Add auto-advance: click Next, scroll to top, capture next page

After each capture the controller asks the vision model to locate a 'Next'
control (returned as normalized screen coordinates), clicks it via OS input,
scrolls to top, and continues the loop. Configurable via SCREEN_LEADS_AUTO_NEXT
/ _NEXT_LOAD_PAUSE / _MAX_AUTO_NEXT, with a per-run safety cap.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

agent/scroller.py

@@ -40,3 +40,18 @@ def scroll_down(amount: int = SCROLL_AMOUNT) -> None:
     pg = _pyautogui()
     pg.scroll(-abs(amount))
     time.sleep(SCROLL_PAUSE)
+
+
+def click_at_fraction(x_frac: float, y_frac: float) -> None:
+    """Click at a position given as fractions (0..1) of the logical screen.
+
+    Fractions are resolution- and Retina-independent: pyautogui reports the
+    logical screen size, and the capture covers the same primary display, so a
+    fraction of the screenshot maps directly to a fraction of the screen.
+    """
+    pg = _pyautogui()
+    w, h = pg.size()
+    x = int(max(0.0, min(1.0, x_frac)) * w)
+    y = int(max(0.0, min(1.0, y_frac)) * h)
+    pg.moveTo(x, y, duration=0.2)
+    pg.click()