Idle language servers stay running while their project root is in use: idle reap is per root, not per server (31.5 GiB of tsserver in swap) #1170

Problem

AFT stops language servers only when a whole project root is idle. A root that has continuous use keeps every server that it started, also servers for package directories that nobody has used for hours. TypeScript servers use approximately 1 GiB each, so the memory increases with each new package directory, and the operating system moves them to swap.

What I measured (aft 0.62.0, one module process, Linux)

• 34 typescript-language-server processes (68 tsserver.js processes), all children of the aft process. Together they used 31.5 GiB of swap and 7 GiB of resident memory. The aft process itself used 0.9 GiB resident and 2.5 GiB swap.
• Each server has a different package directory (the nearest tsconfig.json). Most of the directories are git worktrees in a gitignored directory below the project root. Agents in long sessions in that root create a new worktree for each task.
• Server ages were 0.2 h to 10.7 h. For one server, the last request that AFT sent to it was 8 hours before the measurement.
• In 12.5 h, the module log has one idle lsp reap line (a different root). lsp.idle_minutes is not set, so the default (60) applies.
• I stopped one server with SIGTERM. AFT logged exited TypeScript … (eof) and started a new server on the next request. The diagnostics were correct. Thus a stopped server has no bad effect, but AFT does not stop idle servers itself.

Cause (v0.62.0)

reap_idle_lsp_servers (crates/aft/src/subc/mod.rs:2607) calls shutdown_idle_lsp_at with the root's last_touched. shutdown_idle_lsp_at (crates/aft/src/runtime_drain.rs:4051) compares the idle time of the root with the TTL. Then it stops all clients with take_all_clients(), or it stops none. last_touched changes on each tool call for the root. So while one session in the root makes calls, no server of that root stops. This is also true for a server that got no request for hours. There is no limit on the number of servers for each root.

The standalone path (shutdown_idle_lsp, which uses last_request_at()) has the same all-or-nothing behavior.

Suggested change

• Record the last use for each server (a request sent to it, or a document open, change or save for a file that it serves).
• In the sweep, stop each server whose last use is older than lsp.idle_minutes, also when the root is in use. Use the current background stop (spawn_idle_lsp_reap).
• Possibly: a limit on the number of servers for each root, which stops the server with the oldest last use when a new server starts.

Other observation

A memory check that reads only the aft process (resident memory and swap) does not include the language server children. So that check does not see this increase.