Test 2 was built by copying the raw HTML &xxx markup from the result of test 1. It didn’t suffer from the snowball effect that others see.
Something in the software is not understanding UTF-8.
Test 3: &#E996AA;
The server is mangling UTF-8 sequences for chars outside the 7-bit ASCII range. These consist of multiple bytes in the 80-FF range. The server looks at each byte of the multi-byte UTF-8 sequence and substitutes an HTML entity. The browser then renders the result as several characters of garbage instead of the character intended.
E.g., E2 80 9C is UTF-8 for the left curly double quote. The server substitutes an entity for circumflex lower case a for the E2, the euro sign for 80, and the trademark symbol for 9C.