Owner: you own ONLY argus/vision/__init__.py, argus/vision/marks.py, argus/vision/locate.py, tests/test_vision.py, tests/fixtures/canvas_*.html. Read AGENTS.md, argus/models.py, argus/healing/resolver.py, argus/llm/client.py (LLMClient.json(..., tier="vision", images=[bytes])). Installed: opencv-python-headless, pillow, imagehash, numpy.

Principle: vision is a SECOND perception channel used only where the DOM is insufficient (canvas maps, video, WebGL). The model never emits raw coordinates; it picks a mark id and Argus resolves it.

1. marks.py
   • VisualMark (pydantic): state_key: str (url path + viewport), bbox: [x,y,w,h], center_norm: [fx, fy] (relative to the canvas element), canvas_selector: str, crop_png_b64: str (small 48-96px crop), phash: str (imagehash.phash of the crop), nearby_labels: list[str], hint: str ("drone marker Falcon-1").
   • async def capture_mark(page, canvas_selector, x, y, hint, size=64) -> VisualMark (screenshot the canvas element, crop around (x,y) in canvas coords, compute phash).
2. locate.py — cascade, each returns (x, y, tier, confidence) in PAGE coordinates or None:
   • V0 cached: canvas bbox * center_norm, accepted if the crop at that point has phash distance <= 8.
   • V1 template: cv2.matchTemplate (TM_CCOEFF_NORMED) of the stored crop over a canvas screenshot; accept max score >= 0.80 and second-best peak (outside a 20px radius) at least 0.05 lower.
   • V2 multiscale: repeat V1 at scales 0.8..1.25 (zoomed maps).
   • V3 vlm: draw numbered set-of-marks candidates (up to 12 local maxima of the template response, plus a coarse 4x4 grid if fewer) onto the screenshot with PIL, send to llm.json(tier="vision") asking for {"mark": <id or null>, "confidence": 0..1} given the hint; map id -> coordinate.
   • async def locate(page, mark, llm=None) -> dict | None running V0..V3 in order, never V3 without an llm.
3. Tests (no network, no LLM): tests/fixtures/canvas_map.html draws a map-like canvas with 4 drone icons from a seeded layout (icons as small distinct shapes + labels drawn on canvas; a query param ?shift=40&zoom=1.1 moves/scales everything). Capture a mark on the base layout, then assert V0 finds it unchanged, V1 finds it after shift=40, V2 after zoom=1.15, and that a click at the returned point hits the right icon (the page records the last clicked icon id in window.lastHit). Run .venv/Scripts/python.exe -m pytest -q tests/test_vision.py until green, then print a concise summary. Do not modify other argus modules; describe the integration points (resolver/runner) in your summary.