Fix personal OpenAI setup, prompt presets, and Realtime transcription

Problem

The personal API key flow exposes only a hard-coded model choice and has several integration failures:

• Saved prompt presets appear selected but are never included in Ask or live-summary requests.
• The base prompt forces the assistant to identify itself as Pickle.
• OpenAI Realtime transcription uses the legacy session event/schema and considers the socket ready before OpenAI confirms the session.
• STT forces an English language hint and starts audio capture even when initialization failed.
• Current OpenAI chat models require different token/sampling parameters than the legacy GPT-4.1 request.

Proposed fix

• Persist and apply the active prompt preset.
• Separate custom context from conversation history.
• Add current OpenAI LLM/STT choices and compatible request parameters.
• Migrate transcription to the current session.update transcription schema, wait for session.updated, auto-detect language, and surface setup errors.
• Improve personal-key reuse/storage and streaming response parsing.
• Document OpenAI and fully local Ollama/Whisper setup.

Validation

• Full desktop renderer and web production builds pass.
• Request-shape checks cover legacy/current chat models and Realtime transcription session configuration.