ReasoningService.processText` switch reduced to a registry lookup (~510 LOC removed from ReasoningService).
• OpenAI/custom path stays inline pending careful relocation of OpenAI's endpoint-discovery state (cache, probing, base resolution) into its own module.
• All `logger.logReasoning` event names preserved.

Backward compatibility

• Existing flat store fields (`reasoningModel`, `agentModel`, `meetingReasoningModel`, etc.) untouched.
• Existing localStorage keys, IPC channels, service classes, and logger event names preserved.
• Migration is idempotent and conditional. `agentSystemPrompt` field removal is the only breaking shape change; all consumers updated to use `customPrompts.chatAgent`.

Test plan

• Fresh install: no `_promptsMigrated` flag until store init; afterwards no other prompt keys written.
• Upgrade with custom `customUnifiedPrompt`: parsed once, copied to `customPrompts.cleanup` AND `customPrompts.dictationAgent`, legacy key removed.
• Upgrade with custom `agentSystemPrompt`: copied to `customPrompts.chatAgent`, legacy key removed.
• Upgrade with neither: migration runs once, writes nothing, sets flag.
• PromptStudio edits propagate to chat overlay live (Zustand reactivity).
• Dictate "Hey {agentName}, summarize this" → routes to dictationAgent model + prompt when configured; routes to cleanup model + FULL_PROMPT when not (existing behavior).
• Smoke each provider through dictation pipeline: OpenAI, Anthropic, Gemini, Groq, Local llama, OpenWhispr cloud, LAN, plus one enterprise (Bedrock/Azure/Vertex).
• Each provider logs the same event names as before (grep `ANTHROPIC_START`, etc.).
• Streaming paths (chat overlay) still work — untouched.
• `tsc --noEmit` clean, `npm run lint` clean, `npm run format:check` clean.

Follow-ups (deliberately out of scope)

• Relocate OpenAI endpoint-discovery state (`openAiEndpointPreference`, `probedBases`, `getConfiguredOpenAIBase`, `detectReasoningServerType`, `getOpenAIEndpointCandidates`) into `inferenceProviders/openai.ts`.
• Generic `<InferenceConfigEditor scope="...">` UI to replace 16-prop `` plumbing across SettingsPage, AgentModeSettings, MeetingSettings, DictationAgentSettings (~400 LOC removable).
• Provider override in `ReasoningConfig` so dictationAgent and cleanup can use different non-heuristic providers (currently the model heuristic in `getModelProvider` covers standard model IDs).