Reduce AI request and token usage

[Luckycat133]

What changed

• enforce one configured text-model request per normal persona, Agent, or group turn, with no automatic fallback or hidden retry
• move titles, durable memory capture, Moments background content, quizzes, idiom games, palettes, deterministic math, and learning-state tools to local execution
• bound context and per-Agent output budgets; keep Scholar at one Tavily retrieval plus one bounded synthesis
• keep image and speech APIs behind explicit user actions, and cache repeat TTS playback
• standardize the OpenRouter preset on nvidia/nemotron-3.5-lightning:free
• update English/Chinese README, changelog, and the full review report with request budgets and observed provider behavior

Why

OpenRouter activity showed 276 calls spread across several models. In addition to bounded model-selection QA, the previous architecture had multiple hidden request producers: model-generated titles, LLM memory extraction, group fan-out, recall rewrites, recursive model-backed tools, proactive images, background Moment generation, per-question game generation, and automatic retries. That made request count, latency, and token use hard for users to control.

User impact

• ordinary chat and group turns now have a predictable one-model-request ceiling
• deterministic and background features use zero model requests
• provider failures remain visible and require an explicit manual retry
• cross-topic memory still works through local IndexedDB capture
• explicit search, image generation, and TTS retain their functionality with clear request boundaries

Real UX validation

• Aurora remembered the user's public-speaking plan, fear of forgetting, and green pen across topics; one model request per reply
• Muse translation completed with one model request and no second translation-tool call
• a two-character group selected one responder and sent one model request
• natural-language inventory math returned 66 with zero dynamic requests
• local science trivia and 10 seconds of Moments idle time generated zero dynamic requests
• Scholar used exactly one Tavily request plus one Lightning request and displayed both current OpenRouter official source links
• Pixel explicit image generation used one MiniMax request and zero OpenRouter requests
• repeat Read Aloud playback used the cached audio and made zero additional requests

Checks

• npm test — 31 files, 443 tests passed
• npm run lint — passed
• npm run build — passed
• production-preview browser request tracing — passed for the flows above
• staged diff secret/private-path/large-file checks — passed