feat: improve rate limiting with conservative thresholds and exponent…

[aidenerdogan]

Changes

Rate Limiting Improvements

• Conservative threshold: Wait at 90% of rpm_limit instead of 100%
• Increased safety margin: 2s instead of 1s
• Better logging: Shows current request count (X/Y format)
• Cleanup after waiting: Removes stale timestamps

Exponential Backoff Retry

• Auto-retry on RateLimitError: Up to 3 attempts
• Exponential backoff: 5s → 10s → 20s between retries
• Graceful degradation: Continues on success, fails after max retries

Error Handling

• Exception detection: Checks if batch items are exceptions
• Response validation: Validates that the choices attribute exists before access
• Detailed error messages: Includes batch item index for debugging

Testing

Tested with Mistral API (rpm_limit: 7), generating 285 MCQs without rate limit errors.

Fixes rate-limiting issues when using strict API providers.

[patrickfleith]

LGTM