Interactive Fieldmark bird identification on GitHub Pages

[du000362]

Summary

• publish the complete Bird Detector Python project
• export the trained EfficientNet-B0 checkpoint to an 18 MB ONNX browser model
• add private on-device inference for all 555 NABirds classes
• redesign Fieldmark with responsive light/dark themes, batch uploads, calibrated alternatives, uncertainty details, and CSV export
• add a GitHub Pages workflow and clear local/browser documentation
• refresh the Streamlit interface for Streamlit 1.59

Validation

• pytest -q — 10 passed
• static-bundle validation — passed
• WCAG AA contrast validation — passed
• JavaScript syntax checks — passed
• deterministic and real-photo PyTorch/ONNX parity — passed
• local static HTTP and Streamlit health checks — passed

Privacy and repository scope

• selected photos are processed in browser memory and are not uploaded
• NABirds images, Python checkpoints, reports, and local editor state remain excluded from Git
• only the optimized browser model is versioned