Add research attention analysis endpoint with real CodeGen tokenization

Add research attention analysis endpoints with Q/K/V extraction

Fix ablation study for Code Llama compatibility

Fix model info endpoint for Code Llama compatibility

Add Code Llama 7B support with hardware-aware filtering and ICL timeout fixes

Fix pyarrow compatibility issue with datasets library

Add SWE-bench integration and improve backend routing

Add layer_stride parameter for PromptDiff optimization

Capture complete attention patterns after generation

Fix: Use scaling approach instead of skipping layers

Fix: Refine layer hook output format handling

Fix: Handle single-element tuple outputs in layer hook

Fix: Correct layer hook output format for layer_norm compatibility

Fix: Prevent hook persistence after ablation errors

feat: Add pipeline analyzer and QKV extractor for transformer visualization

Add backend support for ICL emergence analysis

Add ablation support to model service with comprehensive testing

Add CORS support for Vercel production domain

Add API key authentication

Fix backend structure - remove duplicates