Add fast single-counterfactual test runner

Add two-panel action chart, cross-platform fonts, and distractor placement fixes

Fix distractor off-frame placement and task string chart readability

Fix mandatory occlusion targeting wrong object and distractor overlap

Cast LLM-provided float params to int for array indexing

Auto-relocate distractor to empty background region

Use screen blend for lighting perturbation with clamped parameters

Disable contrast delta by default in lighting perturbation

Use gamma correction for lighting perturbation brightness

Fix lighting perturbation whiteout and recolor rainbow artifact

Fix NoneType comparison in cross-run recommendation generator

Refactor diagnostic agent into pluggable registry architecture

Raise default max_counterfactuals from 3 to 7

Restructure README for readability with progressive disclosure

Fix broken counterfactual visualizations with bbox fallback and improved comparison

Fix counterfactual visualization masks highlighting wrong regions

Add semantic representation probes and QK similarity analysis

Add spatial vs object disambiguation phase and improve counterfactual visualizations

Refactor CLI into reusable pipeline components for integrated diagnostic mode

Merge rule-based hypotheses with LLM output and raise budget to 7

Fix mandatory counterfactual tests and LLM confirms_on_change direction

Fix five correctness bugs and improve task_string_swap visualization

Add mandatory signals and counterfactuals for cross-run comparability

Improve diagnostic report clarity and presentation

Add rich console progress output to diagnostic agent

Enhance diagnostic agent with adaptive triage, robust LLM integration, and expert fallbacks

Add fallback detection for task-critical objects missed by primary threshold

Add run comparison reports with recommendations and fix diagnostic robustness

Use SigLIP text encoder for task string diversity embeddings

Update README with diagnostic agent documentation

Add agentic diagnostic system for automated model diagnosis

Add export_data to GPU config and make run scanner recursive

Rewrite README for clarity: add quick start, streamline structure

Rename model health to model internals across the full stack

Update minimum Node.js version to 20.19+ for Vite compatibility

Add Node.js 18+ installation to GPU setup script and clarify README

Add table of contents and fix README section hierarchy

Add web viewer documentation and screenshots to README

Add InfoPopover, fix overlay rendering, and polish compare view UX

Improve LLM UI: tabs, concise prompts, nav cleanup, and UX fixes

Fix missing comma in saliency image mapping dict

Add stats engine, run insights, and fix LLM saliency prompt context

Fix language diff export key mismatch and fall back to PNG images

Use Vite proxy for API requests to eliminate CORS port mismatch

Fix language-diff crash when config uses YAML boolean true

Reorder sidebar: Gradient above Attention, auto-select first available viz on run pick

Add --skip-attention flag to run gradient features without attention extraction

Add localhost:5175 to CORS origins and format cors_origins list

Add image URLs to viz responses and expand comparable viz types

Auto-detect VLM layer count and skip out-of-range layers