feat: configurable LLM API usage rate limiting and automatic 429 retry

[luquiluke]

Summary

Large simulations (300+ agents, 80+ rounds) generate hundreds of LLM calls per run and reliably hit provider rate limits. This PR adds a complete API usage rate limiting layer that prevents simulation crashes and gives users control over throttling behavior.

Changes

Backend — backend/app/utils/llm_client.py

• _TokenBucket class — fixed-window per-minute rate limiter. Tracks RPM and TPM counts, resets every 60s, sleeps until window resets if limit is reached
• _is_rate_limit_error() — detects 429 errors across OpenAI SDK, Anthropic SDK, and string fallback ("429", "rate limit", "too many requests")
• _check_token_bucket() — proactive enforcement before each LLM call; sleeps if over RPM/TPM limits
• chat() retry loop — catches rate limit errors and retries with exponential backoff: wait = min(base_delay * (2 ** attempt), 300). Defaults: base 30s, cap 300s, 3 retries
• chat_json() — updated to accept and pass through rate_limit_config

Backend — backend/app/api/simulation.py

• POST /<simulation_id>/config — new endpoint that merges a rate_limit object into the simulation's simulation_config.json without touching other fields

Backend — simulation scripts (all three)

• backend/scripts/run_twitter_simulation.py
• backend/scripts/run_reddit_simulation.py
• backend/scripts/run_parallel_simulation.py

Each script now:

1. Reads rate_limit config from simulation_config.json before the round loop
2. Wraps env.step() in a retry loop for 429 errors from camel-ai
3. Calls asyncio.sleep(inter_turn_delay_s) after each step

Frontend — frontend/src/api/simulation.js

• updateSimulationConfig() — new API helper that calls POST /api/simulation/{id}/config

Frontend — frontend/src/components/Step3Simulation.vue

• Collapsible Rate Limit Settings panel, visible only in pre-run state (phase === 0)
• 5 controls: Inter-turn Delay slider (0–5000ms), Max Retries, Retry Base Delay, TPM Limit, RPM Limit
• Settings load from and save to localStorage under key mirofish_rate_limit_settings
• updateSimulationConfig() called before startSimulation() on every run

Configuration

All parameters are optional with safe defaults. Set via UI or directly in simulation_config.json:

{
  "rate_limit": {
    "inter_turn_delay_ms": 500,
    "max_retries": 3,
    "retry_base_delay_s": 30,
    "tpm_limit": 0,
    "rpm_limit": 0
  }
}

Set tpm_limit and rpm_limit to 0 to disable proactive throttling and rely on retry-only behavior.

No breaking changes

• LLMClient.__init__ signature unchanged — rate_limit_config is passed per-call, not at construction
• Existing GET /<simulation_id>/config endpoint untouched
• All new behavior is opt-in via config; default behavior is unchanged if rate_limit is absent from config