feat: Industrial-Grade Reliability, Cascading LLM Fallback, and Tiered Service Support

[amaynez]

Overview

This PR transforms the MiroFish engine from a prototype into a robust, production-ready platform. It introduces a multi-layered resilience system for both LLM generations and external service (Zep) interactions, alongside a foundation for tiered service support.

Key Changes

1. Robust LLM Client with Cascading Fallback

• Truncation Detection: Automatically detects if an LLM response was cut off (e.g., due to token limits).
• JSON Repair: Implements smart logic to repair malformed or truncated JSON responses.
• Boost Fallback: If the primary LLM fails or returns broken JSON, the system automatically falls back to a high-capacity "Boost" model (configured via LLM_BOOST_* env vars).

2. Zep Resilience Layer

• Smart Retries & Rate Limiting: Added a sophisticated handling layer to avoid 429 errors and handle quota limits gracefully.
• Robust Paging: Re-implemented graph reading with robust paging to handle large-scale data without timeouts.
• Localized Error Handling: Improved error messages to inform users specifically when Zep quotas are exceeded.

3. Tiered Service Foundation

• Configuration-Driven Polling: Introduced a /config endpoint that allows the backend to control frontend polling behavior.
• Conditional Polling: The UI now dynamically enables/disables automatic graph updates based on the service tier (Free vs. Premium).
• Response Caching: Implemented server-side caching for graph data to optimize performance and reduce API costs.

Why this benefits MiroFish users

• No-Fail Simulations: Simulations are significantly less likely to crash due to minor AI hiccups or transient network issues.
• Clearer Feedback: Users are no longer met with generic "500 Internal Server Error" when external service limits are hit; they get clear, actionable messages.
• Scalability: The engine is now better equipped to handle large documents and complex simulations that previously caused timeouts or JSON parsing errors.

Commit Breakdown

1. feat(utils): implement robust LLM client with cascading fallback and JSON repair
2. feat(zep): add resilience layer with retries, rate limiting, and robust paging
3. feat(graph): refactor core services for high-availability simulations
4. feat(api): add tiered configuration and graph data caching
5. feat(frontend): implement conditional polling and service-tier UI
6. fix(i18n): add localized error messages for service quotas