fix: deduplicate OmniParser init and reuse HeuristicEngine as singleton

[Pranjal0410]

fix: remove duplicate OmniParser initialization and store HeuristicEvaluationEngine as singleton

Problem

Two issues in the current startup and request handling:

1. Duplicate OmniParser initialization in main.py — OmniParserClient() is created and initialized twice during startup (lines 36–37 and lines 44–45), meaning the model is loaded into memory twice unnecessarily.

2. HeuristicEvaluationEngine re-created on every request — In evaluation.py, the /evaluate endpoint creates a new HeuristicEvaluationEngine() and calls await evaluation_engine.initialize() on every single request. This re-initializes the OpenAI client and the RAG knowledge base each time, adding unnecessary latency and memory overhead.

Fix

• main.py: Removed the duplicate OmniParserClient initialization. Stored HeuristicEvaluationEngine on app.state.heuristic_engine as a singleton, matching the existing pattern used for app.state.omniparser_client.

• app/api/routes/evaluation.py: Updated the /evaluate endpoint to use request.app.state.heuristic_engine instead of creating a new engine instance per request.

Files Changed

File	Change
main.py	Removed duplicate OmniParser init; added app.state.heuristic_engine singleton
app/api/routes/evaluation.py	Use singleton engine from app.state instead of re-creating per request

Testing

• Existing integration test (tests/test_integration_real.py) is unaffected — it creates its own engine instance directly.
• The /evaluate endpoint behavior is unchanged; only the initialization strategy differs.