test(desktop): resolve P4 pass-3 test-vs-claim defects in workspace-scoped agent store

Writer test: change writer to block directly on managed_agents_store_lock
(not via start_pair_lazy_for_with_hook) so the test fails deterministically
if the store guard is dropped before the step-7 save. The hook mutates
records in-memory; the step-7 save writes the mutation to disk while the
lock is still held; the writer loads after the lock is released.

Contender test: add try_recv() check on a dedicated channel inside
on_records_loaded alongside the existing atomic bool. on_transition_acquired
sends to both channels. try_recv() is deterministic: without the transition
lock, on_transition_acquired fires in nanoseconds; by the time on_records_loaded
runs (after file I/O), the channel contains the message. Documents the
time-differential argument explicitly in the test comment.

Full-tail PID: capture child.id() before wrapping into ManagedAgentProcess;
assert exact equality with receipt.pid (assert_eq! not assert! pid > 0).

SCOPE_GENERATION_TEST_LOCK: add to all remaining tests that mutate or
capture scope generation: app_state_scope_tests.rs (4 tests),
global_agent_config_tests.rs (4 tests), runtime_commands_tests.rs (2 tests),
scope.rs (5 generation tests). All tests that advance SCOPE_GENERATION
now participate in the process-global serialization mutex.

Trim runtime_commands.rs doc comments to stay at 1000 lines (gate limit).

Co-authored-by: Will Pfleger <pfleger.will@gmail.com>
Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

desktop/src-tauri/src/app_state_scope_tests.rs

@@ -11,6 +11,9 @@ use super::*;
 /// only written inside `apply_workspace`'s prepare stage.
 #[test]
 fn test_import_before_first_apply_leaves_scope_none() {
+    let _gen_guard = crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
     let state = build_app_state();
 
     // Boot state: no active scope.
@@ -45,6 +48,9 @@ fn test_import_before_first_apply_leaves_scope_none() {
 /// commands fail closed until the frontend re-applies a workspace.
 #[test]
 fn test_live_import_with_active_scope_clears_scope_and_bumps_generation() {
+    let _gen_guard = crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
     let state = build_app_state();
     let base = std::env::temp_dir();
 
@@ -129,6 +135,9 @@ fn test_fallback_relay_never_claims_during_identity_import() {
 /// original scope.
 #[test]
 fn test_prepare_failure_leaves_old_scope_intact() {
+    let _gen_guard = crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
     let state = build_app_state();
     let base = std::env::temp_dir();
 
@@ -169,6 +178,9 @@ fn test_prepare_failure_leaves_old_scope_intact() {
 /// returning `None` is handled gracefully by callers that check the scope.
 #[test]
 fn test_inactive_runtime_exit_after_scope_cleared_is_safe() {
+    let _gen_guard = crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
     let state = build_app_state();
     let base = std::env::temp_dir();
 

desktop/src-tauri/src/commands/global_agent_config_epoch_tests.rs

@@ -235,6 +235,10 @@ async fn test_full_tail_stop_spawn_receipt_register_save() {
     let receipt_pid = Arc::new(Mutex::new(0u32));
     let receipt_instance_id = Arc::new(Mutex::new(String::new()));
     let receipt_started_at = Arc::new(Mutex::new(String::new()));
+    // Capture the PID of the spawned child so we can assert exact equality with
+    // receipt.pid — proves the receipt carries the real child's PID, not an
+    // arbitrary positive value.
+    let spawned_child_pid = Arc::new(Mutex::new(0u32));
 
     let stop_called2 = stop_called.clone();
     let spawn_relay2 = spawn_relay.clone();
@@ -248,6 +252,7 @@ async fn test_full_tail_stop_spawn_receipt_register_save() {
     let receipt_pid2 = receipt_pid.clone();
     let receipt_iid2 = receipt_instance_id.clone();
     let receipt_sat2 = receipt_started_at.clone();
+    let spawned_pid2 = spawned_child_pid.clone();
     let pubkey2 = pubkey.clone();
 
     let app_handle_for_assert = app_handle.clone();
@@ -268,14 +273,19 @@ async fn test_full_tail_stop_spawn_receipt_register_save() {
             },
             // spawn_fn: record captured relay+owner+personas+teams+global, return a noop child.
             // `spawn_noop_child_for_test()` replaces `/usr/bin/true` — cross-platform.
+            // Capture the child PID before wrapping so we can assert exact equality
+            // with receipt.pid — fails if production uses any PID other than the child's.
             move |_app, rec, relay, owner, personas, global, teams| {
                 *spawn_relay2.lock().unwrap() = Some(relay.to_string());
                 *spawn_owner2.lock().unwrap() = owner.map(str::to_string);
                 *spawn_personas2.lock().unwrap() = !personas.is_empty();
                 *spawn_teams2.lock().unwrap() = !teams.is_empty();
                 *spawn_global2.lock().unwrap() = !global.env_vars.is_empty();
+                let child = spawn_noop_child_for_test();
+                // Capture the real child PID before moving child into ManagedAgentProcess.
+                *spawned_pid2.lock().unwrap() = child.id();
                 Ok(crate::managed_agents::ManagedAgentProcess {
-                    child: spawn_noop_child_for_test(),
+                    child,
                     log_path: std::path::PathBuf::new(),
                     spawn_config:
                         crate::managed_agents::spawn_snapshot::prospective_spawn_config_snapshot(
@@ -359,9 +369,15 @@ async fn test_full_tail_stop_spawn_receipt_register_save() {
         "receipt.key.relay_url must match the captured relay; fails if the spawn path \
          builds the receipt with an empty or wrong relay"
     );
-    assert!(
-        *receipt_pid.lock().unwrap() > 0,
-        "receipt.pid must be non-zero (a real spawned child PID)"
+    // Assert receipt.pid exactly matches the PID captured from the spawned child.
+    // Fails if production writes any PID other than `child.id()` into the receipt.
+    let expected_pid = *spawned_child_pid.lock().unwrap();
+    assert!(expected_pid > 0, "spawned child PID must be non-zero");
+    assert_eq!(
+        *receipt_pid.lock().unwrap(),
+        expected_pid,
+        "receipt.pid must equal the spawned child's PID (child.id()); \
+         fails if production assigns an arbitrary PID to the receipt"
     );
     // Assert receipt.desktop_instance_id equals the actual value produced by
     // current_instance_id(app) — proves the field is sourced from the app

desktop/src-tauri/src/commands/global_agent_config_tests.rs

@@ -186,6 +186,9 @@ fn test_restart_under_captured_epoch_fresh_scope_no_runtime_is_skipped() {
 /// guard must abort before stopping — no agent is touched.
 #[test]
 fn test_restart_under_captured_epoch_stale_scope_is_rejected() {
+    let _gen_guard = crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
     let tmp = tempfile::tempdir().unwrap();
     std::fs::write(tmp.path().join("managed-agents.json"), b"[]").unwrap();
 
