diff --git a/test/datasets_contract_test.exs b/test/datasets_contract_test.exs
index 3cdc7d8b..f5012289 100644
--- a/test/datasets_contract_test.exs
+++ b/test/datasets_contract_test.exs
@@ -206,11 +206,17 @@ defmodule DatasetsContractTest do
     File.write!(path, ["q,a\n" | rows] ++ ["bad,row,extra\n"])
 
     caller = spawn(fn -> Datasets.csv(path, [:q]) end)
-    walk = wait_for_monitored(caller)
+    walk = wait_for_walk(caller)
     ref = Process.monitor(walk)
+    # Suspended, the walk cannot run to its end and exit on its own, so it
+    # ends only when something kills it. The suspension is lifted if the test
+    # process exits.
+    :erlang.suspend_process(walk)
     Process.exit(caller, :kill)
 
-    assert_receive {:DOWN, ^ref, :process, ^walk, :killed}, 1_000
+    receive do
+      {:DOWN, ^ref, :process, ^walk, reason} -> assert reason == :killed
+    end
   after
     cleanup_tmp("killed-caller.csv")
   end
@@ -434,13 +440,15 @@ defmodule DatasetsContractTest do
     cleanup_tmp("typed-math.jsonl")
   end
 
-  # The process `pid` monitors, once it monitors one.
-  # The caller also monitors other processes for a moment (the file server
-  # while it reads), so the walk is the monitored process running the CSV code.
-  defp wait_for_monitored(pid, tries \\ 500) do
-    walk =
-      case Process.info(pid, :monitors) do
-        {:monitors, monitors} ->
+  # The walk `caller` starts, once it has started it. The caller also monitors
+  # other processes for a moment (the file server while it reads), so the walk
+  # is the monitored process running the CSV code. The caller parses the whole
+  # file before it starts the walk, which takes as long as the machine is slow,
+  # so this waits for as long as the caller is alive.
+  defp wait_for_walk(caller) do
+    case Process.info(caller, :monitors) do
+      {:monitors, monitors} ->
+        walk =
           Enum.find_value(monitors, fn
             {:process, monitored} when is_pid(monitored) ->
               if csv_walk?(monitored), do: monitored
@@ -449,15 +457,15 @@ defmodule DatasetsContractTest do
               nil
           end)
 
-        nil ->
-          nil
-      end
+        if walk do
+          walk
+        else
+          Process.sleep(5)
+          wait_for_walk(caller)
+        end
 
-    if walk || tries == 0 do
-      walk
-    else
-      Process.sleep(5)
-      wait_for_monitored(pid, tries - 1)
+      nil ->
+        flunk("the CSV caller ended before it started a walk")
     end
   end
 
diff --git a/test/parity_sidecar_test.exs b/test/parity_sidecar_test.exs
index 1935d9d9..d462feba 100644
--- a/test/parity_sidecar_test.exs
+++ b/test/parity_sidecar_test.exs
@@ -76,14 +76,19 @@ defmodule Imp.BenchmarkTruth.ParitySidecarTest do
     refute process_group_alive?(group_pid)
   end
 
+  test "the timeout option ends a command that does not finish" do
+    assert {:error, :timeout, %Output{}} =
+             ParitySidecar.run(@python, ["-c", "import time; time.sleep(60)"], timeout: 150)
+  end
+
   test "timeout returns captured output and terminates the process tree" do
     pid_path = tmp_path("timeout")
 
-    assert {:error, :timeout, %Output{} = output} =
-             ParitySidecar.run(@python, process_tree_args(pid_path), timeout: 150)
+    {result, [parent_pid, child_pid, group_pid]} =
+      time_out_ready_tree(process_tree_args(pid_path), pid_path)
 
+    assert {:error, :timeout, %Output{} = output} = result
     assert output.text =~ "tree ready"
-    [parent_pid, child_pid, group_pid] = await_pids(pid_path)
     refute await_alive?(parent_pid)
     refute await_alive?(child_pid)
     refute process_group_alive?(group_pid)
@@ -92,14 +97,13 @@ defmodule Imp.BenchmarkTruth.ParitySidecarTest do
   test "bounded cleanup kills a fully TERM-resistant process tree" do
     pid_path = tmp_path("term-resistant")
 
