UI redesign: consistent design system, card layout, unified colors and typography

Load preset image into patch explorer (not just GT brain)

Fix thread-safety race in brain surface interpolation; bake DINOv2 into Docker image

Add remove button to example presets in editable mode

Add Save as Preset button (--editable-examples flag); wire examples file in entrypoint

Fix Bokeh _pending_writes error by deferring on_change work; guard slider access against stale sessions

Bake CLIP model into Docker image; move active tiles render to background thread

Remove GT brain pre-fill on feature selection — only load when image is selected

Fix GT brain render race conditions with staleness tokens

Show NSD image ID in patch figure title; add image field to example presets

Add precomputed cortical profile cache lookup

Remove pyvista renderer; use nilearn only for all brain renders

Add GT brain render cache, NSD-only phi maps, and parallel precomputation

Add pyvista renderer, precomputed phi map cache, and background-thread steering brain renders

Move cortical profile render to background thread and cache KDTree IDW weights

Refactor: split panels into pure logic + UI layers

Parallelize image loading, add thumbnail cache, and 3-phase search

Make lambda strength indicator larger and clearer in active features display

Hide gallery when CLIP search results are shown

Move lambda below X; use text input to remove spinners

Clean up lambda input: smaller, no spinners, λ label

Remove darkness kwarg incompatible with newer nilearn

Expose lambda steering scale in active features UI

Add nilearn/nibabel/scipy; add get_fmri and beta_std to HTTPDynaDiffLoader

Sync dynadiff_loader.py: add beta_std property, get_fmri, updated timeouts

Fall back to CPU for backbone+SAE inference when no CUDA available

Full sync of explorer/ with local: args, inference, widgets, patch, examples

Sync brain.py: add missing _render_steering_b64, render_fmri_brain_compact_b64, get_dd_fmri, apply_steering_fmri

Comment out new features (CLIP×φ sort, λ input, brain MEI sidecar) until ready

Sort by CLIP×φ, per-feature λ input, brain MEI sidecar support, gallery in UI

Increase reconstruct HTTP timeout to 720s

Show model-loading status on first patch inference request

Fix DynaDiff cold-start UX: immediate queued message, no duplicate runs

Use nsd_img_idx from reconstruct response for GT image fallback

Fix patch image indexing and DynaDiff cold-start indicator

Enable CPU inference for patch backbone+SAE

Fix patch panel: image-not-found error message and GPU status label

Load GT image locally when Modal endpoint returns None

Fix Modal URL: replace underscores with hyphens in endpoint method names

Retry health check on timeout during Modal cold start

Fix orig_dir unbound variable in DynaDiffLoader._load

Fix HTTPDynaDiffLoader URL routing for Modal per-method endpoints

Move DynaDiff to Modal serverless endpoint; switch to CPU-only HF Space

Fix multi-session crash: match modules by file path, not __package__

Fix multi-session module clearing: also remove parent package attribute

Fix multi-session crash: clear Bokeh widget modules between sessions

Switch to modular explorer package (scripts/explorer/)

Sync space code

Sync space code

Fix patch explorer image lookup and DynaDiff loading status