Fix Windows test portability: line-ending translation, locale encoding, and /tmp assumptions

[skymanbp]

Summary

Running the test suites on Windows currently fails in three independent ways. All three are test-infrastructure issues — this PR changes no library code (7 files, +53/−10: six test modules plus a new .gitattributes).

After these fixes, the full suites pass on Windows:

• dataframe:test:tests — Cases: 1097 Tried: 1097 Errors: 0 Failures: 0
• dataframe-fastcsv:test:tests — Cases: 61 Tried: 61 Errors: 0 Failures: 0

(local run: Windows 11, GHC 9.12; fork CI: windows-latest on GHC 9.6.7 and 9.12.2, links below)

Root causes and fixes

1. No .gitattributes → git core.autocrlf rewrites CSV fixtures on checkout

GitHub's Windows runners (and many Windows dev machines) default to core.autocrlf=true, so every text-looking file — including CSV test fixtures — is checked out with CRLF line endings. Byte-level tests then see \r that isn't in the committed fixture. This also silently corrupts fixtures with quoted embedded newlines (e.g. the quotes_and_newlines case).

Fix: add .gitattributes with *.csv -text / *.tsv -text. -text means "no translation of what's in the index", so fixtures that intentionally contain CRLF (e.g. crlf.csv) are preserved byte-for-byte as committed.

2. Text-mode Data.Text.IO in tests → newline translation + locale codepage crashes

A default Haskell handle uses the OS locale encoding and native newline translation. On Windows this breaks roundtrip tests twice over:

• \n is written as \r\n, corrupting output meant for the byte-level CSV reader (embedded quoted newlines come back different);
• non-ANSI text can't be encoded under a non-UTF-8 codepage at all — e.g. under a GBK locale, fast_roundtrip_utf8 dies with commitAndReleaseBuffer: cannot encode character '\676'.

Fix: in test writers, Data.Text.IO → Data.Text.IO.Utf8 (same writeFile signature, byte-mode UTF-8, no new dependencies; provided by text >= 2.1). For the streaming handle in prettyPrintSeparated (fastcsv tests), set the handle explicitly: hSetEncoding handle utf8 + hSetNewlineMode handle noNewlineTranslation.

3. Hardcoded /tmp paths (3 sites)

/tmp doesn't exist on Windows — the path resolves relative to the current drive's root, so the tests only pass by accident if C:\tmp happens to exist, and fail with withFile: does not exist otherwise.

Fix: System.Directory.getTemporaryDirectory (sites: tests/Operations/WriteCsv.hs, tests/Operations/Record.hs; the two property-test paths in dataframe-fastcsv/tests/Properties/Csv.hs were moved under ./tests/data/unstable_csv/ alongside the existing pattern there). Non-test /tmp uses (LazyBenchmark default argument, the huggingface URI predicate) were deliberately left untouched.

Verification

Upstream has no Windows CI lane, so the CI evidence below comes from my fork, where a temporary fork-only workflow (windows-latest + macos-14 × GHC 9.6.7/9.12.2) was stacked under these same fixes. That exact state is preserved on the fix/windows-test-portability-with-ci branch; this PR branch is the identical fixes rebased onto current main without the workflow commit.

• Before (fork CI on unmodified test code): Windows jobs red — https://github.com/skymanbp/dataframe/actions/runs/32095552668
• After (same fixes, pre-rebase state):
  • CI (Windows + macOS), all 4 jobs green — https://github.com/skymanbp/dataframe/actions/runs/32160121891
  • Presubmit (hlint + fourmolu) green — https://github.com/skymanbp/dataframe/actions/runs/32160121946
  • CI (ubuntu) green — https://github.com/skymanbp/dataframe/actions/runs/32160121957
• Local full-suite runs on Windows 11 as quoted above.

Note on Haskell-CI: that workflow also fails on this branch, but it is red on upstream main as well (latest three runs on main all conclude failure, most recent 2026-08-14), so it appears unrelated to this change.

[mchav]

Thanks. This is a great change! Reminds me that I need to split my dev time between unix and Windows.