Thanks for the contribution. I love the idea, but there are quite some change.
I asked both Claude and Gemini to review the PR. Below are what they found. Feel free to discard if the comments are not correct or not importment.

Overall Review
This is a fantastic PR! The architecture for deferring the language server start and introducing the 30-second idle stop is very well thought out. The deduplication logic for rapid starts and the smooth handling of single vs. multi-server modes are great improvements that will definitely help with performance and memory usage.

I did notice a few issues during review—one critical bug regarding Quarto documents and two smaller edge cases—that should be addressed before merging:

1. The Multi-Server Quarto Bug (Critical)

In closeMultiClient, there is a manual cleanup block that triggers when a Quarto document is closed. It loops through all servers, and if a server's remaining tracked documents are all Quarto chunks (.vdoc.r), it wipes the tracking set (this.trackedDocuments.delete(serverKey)) and schedules a stop.

This introduces a cross-document bug:

• If a user has two Quarto files (A.qmd and B.qmd) open in the same workspace, closing A.qmd evaluates to true here (because B.vdoc.r is a Quarto chunk). The server tracking is wiped and the server is killed, unexpectedly breaking LSP features for the still-open B.qmd.
• Furthermore, because isUntitledQuartoDocument returns true for any untitled r or rmd document, simply closing an unrelated Untitled-1.R script will trigger this dangerous Quarto cleanup loop.

Suggested Fix:
The safest and simplest fix is to remove this manual Quarto cleanup block entirely. The Quarto extension already disposes of its virtual documents when the parent .qmd is closed, which automatically triggers onDidCloseTextDocument for the .vdoc.r files, naturally untracking them.

2. Dropped scheme Filter in Single-Server Mode

In the old code, the single-server mode explicitly restricted document selectors to scheme: 'file', scheme: 'untitled', and scheme: 'vscode-notebook-cell'.

The new singleServerDocumentSelector() simplifies this to just { language: 'r' } and { language: 'rmd' }.
Because the scheme filter was dropped, the R language server will now attempt to spin up for any URI scheme—including git:// URIs. This means users might experience the heavy R server spinning up just from clicking to view a Git diff in the source control tab.

Suggested Fix:
Restore the scheme restrictions in singleServerDocumentSelector() to prevent unintended server starts.

3. Missing Client Disposal on Crash (Code Hygiene)

In spawnServer, the child process on('exit') callback now maps to handleClientExit. handleClientExit removes the client from this.clients, but never calls dispose() on it.

Because createClient pushes every new client into extensionContext.subscriptions, these crashed/stale clients will accumulate in the subscriptions array unboundedly until the extension is deactivated.

Suggested Fix:
To prevent accumulating stale event listeners, dispose of the client when it crashes by adding a cleanup call inside handleClientExit:

private handleClientExit(serverKey: string, client: LanguageClient): void {
    if (this.clients.get(serverKey) !== client) { return; }
    this.clients.delete(serverKey);
    this.initSet.delete(serverKey);
    this.cancelIdleStop(serverKey);
    void this.stopStartedClient(client); // Ensures the crashed client is disposed
}

Changes have been made after reviewing the comments.

• For 2. The scheme filter has been restored.

• For 3. As I understand vscode-languageclient should already clean up after a crash, this comment was discarded. But the new change prevents stopped clients from being retained.

• For 1. The bug in dealing with Quarto should be fixed, the correct quarto flow would be

  Single-server mode:

  • A Quarto doc creates .vdoc.r only when R chunks require language server
  • The .vdoc.r starts the single R server.
  • Its parent qmd is tracked as the lifecycle owner.
  • All R, Rmd, and quarto docs share that server.
  • It stops 30 seconds after the last related doc closes.

  Multi-server mode:

  • Each .vdoc.r is mapped to its parent qmd.
  • The parent qmd determines the workspace/server key—not the temp directory.
  • qmds in the same workspace share one server.
  • qmds in different workspaces use separate servers.
  • .vdoc.r requests are routed only to their workspace server
  • Closing one qmd does not affect other open qmds.
  • A workspace server stops 30 seconds after its last R/Rmd/qmd doc closes.

  Untitled Quarto is now handled separately from untitled R/Rmd. In multi-server mode, untitled docs and R chunks from untitled qmd share one untitled server.