Exhale

Resolved root cause CI failures and optimized setup times by 10-100x.

• Replaced pip with uv package installer + cached virtualenvs
• Added package-lock.json, fixed Rust formatting
• Fixed Gradle deprecations
• Integrated Cursor agent for automated CI fixes
• Fixed android-unified workflow

Fixed critical gradient flow issue and completed PyTorch→JAX migration cleanup.

• HJEPA gradient bug: 16 of 19 predictor modules (~5.5M parameters) were not receiving gradients
• Root cause: losses recomputed from targets instead of predictor outputs
• Cleaned up legacy PyTorch training code (−49,824 lines)
• Added XLA compilation cache (15-30 min faster restarts)
• Added extracted training modules: quality_gates, evaluation, training_step

Removed legacy services and consolidated duplicated code across the codebase.

• Renamed semantic matchers/routers for clarity (59,072 lines affected)
• Removed empty stubs, dead code; added static analysis tools (vulture, jscpd)
• Consolidated to central repository, removed legacy services
• Split models.py (2,395 lines) into 18 domain-specific modules
• Consolidated 3 singleton registries into utils/singleton.py

Byzantine security audit identified and fixed critical vulnerabilities.

• RCE vulnerability: Pickle deserialization allowed arbitrary code execution
• Fix: implemented whitelist-only deserializer in safe_pickle.py
• Improved test quality for smarthome controller and integrations
• Added unified crypto roundtrip tests

World model fixes, test additions, and cleanup.

• EFE math fix: Fixed Expected Free Energy calculation
• Added prioritized replay and intrinsic motivation
• Added comprehensive E8/Fano mathematical correctness tests
• Implemented catastrophe_kan.py with KAN layers
• Removed .stagehand-cache from git tracking

Browser automation consolidation for verification tools.

• Removed Puppeteer from MedVerify
• Stagehand is now the sole browser tool
• Part of broader verification platform expansion work