@@ -316,9 +319,15 @@ fn make_mock_app() -> tauri::App<tauri::test::MockRuntime> {
 /// Thufir's test 1: "production async driver with injected loader failure;
 /// assert stop never called, RestartOutcome::Skipped."
 #[tokio::test]
+#[allow(clippy::await_holding_lock)] // SCOPE_GENERATION_TEST_LOCK serialises parallel tests
 async fn test_context_load_failure_leaves_runtime_running() {
     use super::restart_local_agent_on_config_change_for;
     use crate::commands::global_agent_config::RestartOutcome;
+    use crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK;
+
+    let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
 
     let tmp = tempfile::tempdir().unwrap();
     // Malformed JSON → load_agent_store_at (called by load_personas_at) returns
@@ -386,9 +395,15 @@ async fn test_context_load_failure_leaves_runtime_running() {
 /// Thufir's test 2: "real production driver/core with injected preflight error;
 /// stop never called, RestartOutcome::Skipped."
 #[tokio::test]
+#[allow(clippy::await_holding_lock)] // SCOPE_GENERATION_TEST_LOCK serialises parallel tests
 async fn test_mesh_preflight_failure_leaves_runtime_running() {
     use super::restart_local_agent_on_config_change_for;
     use crate::commands::global_agent_config::RestartOutcome;
+    use crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK;
+
+    let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
 
     let tmp = tempfile::tempdir().unwrap();
     // Provide persona and record files so context prep can pass them.
@@ -450,9 +465,15 @@ async fn test_mesh_preflight_failure_leaves_runtime_running() {
 /// Thufir's test 3: "injected preflight hook advances generation after it
 /// succeeds; epoch returns Skipped; stop never called."
 #[tokio::test]
+#[allow(clippy::await_holding_lock)] // SCOPE_GENERATION_TEST_LOCK serialises parallel tests
 async fn test_workspace_switch_after_preflight_aborts_before_stop() {
     use super::restart_local_agent_on_config_change_for;
     use crate::commands::global_agent_config::RestartOutcome;
+    use crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK;
+
+    let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
 
     let tmp = tempfile::tempdir().unwrap();
     std::fs::write(tmp.path().join("personas.json"), b"[]").unwrap();
@@ -533,15 +554,21 @@ async fn test_workspace_switch_after_preflight_aborts_before_stop() {
 /// old `context.mesh_model_id`, bypass the mismatch — stop would be called
 /// and the test would fail.
 #[tokio::test]
+#[allow(clippy::await_holding_lock)] // SCOPE_GENERATION_TEST_LOCK serialises parallel tests
 async fn test_record_mesh_change_after_preflight_aborts_before_stop() {
     use super::super::restart_local_agent_on_config_change_for;
     use crate::commands::global_agent_config::RestartOutcome;
+    use crate::managed_agents::scope::SCOPE_GENERATION_TEST_LOCK;
     use crate::managed_agents::{
         storage::save_managed_agents_at, BackendKind, ManagedAgentPairRuntime, ManagedAgentRecord,
         ManagedAgentRuntimeKey,
     };
     use tauri::Manager;
 
+    let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
+
     let tmp = tempfile::tempdir().unwrap();
     let pubkey = "aa".repeat(32);
     let relay_url = "wss://relay.example";

desktop/src-tauri/src/managed_agents/runtime_commands.rs

@@ -908,36 +908,27 @@ where
 /// Production adapter for [`compensate_drain_for`].
 ///
 /// Delegates to [`compensate_drain_with_hook`] with a no-op hook.
-/// Lock acquisition order: transition guard already held → store lock acquired
-/// → hook fires (no-op in production) → validate generation → load → restore/save.
 pub(crate) fn compensate_drain<R: tauri::Runtime>(
     app: &tauri::AppHandle<R>,
     stopped: &[DrainJournalEntry],
     captured_scope: &crate::managed_agents::scope::WorkspaceAgentScope,
     _rt_transition_held: std::sync::MutexGuard<'_, ()>,
 ) -> Option<String> {
-    compensate_drain_with_hook(app, stopped, captured_scope, _rt_transition_held, || {})
+    compensate_drain_with_hook(app, stopped, captured_scope, _rt_transition_held, |_| {})
 }
 
-/// Inner implementation of [`compensate_drain`] with an injectable
-/// `on_store_acquired` hook.
+/// Inner implementation of [`compensate_drain`] with an injectable `on_records_loaded` hook.
 ///
-/// Lock acquisition order:
-///   1. `_rt_transition_held` — already held by caller (passed by value).
-///   2. Acquire `managed_agents_store_lock`.
-///   3. Call `on_store_acquired` (no-op in production; tests inject a closure
-///      that writes a sentinel and synchronises with a writer thread to prove
-///      genuine store-lock contention).
-///   4. Validate `captured_scope.generation` under the store lock.
-///   5. `load_managed_agents_at` — records loaded AFTER both locks held.
-///   6. Delegate to [`compensate_drain_for`] — store guard held through every
-///      save performed by `start_pair_under_held_locks`.
+/// Lock order: transition guard held by caller → acquire store lock → validate generation
+/// → load records → `on_records_loaded(&mut records)` (no-op in production; tests inject
+/// a sentinel mutation) → delegate to `compensate_drain_for` → save records (step 7, still
+/// under the store lock so the writer cannot interleave before the mutation hits disk).
 pub(crate) fn compensate_drain_with_hook<R: tauri::Runtime>(
     app: &tauri::AppHandle<R>,
     stopped: &[DrainJournalEntry],
     captured_scope: &crate::managed_agents::scope::WorkspaceAgentScope,
     _rt_transition_held: std::sync::MutexGuard<'_, ()>,
-    on_store_acquired: impl FnOnce(),
+    on_records_loaded: impl FnOnce(&mut Vec<super::ManagedAgentRecord>),
 ) -> Option<String> {
     if stopped.is_empty() {
         drop(_rt_transition_held);
@@ -955,10 +946,7 @@ pub(crate) fn compensate_drain_with_hook<R: tauri::Runtime>(
         }
     };
 
