The normalization happens long before API code is reached. All of the Unicode text MediaWiki stores is required to be in normal form C (NFC). The two titles used in your script:

$name1 = 'Test2%E2%80%93_%E0%A6%85%E0%A6%B8%E0%A6%AE%E0%A7%80%E0%A6%AF%E0%A6%BC%E0%A6%BE_%282%29';
$name2 = 'Test2%E2%80%93_%E0%A6%85%E0%A6%B8%E0%A6%AE%E0%A7%80%E0%A7%9F%E0%A6%BE_%282%29';

Are actually the same title in different forms. Try putting the decoded forms in http://www.fontspace.com/unicode/analyzer/. (I think Phabricator also normalizes Unicode on input, so pasting decoded versions here is pointless.)

The first uses:

• য U+09AF Bengali Letter Ya
• ় U+09BC Bengali Sign Nukta

The second uses:

• য় U+09DF Bengali Letter Yya

The first is in NFC, the second is not in any normal form. According to http://www.scarfboy.com/coding/unicode-tool?s=U%2B09DF, U+09DF becomes U+09AF + U+09BC when converted to NFC form.

You should always ensure all of the input to your code is in NFC. Usually the WebRequest class will take care of this for you, but if you bypass it (evil you) or take the data from somewhere else (like file names from filesystem), you need to convert it yourself. UtfNormal\Validator::cleanUp() seems to be the method to do it.