feat: bootstrap clickthrough server, skill docs, and syntax CI

docs/coordinate-system.md

@@ -0,0 +1,43 @@
+# Coordinate System
+
+All interactions ultimately execute in **global pixel coordinates** of the primary monitor.
+
+## Regions
+
+Visual endpoints return a `region` object:
+
+```json
+{"x": 0, "y": 0, "width": 1920, "height": 1080}
+```
+
+This describes where the image sits in global desktop space.
+
+## Grid indexing
+
+- Rows/cols are **zero-based**
+- Cell `(row=0, col=0)` is top-left
+- Each cell has:
+  - `cell_width = region.width / cols`
+  - `cell_height = region.height / rows`
+
+## Cell center formula
+
+Given `(row, col, dx, dy)` where `dx,dy ∈ [-1,1]`:
+
+- `x = region.x + ((col + 0.5 + dx*0.5) * cell_width)`
+- `y = region.y + ((row + 0.5 + dy*0.5) * cell_height)`
+
+Interpretation:
+- `dx = -1` -> left edge of cell
+- `dx = 0` -> center
+- `dx = 1` -> right edge
+- same concept for `dy`
+
+## Recommended agent loop
+
+1. Capture `/screen` with coarse grid
+2. Find candidate cell
+3. If uncertain, use `/zoom` around candidate
+4. Convert target to grid action
+5. Execute `/action`
+6. Re-capture and verify