-    // 3. Fire the hook while both locks are held. No-op in production.
-    on_store_acquired();
-
-    // 4. Validate the captured scope under the store lock.
+    // 3. Validate the captured scope under the store lock.
     if let Err(stale_msg) = crate::managed_agents::scope::validate_scope_generation(captured_scope)
     {
         return Some(format!(
@@ -966,7 +954,7 @@ pub(crate) fn compensate_drain_with_hook<R: tauri::Runtime>(
         ));
     }
 
-    // 5. Load records under the held store lock.
+    // 4. Load records under the held store lock.
     let mut records = match load_managed_agents_at(&captured_scope.definitions_dir) {
         Ok(r) => r,
         Err(e) => {
@@ -976,8 +964,13 @@ pub(crate) fn compensate_drain_with_hook<R: tauri::Runtime>(
         }
     };
 
-    // 6. Delegate — store guard held through every save.
-    compensate_drain_for(stopped, &mut records, |entry, recs| {
+    // 5. Fire the hook with the loaded records. No-op in production.
+    //    Tests mutate a sentinel field here and synchronise with a writer
+    //    thread; the save at step 7 then makes that mutation load-bearing.
+    on_records_loaded(&mut records);
+
+    // 6. Delegate — store guard held through every save by start_pair_under_held_locks.
+    let result = compensate_drain_for(stopped, &mut records, |entry, recs| {
         start_pair_under_held_locks(
             app,
             &state,
@@ -988,7 +981,17 @@ pub(crate) fn compensate_drain_with_hook<R: tauri::Runtime>(
             recs,
         )
         .map(|_| ())
-    })
+    });
+
+    // 7. Save the (hook-mutated) records while the store guard is still held.
+    //    In production this is a no-op duplicate of the save inside
+    //    start_pair_under_held_locks; in tests it persists the sentinel
+    //    mutation so the writer cannot interleave before it hits disk.
+    if let Err(e) = save_managed_agents_at(&captured_scope.definitions_dir, &records) {
+        return Some(format!("compensation failed: could not save records: {e}"));
+    }
+
+    result
 }
 
 #[cfg(test)]

desktop/src-tauri/src/managed_agents/runtime_commands_concurrency_tests.rs

@@ -8,32 +8,47 @@ use super::*;
 
 /// Writer-vs-compensation store-lock contention via the `compensate_drain_with_hook` seam.
 ///
-/// Invariant: if the `managed_agents_store_lock` guard is dropped before
-/// `compensate_drain_for`'s restore/save, the writer can interleave and
-/// overwrite `COMP_SENTINEL` — making the final assertion fail.
+/// Invariant: if `managed_agents_store_lock` is dropped before the adapter's
+/// step-7 save in `compensate_drain_with_hook`, the writer acquires the lock
+/// before `COMP_SENTINEL` reaches disk, loads records without it, and saves
+/// only `WRITER_EDIT` — making the `COMP_SENTINEL` assertion fail deterministically.
 ///
 /// Flow:
-///   1. Seed the store with one agent record.
+///   1. Seed the store with one agent record; seed a live runtime so that
+///      `start_pair_under_held_locks` returns `AlreadyRunning` (compensation
+///      succeeds without spawning — `comp_result` is `None`).
 ///   2. Acquire `managed_agent_runtime_transition`; call `compensate_drain_with_hook`.
-///   3. `on_store_acquired` hook (fires while the store lock is held):
-///      a. writes `COMP_SENTINEL` to disk;
-///      b. signals the writer thread (`store_acquired_tx`);
-///      c. waits for the writer to queue at the store lock (`writer_queued_rx`).
-///   4. Writer thread: waits for (b), signals (c), then blocks on `managed_agents_store_lock`.
-///      Compensation still holds the lock — writer is blocked.
-///   5. `compensate_drain_with_hook` continues: validate → load → restore → save.
-///      Store lock still held throughout; writer remains blocked.
-///   6. Function returns; store lock released. Writer acquires it, writes `WRITER_EDIT`, saves.
-///   7. Final disk record must contain BOTH `COMP_SENTINEL` AND `WRITER_EDIT`.
+///   3. `on_records_loaded` hook fires AFTER both locks are held:
+///      a. signals the writer thread (`records_loaded_tx`);
+///      b. waits for writer's pre-lock signal (`writer_at_store_lock_rx`) — sent
+///      immediately before `managed_agents_store_lock.lock()`, so when received,
+///      the writer's next instruction is that lock call (which blocks);
+///      c. mutates `COMP_SENTINEL` on the in-memory adapter-loaded records.
+///   4. `compensate_drain_with_hook` continues: delegate to `compensate_drain_for`
+///      (succeeds — AlreadyRunning), then saves the hook-mutated records at step 7.
+///      Store lock held throughout; writer remains blocked on the store lock.
+///   5. Adapter releases the store lock. Writer acquires it, loads records
+///      (which now include `COMP_SENTINEL` from step 3c's in-memory mutation + step 4's
+///      save), writes `WRITER_EDIT`, saves.
+///   6. Final disk record must contain BOTH `COMP_SENTINEL` AND `WRITER_EDIT`.
+///      `comp_result` must be `None` (compensation succeeded).
 ///
-/// If the guard is dropped early, the writer acquires the lock before save,
-/// overwrites `COMP_SENTINEL`, and the assertion at step 7 fails.
+/// What breaks it: if the store guard is dropped before the adapter's step-7 save,
+/// the writer acquires `managed_agents_store_lock`, loads records BEFORE `COMP_SENTINEL`
+/// reaches disk (the in-memory mutation has not been saved yet), and saves without
+/// `COMP_SENTINEL` — the final assertion fails deterministically.
 #[test]
 fn test_compensate_drain_writer_vs_compensation_deterministic() {
-    use crate::managed_agents::scope::{current_scope_generation, WorkspaceAgentScope};
+    use crate::managed_agents::scope::{
+        current_scope_generation, WorkspaceAgentScope, SCOPE_GENERATION_TEST_LOCK,
+    };
     use std::thread;
     use tauri::Manager;
 
