Fixes the pnpm linker crash on leftover workspace node_modules

[jeckhart]

Switching a monorepo from nodeLinker: node-modules to pnpm aborts:

➤ · I/O error (Operation not permitted (os error 1))

Repro (yarn 6.0.0-rc.18, bec529a):

# root package.json:   workspaces ["packages/*"], deps { no-deps: 1.0.0 }
# packages/pkg-a:       deps { no-deps: 2.0.0 }
yarn config set nodeLinker node-modules && yarn install
yarn config set nodeLinker pnpm && yarn install

The version conflict leaves no-deps@2.0.0 as a real directory at packages/pkg-a/node_modules/no-deps. The pnpm linker's symlink pass calls fs_rm_file (unlink) on that path, which fails on a directory — EPERM on macOS, EISDIR on Linux — and ok_missing only swallows NotFound. fs_remove_nm clears the root node_modules but never the per-workspace ones.

Fix: use the directory-aware fs_rm at the three symlink sites in link_project_pnpm. Adds a regression test that fails on current main and passes with the fix.

Closes #304

[netlify]

👷 Deploy request for yarn-v6 pending review.

Visit the deploys page to approve it

Name	Link
🔨 Latest commit	d267886

[jeckhart]

@arcanis - Sorry for the ping, I wasn't sure who to address this to from the project docs. I am using yarn 6 preview for some non-critical projects at work and I've found a few bugs. This was one that I thought I could help fix myself.

[arcanis]

Hey John, apologies for the delay, I'll take a look tomorrow!

The main thing I want to check is whether we are truly replacing all previous node_modules, or just those that would happen to conflict with directory entries created by the pnpm linker (thus incorrectly leaving in place hoisted packages).

I somewhat suspect the pnpm linker should use the same fs_sync method as the NM linker rather than a direct fs_rm, to avoid that.