Add Omi QoS tier system for model cost optimization

[beastoin]

Backend LLM costs span multiple providers (OpenAI, Anthropic, Google/Gemini, Perplexity) across 60+ callsites. There is no unified way to control model selection per-feature — models are hardcoded throughout the codebase, making cost optimization and A/B testing impossible without code changes.

Current Behavior

• 60+ LLM callsites across 4 providers with hardcoded model instances
• OpenAI: 15+ features using llm_mini, llm_medium, llm_medium_experiment directly
• Anthropic: chat agent hardcoded to claude-sonnet-4-6
• OpenRouter: persona chat and wrapped analysis hardcoded to specific Gemini/Claude models
• Perplexity: web search hardcoded to sonar-pro
• No mechanism to downgrade/upgrade models per-feature without code changes
• No way to switch cost profiles (e.g., "run everything on cheapest acceptable models")

Expected Behavior

A provider-agnostic QoS profile system where each profile (mini/medium/high) maps every feature to a specific model — potentially different model tiers within the same profile, since some features need more quality than others even in a cost-optimized profile.

Solution

QoS Profiles — each profile is a complete feature→model mapping across all providers:

MODEL_QOS_MINI:
  conv_action_items:  gpt-4.1-nano       (cheapest, structured extraction)
  conv_structure:     gpt-4.1-mini       (needs more quality)
  chat_agent:         claude-haiku-3.5   (cost-optimized chat)
  persona_chat:       gemini-flash-1.5-8b
  ...

MODEL_QOS_MEDIUM:
  conv_action_items:  gpt-4.1-mini
  conv_structure:     gpt-5.1
  chat_agent:         claude-sonnet-4-6
  persona_chat:       claude-3.5-sonnet
  ...

MODEL_QOS_HIGH:
  conv_action_items:  gpt-5.1
  conv_structure:     o4-mini
  chat_agent:         claude-sonnet-4-6
  persona_chat:       gemini-3-flash-preview
  ...

Global switch: MODEL_QOS=mini selects entire profile
Per-feature override: MODEL_QOS_CONV_STRUCTURE=gpt-5.1 overrides one feature

21 features across 4 providers:

• OpenAI (16): conv_action_items, conv_structure, conv_apps, daily_summary, memories, memory_conflict, memory_category, knowledge_graph, chat_responses, chat_extraction, session_titles, goals, notifications, followup, smart_glasses, onboarding
• Anthropic (1): chat_agent
• OpenRouter (3): persona_chat, persona_clone, wrapped_analysis
• Perplexity (1): web_search

Pinned features: fair_use classifier pinned to specific model regardless of profile (accuracy-critical).

Affected Areas

Area	Files	Callsites
QoS core	utils/llm/clients.py	Profile definitions, get_model(), client factories
Conversation processing	utils/llm/conversation_processing.py	5 callsites
Memories	utils/llm/memories.py	4 callsites
Knowledge graph	utils/llm/knowledge_graph.py	2 callsites
Chat	utils/llm/chat.py	10+ callsites
Persona	utils/llm/persona.py	5 callsites
Goals	utils/llm/goals.py	3 callsites
Notifications	utils/llm/notifications.py	2 callsites
Agentic chat	utils/retrieval/agentic.py	1 callsite (Anthropic)
Wrapped	utils/wrapped/generate_2025.py	9 callsites (Gemini)
Other	Various routers/utils	10+ callsites

Impact

Unified cost control across all LLM providers. One env var (MODEL_QOS=mini) to switch the entire backend to cost-optimized models. Per-feature overrides for A/B testing. No user-facing changes.

---

by AI for @beastoin