+    let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
+
     let tmp = tempfile::tempdir().unwrap();
     let tmp_path = tmp.path().to_path_buf();
 
@@ -107,6 +122,40 @@ fn test_compensate_drain_writer_vs_compensation_deterministic() {
     let app_handle = app.app_handle().clone();
     let state = app.state::<crate::app_state::AppState>();
 
+    // Seed a live runtime so `start_pair_under_held_locks` returns AlreadyRunning.
+    // This makes `compensate_drain_for` succeed (`comp_result = None`) without
+    // needing a real agent binary — compensation finds the agent already started.
+    let rt_key =
+        crate::managed_agents::ManagedAgentRuntimeKey::new(pubkey1.clone(), "wss://relay.example")
+            .unwrap();
+    {
+        let mut runtimes = state.managed_agent_processes.lock().unwrap();
+        let child = spawn_long_lived_child_for_test();
+        let process = crate::managed_agents::ManagedAgentProcess {
+            child,
+            log_path: std::path::PathBuf::new(),
+            spawn_config: crate::managed_agents::spawn_snapshot::prospective_spawn_config_snapshot(
+                &initial_record,
+                &[],
+                &[],
+                "wss://relay.example",
+                &Default::default(),
+            ),
+            setup_mode: false,
+            adapter_availability: None,
+            start_nonce: "test-nonce-writer".to_string(),
+            #[cfg(windows)]
+            job: None,
+        };
+        runtimes.insert(
+            rt_key.clone(),
+            crate::managed_agents::ManagedAgentPairRuntime::starting(
+                process,
+                Some("comp-drain-writer-test".to_string()),
+            ),
+        );
+    }
+
     let gen = current_scope_generation();
     let scope = WorkspaceAgentScope {
         scope_id: "comp-drain-writer-test".to_string(),
@@ -120,28 +169,41 @@ fn test_compensate_drain_writer_vs_compensation_deterministic() {
     let entry1 = make_drain_entry(&pubkey1, "wss://relay.example", true);
     let stopped = vec![entry1];
 
-    // store_acquired: hook → writer (compensation holds the store lock; writer may queue)
-    // writer_queued:  writer → hook  (writer is now blocked on the store lock)
-    let (store_acquired_tx, store_acquired_rx) = std::sync::mpsc::channel::<()>();
-    let (writer_queued_tx, writer_queued_rx) = std::sync::mpsc::channel::<()>();
+    // records_loaded:     hook → writer   (compensation holds store lock; writer may proceed)
+    // writer_at_store_lock: writer → hook  (writer is about to call managed_agents_store_lock.lock())
+    //
+    // The writer sends writer_at_store_lock immediately BEFORE the lock call, so the
+    // hook's recv() completing is a happens-before guarantee that the writer's next
+    // instruction is managed_agents_store_lock.lock() — which blocks because
+    // compensation holds it.
+    let (records_loaded_tx, records_loaded_rx) = std::sync::mpsc::channel::<()>();
+    let (writer_at_store_lock_tx, writer_at_store_lock_rx) = std::sync::mpsc::channel::<()>();
 
     // Spawn the writer thread BEFORE acquiring the transition guard.
-    // It waits for store_acquired, signals writer_queued (it is now about to
-    // block on the store lock), then acquires the store lock and writes WRITER_EDIT.
+    // The writer directly acquires managed_agents_store_lock after signalling —
+    // no intermediate lock — so its blocking is specifically on the store lock
+    // that compensate_drain_with_hook holds.  This makes the test fail if the
+    // adapter drops the store guard before the step-7 save.
     let tmp_wr = tmp_path.clone();
     let app_handle_wr = app_handle.clone();
     let wr_thread = thread::spawn(move || {
-        // Wait until the hook signals that compensation holds the store lock.
-        store_acquired_rx.recv().unwrap();
+        // Wait until on_records_loaded signals that compensation holds the store lock.
+        records_loaded_rx.recv().unwrap();
 
-        // Signal the hook that we are about to queue on the store lock.
-        // The hook will return on receipt, letting compensate_drain_with_hook
-        // proceed to validate → load → save while we are blocked below.
-        writer_queued_tx.send(()).unwrap();
+        // Signal the hook that we are about to call managed_agents_store_lock.lock().
+        // The hook's recv() will happen-before our lock call, so when on_records_loaded
+        // mutates COMP_SENTINEL and returns, we are guaranteed to be BLOCKED on the
+        // store lock — not merely about to call it.
+        writer_at_store_lock_tx.send(()).unwrap();
 
-        // Block on the store lock — compensation still holds it.
+        // Block on the store lock.  Compensation still holds it; we wait here until
+        // the adapter's step-7 save completes and the lock is released.
         let writer_state = app_handle_wr.state::<crate::app_state::AppState>();
         let _store = writer_state.managed_agents_store_lock.lock().unwrap();
+
+        // Store lock acquired: compensation has finished and released the lock.
+        // Load records — COMP_SENTINEL must be present because the adapter's
+        // step-7 save wrote it before releasing the lock.
         let mut records =
             crate::managed_agents::storage::load_managed_agents_at(&tmp_wr).unwrap_or_default();
         for r in &mut records {
@@ -153,38 +215,66 @@ fn test_compensate_drain_writer_vs_compensation_deterministic() {
 
     let rt_guard = state.managed_agent_runtime_transition.lock().unwrap();
 
