fix(lsp-client): pass workspace root as spawn cwd

[nl0]

Summary

Language servers that resolve project-level configuration from the current working directory — e.g. Astral's ty searching for .venv, pyright scanning for pyproject.toml, gopls looking up go.mod — fail to discover the correct environment when lsp-mcp-server is launched from a directory that isn't the workspace root (which is the common case when Claude Code is started from the user's shell cwd). The LSP still receives the right rootUri via initialize, but for tools that read getcwd() directly, that's irrelevant.

Manifests as e.g. spurious Cannot resolve imported module <X> errors pointing at the MCP host's shell venv (or uvx's own tool-env), not the project's.

Fix

Pass this._workspaceRoot as cwd to spawn() in LSPClientImpl.initialize. The value is already known and normalized at that point (assigned one line above from the rootUri parameter).

Reproducer

1. In a shell with VIRTUAL_ENV pointing at some unrelated venv, launch an MCP client that spawns lsp-mcp-server.
2. Configure a Python server entry backed by uvx ty server.
3. Open a Python file whose project has .venv in its root.
4. lsp_diagnostics returns unresolved-import errors for packages that are actually present in the project venv.

With this fix, ty sees the project's .venv via getcwd() and the false positives disappear.

Secondary commit

The second commit adds a prepare script so npm install / npx from a git source (e.g. github:…) builds dist/ automatically. Orthogonal to the fix — happy to split off or drop if you'd rather keep the git-install path separate.

Test plan

• npm run build clean
• npm run typecheck clean
• npm test — all 32 existing unit tests pass
• Manual verification: with the patched fork wired into Claude Code via npx github:nl0/lsp-mcp-server#fix/spawn-cwd-to-workspace-root, ty correctly resolves the project's .venv imports; typescript-language-server continues to work (it was already unaffected because it uses initialize.rootUri).

🤖 Generated with Claude Code

[ramfjord]

Hit the same root cause from Ruby/Sorbet via bundle exec srb tc --lsp: without cwd on spawn(), bundle either can't find the Gemfile or resolves the wrong Ruby version through mise's activate env. Confirmed your one-line fix resolves it across two large Rails monorepos. +1 to merging — would be great to drop our local patch.