fix: retry worktree removal on Windows when transient locks delay deletion

[jeremylasne92-hue]

Fixes #24

Problem

On Windows (Git Bash / MSYS), cmux rm fails with Directory not empty when the worktree contains a populated node_modules/. Background processes (npm watchers, esbuild, vitest) hold short-lived file handles inside node_modules after exit; git worktree remove runs its cleanup while a handle is still open → fails. The lock clears within ~1 second but cmux gives up immediately.

Fix

Retry git worktree remove up to 3 times with a 1-second backoff when running on MSYS/MinGW/Cygwin (uname -s matches MINGW*|MSYS*|CYGWIN*). Linux and macOS behavior is unchanged — they fall through to the original failure path with the existing error message.

Verification

On Windows Git Bash (MSYS2/MINGW64, bash 5.2.37, git 2.52.0.windows.1):

$ cd /c/path/to/node-repo
$ cmux new test-rm
# ...worktree created with node_modules/
$ cmux rm test-rm -f
# Before fix: error: failed to delete '...': Directory not empty
#             Failed to remove worktree: test-rm
# After fix:  Removed worktree and branch: test-rm   (succeeds on retry)

On Linux (no cygpath/MSYS detected): the case block is skipped entirely; the function behaves exactly as before. Existing error message and --force hint still trigger when the first attempt fails.

Scope

• One function modified: _cmux_rm
• 27 lines added, 1 removed
• No new dependencies
• No test additions (repo has no test suite per CLAUDE.md)

[jeremylasne92-hue]

Field-tested on a real DANTE IA worktree (Node monorepo with vitest+esbuild). The initial 3x sleep 1 budget (3s total) was insufficient in practice — locks on node_modules native binaries persisted 5-8 seconds after the npm processes exited. Pushed an amendment (a48b817):

• Retry budget bumped to 5x sleep 2 (10s total)
• Last-resort fallback: rm -rf $worktree_dir && git worktree prune when git still refuses to delete a directory it cannot fully empty

This now succeeds reliably end-to-end on Windows for a 1.8 GB node_modules worktree. Linux/macOS branch is still skipped via the uname -s case.

[siennathesane]

I'm also onboard with this one, too!

I have an idea. If you can submit a new PR with a GitHub Actions runner for Windows that will at least bootstrap and run whatever tests are available, then I'm happy to merge #23 and #24 for you 🫶🏼 I just want to make sure things are properly supported.