Tests `std::mem::forget` tempdirs — leaks per test run

[lovebear2008]

In crates/forkd-controller/src/http.rs:1480-1494, test_state does:

let td = tempfile::TempDir::new().unwrap();
let path = td.path().join("state.json");
let snapshot_root = td.path().join("snapshots");
// Leak the TempDir so it survives the test (Drop deletes the dir).
std::mem::forget(td);

The same pattern is repeated in branch_slot_global_cap_blocks (1630-1641) and branch_slot_capacity_recovers_on_drop (1652-1662).

The comment explains the motivation ("survive the test") but the result is that every cargo test run leaves N orphan directories in $TMPDIR per worker. tempfile::TempDir's default location is /tmp on Linux, which is usually tmpfs — so the leak is bounded by reboot, but on CI the cargo test cache can accumulate hundreds of these between runs.

Better shapes

• Bind the TempDir to a 'static location via tempfile::tempdir_in() plus a one-shot OnceCell, so all tests in the file share one cleanup root that's reaped on process exit.
• Or move the leak to a Drop impl on a wrapper that calls remove_dir_all on the leaked path. axum-test patterns usually use this approach.
• Cheapest: keep the TempDir alive by binding it to the returned SharedState (add a field). Drop of AppState then triggers Drop of the TempDir, no leak.

This came up while running the test suite locally — ls /tmp/.tmp* | wc -l jumped by ~10 per cargo test -p forkd-controller.

[WaylandYang]

Thanks for the detailed writeup — the three sites and the suggested shapes made this easy to pick up.

Went with your third option (bind the TempDir to the returned state), since it matches the _td: TempDir pattern tests/http_integration.rs already uses. Fix is in #274.

One data point worth recording: on the current test count the leak is 56 directories per cargo test -p forkd-controller run, not ~10 — it grew with the suite. After the change it is 0, same 71 + 7 tests passing.