feat: native Windows support

[jhodges10]

Closes #35.

I have a developer on Windows, so I took you up on the offer to send a PR. drift already compiled on Windows — what it didn't do was produce a drift.lock that meant the same thing anywhere else. On a fresh Windows checkout of this repo, drift check reported every doc stale.

What was broken

Path separators. Repo-relative paths come out of std.Io.Dir.path.relative with host separators, so drift link docs/a.md src/b.ts recorded doc = "docs\\a.md". Doc discovery matches bindings against git ls-files, which is POSIX on every platform, so a binding never matched the doc it was written for.

Line endings. Git for Windows enables core.autocrlf by default, so the working tree holds CRLF where Linux holds LF. Every fingerprint that reaches raw bytes moved with the checkout rather than the content — the no-grammar fallback and markdown sections directly, and grammar-based fingerprints too, since a line comment's token text runs to end-of-line and swallows the \r.

Together these accounted for all 26 integration-test failures on Windows.

A relink-gate use-after-free. Chasing the above turned up a real bug that isn't Windows-specific. Binding.setField frees the previous value, so the old_sig slice the gate holds across refreshBindingSig dangles, and isDocGateBlocked compares against freed memory. Whether a relink is refused or waved through depends on what the allocator left behind. On Windows it meant drift link <doc> refused every anchor in a doc, including ones whose fingerprint had not moved. It's the second commit and stands alone if you'd rather take it separately.

What's here

• src/repo_path.zig — normalizes a repo-relative path to / where it is produced. Absolute paths keep host form; they never leave the process.
• src/content.zig — collapses CRLF as working-tree files are read. LF-only content hashes unchanged, so existing lockfiles stay valid — no relink needed for anyone on Linux or macOS.
• .gitattributes pinning the working tree to LF. drift no longer needs it, but zig fmt rejects CRLF, so without it a Windows contributor can't format the repo. Content is already LF in the index, so git add --renormalize . is a no-op here.
• A test-windows CI job on windows-latest running the same build, test suite, and drift lint as the Linux job. Cross-compiling Windows without ever running it is what let this sit.
• x86_64-windows and aarch64-windows in the CI and release matrices, cross-compiled from Linux like the other targets, packaged as .zip since Windows opens those without extra tooling.
• Regression tests that fail without the fixes: a CRLF-vs-LF freshness test (covering a grammar-backed file, a no-grammar file, and a markdown heading) and a lockfile-separator test. Both run on every platform.
• Docs: Decision 15, a DESIGN section on cross-platform identity, a README install section for Windows, RELEASING and CLI updates.

Verification

On Windows 11, Zig 0.16.0:

Build Summary: 11/11 steps succeeded; 81/81 tests passed
drift check                      → exit 0
zig build gen-check-schema       → schema matches what's committed

Both Windows targets cross-compile clean. Test count went 79 → 81 (the 26 Windows failures are gone; the added tests are net-new).

I have not touched the install.sh at drift.fp.dev, so the README points Windows users at the release zip and at building from source. Happy to follow up there if you want it to handle Windows too.

🤖 Generated with Claude Code

[jhodges10]

@laulauland I missed your initial response to my github issue, but I just went ahead and did the Windows build!