fix: include decrypted sequence comments in the MAC (fixes #2243)

[cbcoutinho]

What

Fixes a regression in 3.13.2 where sops fails to decrypt any YAML/JSON file
containing a comment inside a sequence, reporting MAC mismatch. This
affects files written by ≤3.13.1 and files 3.13.2 encrypts itself.

Fixes #2243.

Root cause

A comment in a sequence is encrypted as an ENC[...,type:comment] element and
restored as a sops.Comment node when the file is loaded again.

#2227 changed the MAC-inclusion test in Tree.Decrypt from inspecting the
decrypted result to inspecting the input node:

-if _, ok := v.(Comment); !ok {
+if !isComment {
     bytes, err := ToBytes(v)
     hash.Write(bytes)
 }

A comment that decrypts successfully becomes a non-Comment value, so the old
check included it in the MAC. Tree.Encrypt still folds the comment in (at
encrypt time it is a non-Comment node), so the encrypt-side and decrypt-side
MACs disagree for any sequence containing a comment.

Fix

Restore the pre-#2227 behaviour (test the decrypted value's type) while keeping
the non-string panic guard #2227 introduced. One-line semantic change.

Reproduction

# repro.yaml
items:
  # - commented@example.com
  - real@example.com

$ sops -v3.13.1 -e repro.yaml > repro.enc.yaml
$ sops -v3.13.1 -d repro.enc.yaml   # OK
$ sops -v3.13.2 -d repro.enc.yaml   # MAC mismatch

A comment on a mapping, or a sequence with no comment, is unaffected.

Test

Adds TestMACWithCommentInSequence (YAML store round-trip: load → encrypt →
emit → reload → decrypt, asserting the MACs match). It fails on main and
passes with this change. go test ./... for the root, stores, decrypt and cmd
packages passes; gofmt/go vet clean.

---

This PR was generated with the help of AI, and reviewed by a Human

[felixfontein]

Thanks for reporting this and creating a fix! There are a few issues with the test and the comment for the fix, though.

[cbcoutinho]

Thanks for the review @felixfontein! Pushed 1aa8912 addressing all four points:

• sops.go comment — replaced with your suggested wording. You're right: the plaintext comment in a sequence is a Comment node, but Encrypt turns it into an ordinary (non-Comment) string element whose metadata marks it a comment; on Decrypt that string decrypts back to a Comment. Hence the MAC decision has to look at the decrypted result's type, not isComment on the input.
• Test location — moved out of the YAML store and into sops_test.go (package sops).
• Build tree from SOPS objects — the test now constructs branches directly with a Comment inside a slice, no store involved.
• No serialize/deserialize — it encrypts and then decrypts the same tree (Encrypt mutates in place), and additionally asserts the sequence comment became a non-Comment string after encryption, so the format is checked explicitly. A small commentCipher in the test models the AES cipher's comment-metadata round-trip without pulling in the aes package (which would be an import cycle in package sops).

Verified the test fails on the pre-fix code and passes with the fix.

[felixfontein]

@cbcoutinho thanks for fixing this!