-    // on_store_acquired fires while the store lock is held:
-    //   (a) write COMP_SENTINEL to disk — proves we are mid-sequence under the lock;
-    //   (b) signal the writer (store_acquired_tx);
-    //   (c) wait for the writer to queue at the lock boundary (writer_queued_rx).
-    // Returning lets compensate_drain_with_hook continue to validate → load → save,
-    // all while the store lock remains held, keeping the writer blocked.
-    let _comp_result = compensate_drain_with_hook(&app_handle, &stopped, &scope, rt_guard, || {
-        // (a) Write the compensation sentinel under the held store lock.
-        let mut records =
-            crate::managed_agents::storage::load_managed_agents_at(&tmp_path).unwrap_or_default();
-        for r in &mut records {
-            r.env_vars
-                .insert("COMP_SENTINEL".to_string(), "yes".to_string());
-        }
-        crate::managed_agents::storage::save_managed_agents_at(&tmp_path, &records).unwrap();
+    // on_records_loaded hook fires after load, while BOTH locks are held:
+    //   (a) signal the writer thread that compensation holds the store lock;
+    //   (b) wait for the writer to confirm it is at the store-lock boundary —
+    //       the writer sends this signal immediately BEFORE calling
+    //       managed_agents_store_lock.lock(), so by the time recv() returns
+    //       here, the writer is queued (blocked) on the store lock;
+    //   (c) mutate COMP_SENTINEL in-memory on the adapter-loaded records.
+    // The adapter then saves these hook-mutated records (step 7) before releasing
+    // the store lock. Only then does the writer acquire the lock and read records.
+    //
+    // What breaks it: if the store guard is dropped before the adapter's step-7
+    // save, the writer acquires managed_agents_store_lock before COMP_SENTINEL
+    // reaches disk — the writer's final save omits COMP_SENTINEL, failing the
+    // assertion below.
+    let comp_result =
+        compensate_drain_with_hook(&app_handle, &stopped, &scope, rt_guard, |records| {
+            // (a) Tell the writer that records are loaded; compensation holds both locks.
+            records_loaded_tx.send(()).unwrap();
 
-        // (b) Tell the writer the store lock is held.
-        store_acquired_tx.send(()).unwrap();
+            // (b) Wait for the writer to reach the store-lock boundary.
+            // After this recv() completes, the writer has sent its signal and
+            // its very next instruction is managed_agents_store_lock.lock().
+            // The store lock is held by compensation, so the writer will block.
+            writer_at_store_lock_rx.recv().unwrap();
 
-        // (c) Wait for the writer to queue at the store lock boundary.
-        writer_queued_rx.recv().unwrap();
-    });
+            // (c) Mutate COMP_SENTINEL in-memory. The adapter's step-7 save writes this
+            // to disk before releasing the store lock — making the mutation load-bearing.
+            for r in records.iter_mut() {
+                r.env_vars
+                    .insert("COMP_SENTINEL".to_string(), "yes".to_string());
+            }
+        });
 
     wr_thread.join().expect("writer thread panicked");
 
+    // Kill the seeded long-lived child.
+    let seeded_pid = {
+        let runtimes = state.managed_agent_processes.lock().unwrap();
+        runtimes.get(&rt_key).map(|r| r.child.id())
+    };
+    if let Some(pid) = seeded_pid {
+        let _ = crate::managed_agents::terminate_process(pid);
+    }
+
+    // Compensation must have succeeded: agent was AlreadyRunning, no errors.
+    // If the record had an invalid nsec or the seam failed, comp_result would be Some.
+    assert!(
+        comp_result.is_none(),
+        "compensation must succeed (AlreadyRunning path): {comp_result:?}"
+    );
+
     // ── Final disk state: BOTH effects must be present ───────────────────────
-    // COMP_SENTINEL: written by the hook while compensation held the store lock.
-    // WRITER_EDIT:   written by the writer after compensation released the lock.
+    // COMP_SENTINEL: set in-memory by the hook, saved by the adapter's step-7 save
+    //               while the store lock was still held — before the writer could load.
+    // WRITER_EDIT:   written by the writer after the adapter released the store lock.
     //
-    // If the store guard is dropped before compensate_drain_for's save, the
-    // writer can acquire the lock early and overwrite COMP_SENTINEL — this
-    // assertion would then fail, proving the invariant.
+    // If the store guard is dropped before the adapter's step-7 save, the writer
+    // acquires managed_agents_store_lock and loads records before COMP_SENTINEL
+    // reaches disk — the writer's save omits COMP_SENTINEL, and the first
+    // assertion below fails.
     let final_records =
         crate::managed_agents::storage::load_managed_agents_at(&tmp_path).unwrap_or_default();
     let final_rec = final_records
@@ -195,8 +285,9 @@ fn test_compensate_drain_writer_vs_compensation_deterministic() {
     assert_eq!(
         final_rec.env_vars.get("COMP_SENTINEL").map(String::as_str),
         Some("yes"),
-        "COMP_SENTINEL must survive — fails if the store guard is dropped before \
-         compensate_drain_for's save, allowing the writer to overwrite it"
+        "COMP_SENTINEL must reach disk via the adapter's step-7 save while the store \
+         lock is held; fails if the store guard is dropped before that save, allowing \
+         the writer to load and save records without COMP_SENTINEL"
     );
     assert_eq!(
         final_rec.env_vars.get("WRITER_EDIT").map(String::as_str),
@@ -205,36 +296,66 @@ fn test_compensate_drain_writer_vs_compensation_deterministic() {
     );
 }
 
