Age.TestKit: 61/143 CCTV vectors fail on Windows due to Git autocrlf line-ending normalization

[petre-c]

Summary

Age.TestKit/CctvTestRunner fails 61 of 143 wire-format reference vectors on Windows. The same vectors pass cleanly on Linux (143/143). The failures are not crypto bugs — they're caused by Git's default core.autocrlf=true on Windows checkouts normalizing the LF-only testdata .txt files to CRLF, which corrupts byte-exact reference vectors.

Repro

On a fresh Windows checkout of pscheid92/AgeSharp@main (no patches):

git clone https://github.com/pscheid92/AgeSharp
cd AgeSharp
dotnet test Age.TestKit/Age.TestKit.csproj -c Release --nologo

Result:

Failed!  - Failed: 61, Passed: 82, Skipped: 0, Total: 143

On a Linux runner (e.g. GitHub Actions ubuntu-latest):

Passed!  - Failed: 0, Passed: 143, Skipped: 0, Total: 143

Root cause

testdata/*.txt files are byte-exact age v1 reference vectors. The CCTV runner reads them and compares against expected hashes / decryption outputs. On Windows, Git's default core.autocrlf=true converts LF to CRLF on checkout, breaking the byte-exactness:

$ file Age.TestKit/testdata/armor_empty.txt
Age.TestKit/testdata/armor_empty.txt: ASCII text, with CRLF line terminators

Hex dump confirms 0d 0a after every line — should be 0a only:

00000000: 6578 7065 6374 3a20 6865 6164 6572 2066  expect: header f
00000010: 6169 6c75 7265 0d0a 6669 6c65 206b 6579  ailure..file key

Failure distribution

The 61 failures span every test category — not just armored ones — because the metadata parser is also affected:

Test prefix	Failing
armor_*	29
stanza_*	9
x25519_*	8
stream_*	7
hybrid_*	5
scrypt_*	1
hmac_*	1
empty	1
Total	61

Suggested fix

Add a .gitattributes file that marks the testdata as binary so Git leaves it alone on every platform:

Age.TestKit/testdata/** binary
Age.TestKit/testdata/**/*.txt binary

Or scope it tighter to just the .txt files in that directory.

After the .gitattributes lands, existing Windows clones need to re-fetch the files to get the LF version:

git rm --cached -r Age.TestKit/testdata
git checkout HEAD -- Age.TestKit/testdata

Environment

• AgeSharp main (unpatched, commit at time of repro: latest)
• .NET SDK 10.0.203 on Windows 11
• Git core.autocrlf=true (Windows default)

Why I noticed

This came up while validating an unrelated patch (#2 — WASM cipher swap). I initially attributed the 61 failures to pre-existing upstream bugs (because they showed up on both patched and unpatched checkouts on Windows). Re-running the same code on Linux CI revealed 143/143 pass — at which point the platform pattern was obvious.

Happy to send a PR with the .gitattributes fix if you'd like.

[davidmatson]

I hit this problem as well and worked around via .gitattributes here:
https://github.com/davidmatson/AgeSharpNetStandard

But I think the better fix would be to use a different file extension for these test files - age files are generally binary (unless armored), so the .txt extension might be better replaced by .age, .agetest, .bin, or similar.