This problem is often caused by layout shifting
Consider the steps a browser takes when loading a page:
- Step 1: Create the Document Object Model (DOM) and CSS Object Model (CSSOM)
- Step 2: Render tree is created - this contains only the nodes required to render the page. Layout computes the exact position and size of each object
- Step 3: From the render tree, the browser then computes the geometry of the layout and elements.
- Step 4: Browser paints the content (pixel by pixel) to create the visual representation seen by the user on the screen.

Learn more about how browsers render webpages.
When the browser is re-painting that's when layout shift happens. It looks like the content is being pushed down the page.
So what's most likely happening is that the browser is catching the (slightly) larger banner image above the fold during re-painting at the same time the last LCP is reported.
It's seen above the fold, but gets pushed down by the layout shift. Connection variability would explain the randomness.
I wrote an answer on How Chrome Measures LCP that will explain what I mean. I believe the OP was experiencing the same thing you are, except in your case, it is possible that you're seeing this because of the tool you're using.
You mention that you have a script that scrolls the page until the end to look for CLS issues. In my answer on how Chrome measures LCP, notice the last bullet in my explanation:
- Once the user starts interacting with the page, the last reported largest content is the Largest Contentful Paint time.
So the first thing you should do is try testing your site using a different tool that isn't doing anything like that. Personally, I use WebPageTest. In the Web Vitals report it will show you the LCP image and provide some useful data about it that I don't see in other tools.
If it's not the tool, you still know what you have to do: get rid of CLS.
Additional Resources
See my answer to a similar question where I throttle my network to show the CLS happening and another one about identifying the cause of Cumulative Layout Shift.