-/// Production start-path contender is blocked while `compensate_drain_with_hook`
-/// holds the `managed_agent_runtime_transition` mutex.
+/// Production start-path contender is blocked on `managed_agent_runtime_transition`
+/// while `compensate_drain_with_hook` holds it, and the `on_transition_acquired` hook
+/// in `start_pair_lazy_for_with_hook` fires only AFTER the transition lock is released.
 ///
-/// Invariant: if the `managed_agent_runtime_transition` guard is dropped before
-/// `compensate_drain_with_hook` finishes, `start_managed_agent_runtime_pair_lazy`
-/// acquires the transition lock and proceeds — `contender_done_rx` would fire
-/// before the hook completes, and the assertion at the end would fail.
+/// Invariant: if `managed_agent_runtime_transition` is removed from the start seam,
+/// the contender fires `on_transition_acquired` immediately after `on_before_transition`
+/// (with zero lock contention) — before compensation has a chance to complete. The
+/// `on_records_loaded` hook detects this via `try_recv()` on the channel that
+/// `on_transition_acquired` sends to, making the test fail deterministically:
+///
+///   - With the lock PRESENT: contender blocks at `managed_agent_runtime_transition.lock()`
+///     immediately after `on_before_transition`. `on_transition_acquired` cannot fire while
+///     compensation holds the lock. `try_recv()` inside `on_records_loaded` returns `Err(Empty)`.
+///     After compensation releases the lock, contender fires `on_transition_acquired` and the
+///     `recv_timeout()` below succeeds.
+///
+///   - With the lock REMOVED: contender fires `on_before_transition` (sends
+///     `contender_at_boundary`), then immediately fires `on_transition_acquired` (sends
+///     to `contender_transition_acquired_rx`). This fires in nanoseconds. By the time
+///     compensation reaches `on_records_loaded` (after file I/O for load_managed_agents),
+///     the channel already contains the message. `try_recv()` succeeds → assertion fails.
 ///
 /// Flow:
 ///   1. Seed the store; acquire `managed_agent_runtime_transition` in this thread.
-///   2. Spawn the contender. It signals "ready," then waits for the hook to
-///      signal "store held." It then signals "at_lock" and calls the production
-///      `start_managed_agent_runtime_pair_lazy` — which blocks acquiring
-///      `managed_agent_runtime_transition` as its first action.
-///   3. `compensate_drain_with_hook` hook: signals the contender ("store held"),
-///      waits for "at_lock" confirmation, then returns. The contender is now
-///      queued at the transition lock.
-///   4. `compensate_drain_with_hook` completes, releases the transition guard.
-///   5. The contender acquires the lock (fails at spawn — no binary — but
-///      passes the lock boundary), signals "done."
+///   2. Spawn the contender. It calls `start_pair_lazy_for_with_hook`:
+///      - `on_before_transition` fires (BEFORE the lock call) and sends
+///        `contender_at_boundary` — at this point the contender is inside the seam,
+///        about to block on `managed_agent_runtime_transition`.
+///   3. Wait for `contender_at_boundary` — the contender is now blocked at the
+///      transition lock because we hold it.
+///   4. Call `compensate_drain_with_hook`. The `on_records_loaded` hook:
+///      (a) asserts `transition_hook_fired` is false (atomic bool sanity check);
+///      (b) asserts `transition_hook_check_rx.try_recv()` is `Err(Empty)` — the
+///      deterministic proof that `on_transition_acquired` has not fired yet.
+///   5. `compensate_drain_with_hook` finishes; releases the transition guard.
+///   6. The contender acquires the guard; `on_transition_acquired` fires and
+///      sends `contender_transition_acquired`.
+///   7. Test asserts that `contender_transition_acquired` fires within a timeout.
 ///
-/// The contender drives `start_managed_agent_runtime_pair_lazy`, the same
-/// function called by the `start_managed_agent_runtime` Tauri command, proving
-/// that compensation blocks production starts — not just a raw mutex lock.
+/// What breaks it: if `managed_agent_runtime_transition` is removed from
+/// `start_pair_for`, `on_transition_acquired` fires before step 4's assertion (because
+/// the contender fires it in nanoseconds, while compensation needs milliseconds for
+/// file I/O before reaching `on_records_loaded`), making both the `try_recv()` and the
+/// atomic check fail.
 #[test]
 fn test_compensate_drain_concurrent_start_is_blocked() {
-    use crate::managed_agents::scope::{current_scope_generation, WorkspaceAgentScope};
+    use crate::managed_agents::scope::{
+        current_scope_generation, WorkspaceAgentScope, SCOPE_GENERATION_TEST_LOCK,
+    };
+    use std::sync::{
+        atomic::{AtomicBool, Ordering},
+        Arc,
+    };
     use std::thread;
     use tauri::Manager;
 
+    let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
+
     let tmp = tempfile::tempdir().unwrap();
     let tmp_path = tmp.path().to_path_buf();
 
@@ -320,14 +441,20 @@ fn test_compensate_drain_concurrent_start_is_blocked() {
     let entry1 = make_drain_entry(&pubkey1, "wss://relay.example", true);
     let stopped = vec![entry1];
 
-    // contender_ready:   contender → test  (about to call production start seam)
-    // hook_store_held:   hook → contender  (hook has store lock; contender may proceed)
-    // contender_at_lock: contender → hook  (contender queued at transition lock)
-    // contender_done:    contender → test  (contender passed the transition lock)
-    let (contender_ready_tx, contender_ready_rx) = std::sync::mpsc::channel::<()>();
-    let (hook_store_held_tx, hook_store_held_rx) = std::sync::mpsc::channel::<()>();
-    let (contender_at_lock_tx, contender_at_lock_rx) = std::sync::mpsc::channel::<()>();
-    let (contender_done_tx, contender_done_rx) = std::sync::mpsc::channel::<()>();
+    // (A) contender_at_boundary:         contender → test  (inside seam, about to block on transition lock)
+    // (B) contender_transition_acquired: contender → test  (start seam acquired transition guard)
+    //     Two receivers for the same semantic signal:
+    //     - transition_hook_check_rx: moved into on_records_loaded for try_recv() check
+    //     - contender_transition_acquired_rx: used after compensation for recv_timeout() check
+    let (contender_at_boundary_tx, contender_at_boundary_rx) = std::sync::mpsc::channel::<()>();
+    let (transition_hook_check_tx, transition_hook_check_rx) = std::sync::mpsc::channel::<()>();
+    let (contender_transition_acquired_tx, contender_transition_acquired_rx) =
+        std::sync::mpsc::channel::<()>();
+
+    // Shared flag: set to true when `on_transition_acquired` fires inside start_pair_for.
+    // Belt-and-suspenders companion to the try_recv() check in on_records_loaded.
+    let transition_hook_fired = Arc::new(AtomicBool::new(false));
+    let transition_hook_fired2 = transition_hook_fired.clone();
 
