fix(pm): windows clone error

[elrrrrrrr]

Summary

This PR started as a small "preserve clone error chain" fix and grew, in a good way: the new e2e coverage immediately surfaced a real Windows concurrency bug in CLONE_CACHE, which is also fixed here.

Three commits:

1. fix(pm): preserve clone error chain instead of swallowing it

   • clone_package_once() now returns Result<()> instead of bool, so the full error chain reaches the caller via ? instead of being collapsed into a generic "{name} clone failed" bail
   • The warn log uses {:#} (alternate Display) to print the complete anyhow chain, not just the top-level message
   • install.rs and the pipeline pre-clone worker propagate / log the real error

2. test(pm): add eggjs/egg next branch smoke to Windows e2e

   • Mirrors the existing Linux/macOS bash case into e2e/utoo-pm.ps1 so Windows CI also exercises ut install --from pnpm against a complex pnpm-workspace monorepo (@langchain/core with deeply nested node_modules)
   • Added because Windows has been intermittently flaky and this surface was uncovered

3. fix(pm): normalize path separators in CLONE_CACHE keys on Windows

   • Surfaced by commit 2: the new e2e job failed on Windows with ERROR_SHARING_VIOLATION (os error 32) cloning camelcase and uuid into @langchain/core/node_modules/
   • Root cause: OnceMap dedupes by target path string, but install.rs constructs targets with forward slashes (lockfile-derived) while pipeline workers go through Path::join, which injects \ on Windows. node_modules/foo/bar and node_modules/foo\bar are distinct strings → dedup fails → both clone tasks race on the same real directory file-by-file → second writer hits a sharing violation. wait_clone_if_pending is also defeated for the same reason.
   • Fix: switch the OnceMap key from String to PathBuf, run target paths through Path::components().collect() on Windows. components() parses both / and \ as separators and collect::<PathBuf>() rebuilds with the OS-preferred separator — gives a stable key without any extra crate or replace() hack. Unix paths skip the normalization and use to_path_buf() directly.
   • Includes a Windows-only regression test asserting cache_key("a/b") == cache_key("a\\b").

Without commit 1, the diagnosis above wouldn't have been possible

The whole investigation was unblocked by the error chain fix: before commit 1 the failure would have shown up as "camelcase clone failed" in the install bail, with no os error 32 to grep for. Commit 1's tracing::warn! with {:#} plus bail! propagation made the exact NTFS error visible in CI logs the first time the new e2e job ran.

Test plan

• cargo build -p utoo-pm clean
• cargo clippy -p utoo-pm --all-targets -- -D warnings --no-deps clean
• cargo test -p utoo-pm cloner:: (16 pass)
• CI passing: linux/macos/windows utoopm-ci + utoopm-e2e (incl. new eggjs/egg smoke on Windows)
• CI rerun confirms Windows stability (no flaky)

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request refactors the auto-update mechanism to include a 24-hour cooldown after failed attempts and improves dependency cloning by propagating errors instead of returning booleans. Key changes include updating Cargo.lock with new dependencies, simplifying global state in the update helper, and enhancing user feedback during the update process. Feedback was provided regarding the loss of error context in the package cloning logic and the lack of error handling when saving the version cache.