-    assert {:error, :timeout, %Output{}} =
-             ParitySidecar.run(
-               @python,
-               process_tree_args(pid_path, leader_ignore_term: true, child_ignore_term: true),
-               timeout: 150
-             )
+    {result, [parent_pid, child_pid, group_pid]} =
+      time_out_ready_tree(
+        process_tree_args(pid_path, leader_ignore_term: true, child_ignore_term: true),
+        pid_path
+      )
 
-    [parent_pid, child_pid, group_pid] = await_pids(pid_path)
+    assert {:error, :timeout, %Output{}} = result
     refute await_alive?(parent_pid)
     refute await_alive?(child_pid)
     refute process_group_alive?(group_pid)
@@ -108,18 +112,14 @@ defmodule Imp.BenchmarkTruth.ParitySidecarTest do
   test "Port exit does not hide a TERM-resistant child with redirected stdio" do
     pid_path = tmp_path("mixed-tree")
 
-    assert {:error, :timeout, %Output{text: text}} =
-             ParitySidecar.run(
-               @python,
-               process_tree_args(pid_path,
-                 child_ignore_term: true,
-                 child_redirect_stdio: true
-               ),
-               timeout: 150
-             )
+    {result, [parent_pid, child_pid, group_pid]} =
+      time_out_ready_tree(
+        process_tree_args(pid_path, child_ignore_term: true, child_redirect_stdio: true),
+        pid_path
+      )
 
+    assert {:error, :timeout, %Output{text: text}} = result
     assert text =~ "tree ready"
-    [parent_pid, child_pid, group_pid] = await_pids(pid_path)
     refute await_alive?(parent_pid)
     refute await_alive?(child_pid)
     refute process_group_alive?(group_pid)
@@ -203,11 +203,11 @@ defmodule Imp.BenchmarkTruth.ParitySidecarTest do
     #{if leader_ignore_term, do: "signal.signal(signal.SIGTERM, signal.SIG_IGN)", else: ""}
     child_code = #{inspect(child_code(child_ignore_term))}
     child = subprocess.Popen([sys.executable, "-c", child_code]#{child_stdio})
+    print("tree ready", flush=True)
     with open(#{inspect(pid_path)}, "w", encoding="utf-8") as handle:
         handle.write(f"{os.getpid()}\\n{child.pid}\\n{os.getpgrp()}\\n")
         handle.flush()
         os.fsync(handle.fileno())
-    print("tree ready", flush=True)
     time.sleep(60)
     """
 
@@ -220,25 +220,47 @@ defmodule Imp.BenchmarkTruth.ParitySidecarTest do
 
   defp child_code(false), do: "import time; time.sleep(60)"
 
-  defp await_pids(path, attempts \\ 100)
-  defp await_pids(_path, 0), do: flunk("sidecar did not publish process ids")
+  # A timeout is a timer in the process that owns the sidecar's port, counted
+  # from the start. A timer short enough for a fast test fires, on a loaded
+  # machine, before Python has started the tree, and a long one makes every
+  # run wait it out. So these tests run the tree with no timer, wait until it
+  # is ready, and send its owner the message the timer sends.
+  defp time_out_ready_tree(args, pid_path) do
+    task = Task.async(fn -> ParitySidecar.run(@python, args) end)
+    [parent_pid | _rest] = pids = await_pids(pid_path)
+    send(port_owner(parent_pid), :command_timeout)
+    {Task.await(task, :infinity), pids}
+  end
+
+  defp port_owner(os_pid) do
+    Enum.find_value(Port.list(), fn port ->
+      with {:os_pid, ^os_pid} <- Port.info(port, :os_pid),
+           {:connected, owner} <- Port.info(port, :connected) do
+        owner
+      else
+        _other -> nil
+      end
+    end) || flunk("no port runs OS process #{os_pid}")
+  end
 
-  defp await_pids(path, attempts) do
+  # The sidecar writes its process ids once the tree is running, which takes
+  # as long as the machine is slow.
+  defp await_pids(path) do
     case File.read(path) do
       {:ok, contents} ->
         case contents |> String.split("\n", trim: true) |> Enum.map(&Integer.parse/1) do
           [{parent, ""}, {child, ""}, {group, ""}] -> [parent, child, group]
-          _other -> retry_pids(path, attempts)
+          _other -> retry_pids(path)
         end
 