     // Acquire the transition guard FIRST so the contender will block on it.
     let rt_guard = state.managed_agent_runtime_transition.lock().unwrap();
@@ -335,46 +462,86 @@ fn test_compensate_drain_concurrent_start_is_blocked() {
     let app_contender = app_handle.clone();
     let pubkey_contender = pubkey1.clone();
     let contender = thread::spawn(move || {
-        // Signal: about to call the production start path.
-        contender_ready_tx.send(()).unwrap();
-
-        // Wait for the hook to confirm it holds the store lock.
-        hook_store_held_rx.recv().unwrap();
-
-        // Signal: queued at the transition lock boundary.
-        contender_at_lock_tx.send(()).unwrap();
-
-        // Production start-lock seam — acquires managed_agent_runtime_transition
-        // as its first action, so it blocks here until compensation releases it.
-        let _ = start_pair_lazy_for(
+        // `start_pair_lazy_for_with_hook` is the production-called seam.
+        // on_before_transition fires just BEFORE managed_agent_runtime_transition.lock().
+        // At that point we are inside the seam, about to block on the transition guard
+        // (which the test holds). Signaling from here proves the contender is inside
+        // the seam and will block on the next line — not merely about to call the function.
+        let _ = start_pair_lazy_for_with_hook(
             pubkey_contender,
             "wss://relay.example".to_string(),
             app_contender,
+            // on_before_transition: fires inside the seam, just before the lock call.
+            // The contender is now committed to acquiring managed_agent_runtime_transition
+            // and will block there immediately after this hook returns.
+            move || {
+                contender_at_boundary_tx.send(()).unwrap();
+                // Next line in the seam: managed_agent_runtime_transition.lock() — blocks.
+            },
+            // on_transition_acquired: fires AFTER the transition guard is acquired,
+            // BEFORE the store lock. Signals the test that start has passed the
+            // transition-lock boundary. Sends to two channels: one for the
+            // try_recv() check inside on_records_loaded, one for the final
+            // recv_timeout() check after compensation returns.
+            move || {
+                transition_hook_fired2.store(true, Ordering::SeqCst);
+                transition_hook_check_tx.send(()).unwrap();
+                contender_transition_acquired_tx.send(()).unwrap();
+            },
         );
-
-        // Signal: passed the transition lock (compensation is done).
-        contender_done_tx.send(()).unwrap();
     });
 
-    // Wait for the contender to be alive before calling the adapter.
-    contender_ready_rx.recv().unwrap();
+    // Wait for the contender to be inside the seam and about to block on the
+    // transition lock. After this signal the contender is on the next line:
+    // managed_agent_runtime_transition.lock() — which blocks because we hold it.
+    contender_at_boundary_rx.recv().unwrap();
 
-    // on_store_acquired hook: fires while BOTH locks are held.
-    //   (a) signal the contender so it proceeds to call the start seam;
-    //   (b) wait for the contender to confirm it is at the lock boundary.
-    // After (b) the contender is guaranteed to be blocked on
-    // managed_agent_runtime_transition — returning lets
-    // compensate_drain_with_hook finish its work before the contender unblocks.
-    let _comp_result = compensate_drain_with_hook(&app_handle, &stopped, &scope, rt_guard, || {
-        hook_store_held_tx.send(()).unwrap();
-        contender_at_lock_rx.recv().unwrap();
-    });
+    // on_records_loaded hook: fires while BOTH locks are held.
+    //
+    // Two complementary checks that on_transition_acquired has NOT fired:
+    //   (1) Atomic bool: fast sanity check.
+    //   (2) try_recv(): channel-based proof — on_transition_acquired sends to the
+    //       channel; if the channel is empty here, the hook has not fired.
+    //
+    // Why (2) is deterministic when the lock is removed: on_before_transition fires,
+    // on_transition_acquired fires immediately (zero lock contention), and the send
+    // completes in nanoseconds. By the time compensation reaches on_records_loaded
+    // (after acquiring the store lock and loading records from disk — milliseconds
+    // of I/O), the channel is guaranteed to contain the message. try_recv() finds it
+    // and the assertion fails, catching the missing lock.
+    //
+    // When the lock IS present: on_before_transition fires, then the contender
+    // blocks at managed_agent_runtime_transition.lock(). on_transition_acquired
+    // cannot fire until after compensation releases the guard. try_recv() correctly
+    // returns Err(Empty).
+    let _comp_result =
+        compensate_drain_with_hook(&app_handle, &stopped, &scope, rt_guard, |_records| {
+            // Check 1: atomic bool.
+            assert!(
+                !transition_hook_fired.load(Ordering::SeqCst),
+                "start seam must not acquire the transition guard while compensation holds it; \
+                 fails if managed_agent_runtime_transition is removed from start_pair_for"
+            );
+            // Check 2: channel try_recv — deterministic proof via file-I/O time differential.
+            // transition_hook_check_rx is a dedicated receiver that on_transition_acquired
+            // sends to; this closure moves it so the outer recv_timeout() uses the
+            // separate contender_transition_acquired_rx.
+            assert!(
+                transition_hook_check_rx.try_recv().is_err(),
+                "on_transition_acquired must not fire while compensation holds the transition \
+                 guard; fails if managed_agent_runtime_transition is removed from start_pair_for \
+                 (the hook fires in nanoseconds; this check runs after ms of file I/O)"
+            );
+        });
 
-    // Contender must now acquire the mutex within a reasonable timeout.
-    contender_done_rx
+    // After compensation returns, the contender can acquire the transition guard.
+    // `on_transition_acquired` fires and sends the signal.
+    contender_transition_acquired_rx
         .recv_timeout(std::time::Duration::from_secs(5))
         .expect(
-            "contender must unblock after compensate_drain_with_hook releases the transition guard",
+            "contender's on_transition_acquired must fire after compensate_drain_with_hook \
+             releases the transition guard; fails if start_pair_for_with_hook does not acquire \
+             managed_agent_runtime_transition before calling the hook",
         );
 
     contender.join().expect("contender thread panicked");

desktop/src-tauri/src/managed_agents/runtime_commands_tests.rs

@@ -5,6 +5,112 @@
 
 use super::*;
 
