# GroundingPI Studio r8

User-requested layout and visualization revision, 2026-09-16. The same UI files
serve the standalone demo and the private Hugging Face Space.

## Changes

- Container-relative brand sizing prevents the GroundingPI title from clipping.
- Bounded sidebar, consistent canvas/query gutters, visible desktop action, and
  stacked mobile controls. Native image sharing is explicitly disabled.
- Request identifiers let the browser distinguish consecutive identical outputs.
  The immediate client acknowledgement owns a separate node; it never replaces
  Gradio/Svelte-managed status DOM. Completion/error clears the timer and inert state.
- Category chips sit next to the canvas workflow. Selection shows original pixels,
  exact returned geometry, category-colored focus with a contrast halo, instance
  navigation and a magnified crop. Closing focus restores every prediction.
- Dense boxes use fine edges and short corner marks, with no filled interiors.
  No prediction is suppressed, merged, moved, deduplicated or replaced.
- Box labels appear in sparse and dense scenes. Each exact category gets at most
  one on-image name across its instances. Labels avoid each other and small boxes;
  nearby free space and short leader lines handle crowded anchors. The category
  inspector always retains every category, including labels that cannot physically
  fit on a tiny canvas. Font weight is explicitly set for legible raster text.
- Empty detail panels stay hidden; raw prompt/output remain available in accordions.

## Contract

The 27 examples, 9 task routes, protocol parser, prompts, native inference backend,
model revision and generated coordinates are unchanged. Dataset annotations are
never substituted for predictions. The original r7 manifest remains as provenance;
`ui_studio_r8_lock.json` records the reviewed new UI bytes and unchanged case contract.

## Verification scope

Console request lifecycle, protocol/render regressions and packaging lock checks;
browser review at mobile, intermediate and desktop widths; success, error and repeat
requests; category/instance focus, zoom/reset, example switching and upload reset.
Local visual regression uses a separate, non-deployed HTTP fixture replaying a
recorded real 27-target HF response. Its latency is test latency. The deployed HF
Space is checked with a fresh real model request after release; these checks are
qualitative UI acceptance, not a benchmark.