[andreBurnt]

Summary (Meta Global AI Developer Hackathon — Oct 26–Nov 4, 2026)

Adds Meta SAM 3.1 Promptable Concept Segmentation (POST https://api.meta.ai/v1/responses) as a native visual grounding fallback in linc-nerve for screens where uiautomator dump returns an empty or unlabelled UI hierarchy (custom Compose/Flutter/WebView canvases, games, and Wear OS watch surfaces).

What was built during the Hackathon window (after tag pre-hackathon-2026-10-26)

• src/linc_nerve/core/grounding.py:
  • ground_visual_targets(image_bytes, prompt, screen_size, api_key): Calls sam-3.1 on Stage-1 screenshots, parses wire records via meta-sam-parser==0.1.0, decodes 1-bit masks to compute true mask center-of-mass, and scales coordinates into native device tap space (1080x2340 phone, 432x432 Wear OS watch, iOS logical points).
  • render_grounded_overlay(...): Renders translucent green bounding box + red centroid crosshair + confidence badge for visual verification.
  • should_suggest_visual_grounding(...): Detects empty trees, failed dumps, or unlabelled clickable Canvas controls.
• src/linc_nerve/mcp/ui_tools.py: New nerve_ground(device, prompt, annotate=False) MCP tool + automatic visual_grounding_hint on nerve_elements and nerve_state.
• src/linc_nerve/mcp/input_tools.py & src/linc_nerve/mcp/_core.py: Semantic nerve_tap(device, prompt=...) resolution integrated before coordinate guards and action logging.
• tests/test_grounding.py: Unit and contract tests against live sam-3.1 wire output.

[andreBurnt]

Closing direct mirror PR — releasing via canonical linc-nerve -> linc-publish pipeline.