I repeated the same download procedure that I used originally and got different data.

$ for i in md5 sha1; do openssl dgst -$i CSelTest.pm CSelTest.pm-20160
+902a_ORIGINAL_CODE; done
MD5(CSelTest.pm)= 2a721b1bdc7e0ba01a9620c2cf61b171
MD5(CSelTest.pm-20160902a_ORIGINAL_CODE)= 2fc9805a057076e10117e0fc710f
+6321
SHA1(CSelTest.pm)= 092f86a1c1c03523c2bb9f459ef353a757dfa8b6
SHA1(CSelTest.pm-20160902a_ORIGINAL_CODE)= 49bc50c3b30e817be950a9bcda9
+973baa5b85309
[download]

The new code was four bytes shorter ...

$ ls -l CSelTest.pm CSelTest.pm-20160902a_ORIGINAL_CODE
-rw-r--r--  1 ken  staff  3224  3 Sep 00:16 CSelTest.pm
-rw-r--r--  1 ken  staff  3228  2 Sep 19:14 CSelTest.pm-20160902a_ORIG
+INAL_CODE
[download]

... due to replacing "et ai", wherever that came from, with presumably the missing closing parenthesis:

$ diff CSelTest.pm CSelTest.pm-20160902a_ORIGINAL_CODE
43c43
<                   )?
---
>                   et ai?
[download]

Anyway, I ran your six tests and got the same results except for the second one: correctly output '"="', instead of '"eq"'.

$ perl -MCSelTest -MData::Dump -E 'dd CSelTest::parse_csel(q{ [attr=1]
+ })'
[[{ name => "attr" }], "=", 1]
[download]

Update: Added " [Part 1 of 2]" to the title to differentiate this node, "Re^3: Weirdness (duplicated data) while building result during parsing using regex", from another of the same name, "Re^3: Weirdness (duplicated data) while building result during parsing using regex" (which will have " [Part 2 of 2]" appended).