       _other ->
-        retry_pids(path, attempts)
+        retry_pids(path)
     end
   end
 
-  defp retry_pids(path, attempts) do
+  defp retry_pids(path) do
     Process.sleep(20)
-    await_pids(path, attempts - 1)
+    await_pids(path)
   end
 
   defp await_alive?(pid, attempts \\ 100)
diff --git a/test/test_helper.exs b/test/test_helper.exs
index 4e459308..9b0044e2 100644
--- a/test/test_helper.exs
+++ b/test/test_helper.exs
@@ -78,5 +78,14 @@ external_excludes =
 
 ExUnit.configure(exclude: external_excludes)
 
+# ReqLLM reads its model catalog from LLMDB, which loads it on first use under
+# a :global.trans lock. A process that finds the lock taken sleeps a random
+# backoff that doubles up to 8 s before it asks again, so when the async tests
+# at the start of the suite all make their first model lookup at once, some
+# wait several times as long as the load itself (a 1.4 s load, waits of up to
+# 4.6 s among 16 first lookups on an idle machine), and on a loaded runner past
+# the 5 s of a Task.await. Loading it here, before any test runs, leaves nothing to contend for.
+{:ok, _catalog} = LLMDB.load()
+
 Imp.Test.OwnLog.install()
 ExUnit.start()
diff --git a/test/timed_out_connection_test.exs b/test/timed_out_connection_test.exs
index d84a4bf8..965fa9be 100644
--- a/test/timed_out_connection_test.exs
+++ b/test/timed_out_connection_test.exs
@@ -23,21 +23,24 @@ defmodule Imp.TimedOutConnectionTest do
   }
 
   test "the call after a timed-out call is answered on a new connection" do
+    test_pid = self()
     {:ok, seen} = Agent.start_link(fn -> [] end)
 
     # Bandit serves each HTTP/1 connection from one process, so the handler's
-    # pid names the connection a request came in on. The first request is held
-    # past the client's timeout; every later one is answered at once.
+    # pid names the connection a request came in on. The first request is not
+    # answered while the test runs, so the first call times out however slow
+    # the machine is, and its connection stays busy: a request written behind
+    # it would not be answered either. Every later request is answered at once.
     base_url =
       Imp.Test.LocalHTTP.start(fn _request ->
         connection = self()
         count = Agent.get_and_update(seen, &{length(&1) + 1, &1 ++ [connection]})
 
         if count == 1 do
+          test_ref = Process.monitor(test_pid)
+
           receive do
-            :release -> :ok
-          after
-            2_000 -> :ok
+            {:DOWN, ^test_ref, :process, _pid, _reason} -> :ok
           end
         end
 
@@ -61,13 +64,16 @@ defmodule Imp.TimedOutConnectionTest do
     # connection only when it draws the same shard. Choosing the first shard
     # for every request makes that meeting certain.
     finch = [name: ReqLLM.Application.finch_name(), pool_strategy: &hd/1]
-    opts = [timeout: 200, max_retries: 0, req_http_options: [finch: finch]]
+    opts = [max_retries: 0, req_http_options: [finch: finch]]
     messages = [%{role: :user, content: "ping"}]
 
-    assert {:error, %Imp.LMError{retryable: true}} = Imp.LM.generate(lm, messages, opts)
-    # Only the first call has to time out. The second gets room to be answered
-    # on a loaded machine; landing on the stuck connection still fails below.
-    assert {:ok, _reply} = Imp.LM.generate(lm, messages, Keyword.put(opts, :timeout, 5_000))
+    assert {:error, %Imp.LMError{retryable: true}} =
+             Imp.LM.generate(lm, messages, Keyword.put(opts, :timeout, 200))
+
+    # The second call keeps ReqLLM's own receive timeout. On a new connection
+    # it is answered at once; on the stuck one it is not answered, and fails
+    # when that timeout runs out.
+    assert {:ok, _reply} = Imp.LM.generate(lm, messages, opts)
 
     assert [first, second] = Agent.get(seen, & &1)
     refute first == second