fix: make loop detection thresholds configurable via config.yaml

[Ricardo-M-L]

Summary

• Problem: LoopDetectionMiddleware uses hardcoded thresholds (warn_threshold=3, hard_limit=5) that are not exposed in config.yaml. This causes false positives with smaller context window models (e.g. DeepSeek) on legitimate retry-heavy tasks.
• Fix: Add a loop_detection section to config.yaml with four tunable parameters (warn_threshold, hard_limit, window_size, max_tracked_threads). Both the create_deerflow_agent factory and the make_lead_agent application factory now read these values from config.
• Backwards compatible: Default values match the original hardcoded constants (3, 5, 20, 100), so existing setups are unaffected.

Changes

File	Description
config.example.yaml	New loop_detection section with documented defaults
deerflow/config/loop_detection_config.py	New Pydantic config model with validation (ge constraints)
deerflow/config/app_config.py	Added loop_detection field to AppConfig
deerflow/agents/features.py	Added loop_detection field to RuntimeFeatures
deerflow/agents/factory.py	Wire config values into LoopDetectionMiddleware constructor
deerflow/agents/lead_agent/agent.py	Wire config values into LoopDetectionMiddleware constructor
backend/tests/test_loop_detection_config.py	229-line test suite covering config defaults, validation, custom thresholds, and DeepSeek-friendly scenarios

Example usage

# config.yaml — raise thresholds for DeepSeek or retry-heavy workflows
loop_detection:
  warn_threshold: 5
  hard_limit: 10

Fixes #2116

Test plan

• Existing test_loop_detection_middleware.py tests pass unchanged (backwards compat)
• New test_loop_detection_config.py tests verify: default values, custom values, Pydantic validation, higher thresholds avoiding false positives, DeepSeek-friendly config scenario
• Config loading: AppConfig correctly parses loop_detection from YAML and falls back to defaults when omitted

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}