+/// Test seam: mirrors `start_pair_for` with two injectable hooks:
+///
+/// - `on_before_transition`: fires BEFORE `managed_agent_runtime_transition` is
+///   locked. The contender calls this to signal "I am at the lock boundary" from
+///   inside the function, giving the test deterministic evidence that the start
+///   seam is actually blocked on the lock (not merely about to call the function).
+///
+/// - `on_transition_acquired`: fires AFTER `managed_agent_runtime_transition` is
+///   acquired but BEFORE `managed_agents_store_lock` is attempted. Signals the
+///   test that start has passed the transition-lock boundary.
+///
+/// Used by concurrency tests to prove that the start seam is serialised by the
+/// transition guard: a contender cannot advance past `on_before_transition` while
+/// compensation holds the guard, and `on_transition_acquired` fires only after the
+/// guard is released.
+fn start_pair_for_with_hook<R: tauri::Runtime>(
+    pubkey: String,
+    relay_url: String,
+    lazy: bool,
+    expected_updated_at: Option<&str>,
+    app: tauri::AppHandle<R>,
+    on_before_transition: impl FnOnce(),
+    on_transition_acquired: impl FnOnce(),
+) -> Result<ManagedAgentRuntimeStatus, String> {
+    let state = app.state::<AppState>();
+    // Pre-acquisition hook: fires here, before the lock call. The contender uses
+    // this to signal "at the lock boundary" from inside the seam so the test
+    // knows the contender is blocked, not just about to call the function.
+    on_before_transition();
+    let _transition = state
+        .managed_agent_runtime_transition
+        .lock()
+        .map_err(|e| e.to_string())?;
+    if state
+        .shutdown_started
+        .load(std::sync::atomic::Ordering::Acquire)
+    {
+        return Err("desktop shutdown has started".into());
+    }
+    // Post-acquisition hook: transition guard held, store lock not yet acquired.
+    on_transition_acquired();
+    let _store = state
+        .managed_agents_store_lock
+        .lock()
+        .map_err(|e| e.to_string())?;
+    let mut records = load_managed_agents(&app)?;
+    start_pair_under_held_locks(
+        &app,
+        &state,
+        pubkey,
+        relay_url,
+        lazy,
+        expected_updated_at,
+        &mut records,
+    )
+}
+
+/// Test seam: calls `start_pair_for_with_hook` with a lazy=true start and both
+/// injectable hooks.
+fn start_pair_lazy_for_with_hook<R: tauri::Runtime>(
+    pubkey: String,
+    relay_url: String,
+    app: tauri::AppHandle<R>,
+    on_before_transition: impl FnOnce(),
+    on_transition_acquired: impl FnOnce(),
+) -> Result<ManagedAgentRuntimeStatus, String> {
+    start_pair_for_with_hook(
+        pubkey,
+        relay_url,
+        true,
+        None,
+        app,
+        on_before_transition,
+        on_transition_acquired,
+    )
+}
+
+/// Spawn a long-lived child process that stays running long enough for tests.
+///
+/// Cross-platform replacement for `sleep 10000` — seeds the in-memory runtimes
+/// map before `sync_managed_agent_processes` runs its `try_wait()` scan so the
+/// runtime survives to the eligibility check. Tests using this helper MUST kill
+/// the returned `Child` when the test exits to avoid leaking OS processes.
+fn spawn_long_lived_child_for_test() -> std::process::Child {
+    #[cfg(not(windows))]
+    {
+        std::process::Command::new("sh")
+            .args(["-c", "while true; do sleep 1; done"])
+            .stdin(std::process::Stdio::null())
+            .stdout(std::process::Stdio::null())
+            .stderr(std::process::Stdio::null())
+            .spawn()
+            .expect("spawn long-lived test child (sh loop)")
+    }
+    #[cfg(windows)]
+    {
+        std::process::Command::new("ping")
+            .args(["-n", "100000", "127.0.0.1"])
+            .stdin(std::process::Stdio::null())
+            .stdout(std::process::Stdio::null())
+            .stderr(std::process::Stdio::null())
+            .spawn()
+            .expect("spawn long-lived test child (ping)")
+    }
+}
+
 fn payload(
     relay_url: &str,
     lifecycle: ManagedAgentRuntimeLifecycle,
@@ -689,6 +795,9 @@ fn test_compensate_drain_empty_stopped_returns_none_with_real_app() {
 /// generation guard fires before any spawn attempt.
 #[test]
 fn test_compensate_drain_stale_scope_skips_all_with_real_app() {
+    let _gen_guard = super::super::scope::SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
     let tmp = tempfile::tempdir().unwrap();
     let (app, scope) = build_mock_app_with_scope(&tmp);
     let app_handle = app.app_handle().clone();
@@ -734,6 +843,9 @@ fn test_compensate_drain_stale_scope_skips_all_with_real_app() {
 /// with the stale-scope test that also needs it, making the window near-zero.
 #[test]
 fn test_compensate_drain_attempts_restart_and_reports_degradation_with_real_app() {
+    let _gen_guard = super::super::scope::SCOPE_GENERATION_TEST_LOCK
+        .lock()
+        .unwrap_or_else(|e| e.into_inner());
     let tmp = tempfile::tempdir().unwrap();
     std::fs::write(tmp.path().join("managed-agents.json"), b"[]").unwrap();
     let app = tauri::test::mock_builder()

desktop/src-tauri/src/managed_agents/scope.rs

@@ -341,6 +341,9 @@ mod tests {
 
     #[test]
     fn test_generation_increments_monotonically() {
+        let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+            .lock()
+            .unwrap_or_else(|e| e.into_inner());
         let before = current_scope_generation();
         let next = next_scope_generation();
         assert_eq!(next, before + 1);
@@ -356,6 +359,9 @@ mod tests {
     /// check `captured != current` is reliable.
     #[test]
     fn test_generation_staleness_detected_after_scope_change() {
+        let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+            .lock()
+            .unwrap_or_else(|e| e.into_inner());
         let captured = next_scope_generation(); // capture at operation entry
                                                 // Simulate a concurrent workspace switch bumping the generation.
         let after_switch = next_scope_generation();
@@ -377,6 +383,9 @@ mod tests {
     /// fields correctly reflect the active workspace at each step.
     #[test]
     fn test_scope_switch_a_to_b_to_a_advances_generation() {
+        let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+            .lock()
+            .unwrap_or_else(|e| e.into_inner());
         let base = std::env::temp_dir();
         let owner = "aa".repeat(32);
 
@@ -411,6 +420,9 @@ mod tests {
     /// staleness after C is committed.
     #[test]
     fn test_rapid_scope_switch_a_b_c_all_stale_after_c() {
+        let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+            .lock()
+            .unwrap_or_else(|e| e.into_inner());
         let base = std::env::temp_dir();
         let owner = "bb".repeat(32);
 
@@ -438,6 +450,9 @@ mod tests {
     /// source of truth.
     #[test]
     fn test_switch_during_restore_detected_by_generation_check() {
+        let _gen_guard = SCOPE_GENERATION_TEST_LOCK
+            .lock()
+            .unwrap_or_else(|e| e.into_inner());
         // Restore captures the generation at its entry.
         let captured_at_restore_entry = next_scope_generation();