Add display mode toggle to compare view and .env.example

Add display mode toggle for heatmap visualizations, OpenAI base URL support, and .env config

Fix compare runs: fetch all viz types in parallel and render chart data

Add web viewer with structured data export and serve subcommand

Add zoom and pan controls to image viewers in web frontend

Improve visualization interpretability and add --image-map flag

Add --config flag and enable all features in gpu.yaml

Add clone-and-setup script and auto-checkout extended-attribution in setup

Use system Python 3.10+ instead of installing Python 3.11 via deadsnakes

Add GPU config and Ubuntu 20.04 setup script for RTX 3080 laptop

Add extended attribution features: per-step cross-attention, GradCAM connector/VLM layers, vision-vs-state ablation, per-action-dim saliency, and language-diff comparison

Add GradCAM worked example with step-by-step numerical walkthrough

Add SmoothGrad option for noise-reduced saliency maps

Rearrange grid: interpretable overlays at bottom, raw maps at top

Update example grid asset with 7-row gradient attribution output

Add gradient-based attribution (saliency + GradCAM) for SmolVLA

Refactor inspect_attention.py into smolvla_inspect package

Add --task flag to override language instruction from CLI

Refresh README assets with threshold-corrected heatmaps

Add percentile thresholding and per-head baseline subtraction for SigLIP heatmaps

Crop padding patches from SigLIP attention heatmaps

Sync README examples with defaults.yaml and add health report assets

Add interpretability roadmap to README

Add model health diagnostics and architecture documentation

Fix co-attention arrow alignment in architecture diagram

Regenerate architecture diagram to show full pipeline

Update README to reflect current visualization pipeline

Improve attention visualization: cyan co-attention, clearer labels, quieter loading

Move supplementary docs into docs/ directory

Initial commit: SmolVLA attention visualizer