Per-position accent borders on the use-case pills

GPU spam guard (serialize + button lock), live progress tickers, describe-first layout, real-machine bindings, chart fix for custom VRAM

Fix spec-parser GPU window; clearer chart colours, unclipped label

Paste box goes live: fine-tuned spec parser fills the form from messy text

Load speed predictor via xgboost native format

Speed predictions now come from a model trained on 6.6k real measurements (17.5% median error on unseen hardware), physics formula beyond its measured range

Speed predictions with receipts: bandwidth roofline, real-runs chart, honest provenance

Multi-goal checks (union buying advice), fix vision/audio quality bars, honest niche fallback

Add einops for the Mamba hub kernel

Fix kernels version clash: pin to transformers' declared range

Real catalogue: 83 verified models, buy-advice mode, live model lookup, license-aware cards

Speed up explainer: prebuilt Mamba kernels (transformers 5.8 + kernels hub), tighter answer budget

Deploy FitCheck: engine + Nemotron model brick on ZeroGPU

Deploy FitCheck: engine + Nemotron model brick on ZeroGPU

Deploy FitCheck UI brick

initial commit