fix: Grad-CAM works in ensemble mode (shim ensemble pieces into expected interface)

fix: re-enable DermLIP via timm Eva (EVA-01-Base, strict load)

move fallback checkpoint off Drive, pull from HF Model repo Kabirgrover/dermtriage-weights

remove orphaned DermLIP disabled-stub block (syntax error)

use duplicated Drive file ID for fallback checkpoint (fresh quota)

disable DermLIP (EVA-ViT mismatch); graceful fallback on Drive quota exhaustion

DermLIP loader v2: build vision tower only (skip BERT text encoder mismatch)

patched DermLIP loader: strip unknown config fields, manually load weights (Path B follow-through)

MedGemma prompt: decision-oriented opening, 3-sentence cap, no numbered lists, no filler

flat corners + demo-specific explanation handler (don't toggle Step 3 in demo mode)

custom dark navy theme (Base + comprehensive CSS): wider layout, polished cards, locked palette

switch to gr.themes.Default() — better HF dark-mode contrast than Soft()

revert title/intro/footer back to gr.Markdown — native is theme-aware, my gr.HTML was the broken part

title/intro/disclaimer: gr.HTML with CSS vars instead of gr.Markdown (.prose dimness fix)

clean architecture: pure Gradio defaults + CSS vars throughout custom HTML, drop all hardcoded colors and pastels

force Markdown .prose to use block_label_text_color (same var component labels use, works in both modes)

back to Gradio defaults: drop LIGHT_THEME, make triage card theme-aware via CSS variables

use Gradio theme.set() to override _dark variant variables (proper fix for invisible text)

fix invisible text: set explicit dark text color + color-scheme:light on container

force white background on container body so title is readable when HF chrome is dark

fix Grad-CAM: preprocess_pil returns (tensor, pil) tuple, not just tensor

strip custom CSS/JS; use Gradio defaults; rename gallery 'Curated DDI' to 'Example cases'

try standard open_clip_torch>=2.30 for DermLIP (Path B ensemble attempt)

force light mode aggressively via JS + CSS overrides

wire Grad-CAM heatmap into demo flow (classification -> heatmap -> rationale)

flip pilot mode to demo flow: hide Step 1/3, auto-run on upload, strip mode label, force light theme

pin huggingface_hub<1.0 (HfFolder removed in 1.0, still imported by gradio[oauth])

bump Gradio to 5.0.0, drop now-unneeded compat pins

bump Gradio to 5.0.0

pin jinja2==3.1.4 to dodge unhashable-dict cache regression

patch _json_schema_to_python_type to short-circuit bool schemas

monkey-patch gradio_client bool-schema bug

drop gradio_client pin (strict-pinned by gradio)

pin gradio_client>=1.5.0 to fix bool-schema TypeError

bump gradio sdk to 4.44.1

pre-emptive deps: pandas, tqdm, matplotlib, accelerate

add scikit-learn (transitive import via src.__init__)

pin huggingface_hub<0.31 for Gradio 4.44 HfFolder compat

pin Python 3.11 to fix audioop import on 3.13

fix HF Space build: drop dead open_clip fork dep

updated for new HF space

Add DermTriage Gradio demo app

initial commit