Why does SSIM fail to predict image quality on cartoon images?

I was planning on doing an image codec comparison thread going from reasonable 0.93 to 0.96 SSIM range showcasing how much file size vs quality goes up across different image codecs. In theory anything above a 0.90 SSIM but below a 0.95 SSIM score is correlated with subjectively good quality. Not great quality, just good, like a quarter pounder from the clown instead of a properly made cheeseburger from a 5 star restaurant. Again in theory the closer you get to 0.99 SSIM the more overkill the lossy compression with an expected huge file size increases with little quality improvements.

A quality parameter of 50 with webp should not be getting anywhere near 0.99 SSIM...

Here's the webp file: https://files.catbox.moe/juhk01.webp

And picrel is the source file used for the test.

JPG result doesn't make sense either. 0.93 SSIM and it looks worse than dogshit. It SHOULD be like 0.8 SSIM instead.

>WEBP moron learns that image quality metrics are suited for specific tasks. And that encoders can be TUNED TO SWAY SCORES.
Retard.

>>105806047
What the fuck am I supposed to use then? PSNR failed too. How the fuck was I supposed to know that there's a special snowflake quality metric dedicated to fucking cartoons?

>>105806047
Also if webp is cheating in SSIM then why did >>105805981 happen?