[v2] Test suite fails on Windows: 3 independent causes (CRLF checkout, symlink EPERM, POSIX signals) #3117

What happened?

Running pnpm -r test on a fresh clone of main (39ae8e6) on Windows gives 5 failures in 3 packages. Build passes and everything else is green. CI only runs on ubuntu-latest, so this doesn't show up there.

Each failure has a different cause:

1. core-internal — schema twin byte-identity (2 tests)
test/wire/schemaTwinConformance.test.ts

AssertionError: byte size drifted for 2025-11-25 — the vendored twin was rewritten: expected 178384 to be 174326

The repo has no .gitattributes, and Git for Windows defaults to core.autocrlf=true, so the twins are checked out with CRLF. 2025-11-25.schema.json has 4058 lines and is exactly 4058 bytes larger than the committed file. Same for 2026-07-28 (185437 vs 181474).

Screenshot of the core-internal run (both twins fail):

2. codemod — symlink cycle test (1 test)
test/v1-to-v2/packageJsonUpdater.test.ts > "survives a directory symlink cycle without following it"

Error: EPERM: operation not permitted, symlink '...\codemod-manifests-DQY0E9' -> '...\codemod-manifests-DQY0E9\src\loop'

Creating a directory symlink on Windows needs admin rights or Developer Mode, so the test fails during setup, before the codemod runs.

3. test/e2e — transport:stdio:shutdown-escalation (2 tests)
scenarios/stdio.test.ts:208

AssertionError: expected '[stdio-server] ready\n' to contain '[stdio-server] sigterm ignored'

On Windows there are no POSIX signals: per the Node docs, subprocess.kill('SIGTERM') means the process "will always be killed forcefully and abruptly (similar to 'SIGKILL')". So the fixture can never ignore SIGTERM. The transport itself works (the child does die); it's the test's premise that doesn't hold on Windows.

Possible fixes, happy to send them as PRs if this sounds right:

1. Add a .gitattributes that keeps the twins as LF (e.g. packages/core-internal/test/corpus/schema-twins/* -text). Or a repo-wide * text=auto eol=lf if you prefer.
2. Use 'junction' instead of 'dir' on win32 in that test. Junctions don't need privileges and still form the cycle, so the test keeps covering Windows.
3. Skip the shutdown-escalation scenario on win32, or on Windows only assert that the child is dead after close().

Would you prefer one PR or three separate ones? Separately, would adding a windows-latest job to the test matrix be welcome?

What did you expect?

The suite to pass on Windows, or the Windows-only cases to be skipped explicitly.

Code to reproduce

git clone https://github.com/modelcontextprotocol/typescript-sdk.git
cd typescript-sdk
pnpm install --frozen-lockfile
pnpm build:all
pnpm --filter @modelcontextprotocol/core-internal test
pnpm --filter @modelcontextprotocol/codemod test
pnpm --filter @modelcontextprotocol/test-e2e test

SDK version

main @ 39ae8e6 (2026-10-27) — Windows 11 Pro 26100, Node 24.11.1, pnpm 10.26.1, git 2.52.0.windows.1 (core.autocrlf=true)

Area

Other