diff --git a/scripts/run-test-wave.py b/scripts/run-test-wave.py
index ce7e3d344..390cb666d 100755
--- a/scripts/run-test-wave.py
+++ b/scripts/run-test-wave.py
@@ -40,6 +40,15 @@
 # runner, survives only because it was already listed here.
 SLOW_SUITES = frozenset(("incremental", "store_arch", "daemon_runtime", "cli"))
 POLL_SECONDS = 0.05
+# Floor for how long the external Windows kill helper (taskkill.exe /T /F) may
+# take to answer. This is deliberately NOT --kill-grace: that flag budgets how
+# long a doomed process may take to die, while this budgets spawning the helper
+# on a loaded runner, which routinely exceeds a second. Wiring the two together
+# made a small kill grace flake the whole wave -- a timed-out taskkill is
+# reported as "could not prove cleanup", which raises out of the wave loop and
+# re-enters cleanup with the leader already dead. kill_grace still governs every
+# actual death wait. The descendant probe has its own budget below.
+WINDOWS_HELPER_TIMEOUT_SECONDS = 30
 
 # WHY: the Windows descendant probe below is a cold `powershell.exe` + CIM
 # start. On a GitHub Windows runner that routinely costs seconds -- interpreter
@@ -175,6 +184,10 @@ def start_suite(
     )
 
 
+def windows_helper_timeout(kill_grace: int) -> int:
+    return max(kill_grace, WINDOWS_HELPER_TIMEOUT_SECONDS)
+
+
 def windows_tree_cleanup_blocker(pid: int) -> str | None:
     """Why `pid`'s tree cannot be called clean, or None when it provably is.
 
@@ -256,7 +269,7 @@ def terminate_process_tree(active: ActiveSuite, kill_grace: int) -> None:
                 stdin=subprocess.DEVNULL,
                 stdout=subprocess.DEVNULL,
                 stderr=subprocess.DEVNULL,
-                timeout=kill_grace,
+                timeout=windows_helper_timeout(kill_grace),
             )
         except (OSError, subprocess.TimeoutExpired):
             completed = None
diff --git a/tests/test_parallel_harness_contract.sh b/tests/test_parallel_harness_contract.sh
index 31854b995..3e0c6c731 100755
--- a/tests/test_parallel_harness_contract.sh
+++ b/tests/test_parallel_harness_contract.sh
@@ -366,6 +366,7 @@ python3 - "$scheduler" "$fixture" "$(command -v python3)" <<'PY'
 from __future__ import annotations
 
 import ctypes
+import importlib.util
 import os
 import pathlib
 import signal
@@ -385,6 +386,36 @@ release = barrier / "timeout_exit_race.release"
 descendant_path = fixture / "descendant.pid"
 
 
+def scheduler_wait_budget() -> int:
+    """Seconds to allow the scheduler to finish refusing.
+
+    Generous on purpose: the scheduler's refusal is the asserted state, and this
+    bound only has to exceed its worst case; it never decides the verdict.
+
+    On POSIX the refusal is signal-driven and lands well inside --kill-grace.
+    On Windows it costs external helper spawns -- taskkill.exe, and powershell.exe
+    for the descendant probe -- which the scheduler deliberately budgets on their
+    own rather than with --kill-grace. Read those budgets from the scheduler
+    instead of restating them: hard-coding a number here silently turns a slow
+    runner into a harness failure the moment the two drift apart. The refusal
+    path can spend the budgets twice -- once in the wave loop, once in the
+    cleanup re-entry -- so allow both passes plus interpreter startup.
+    """
+    if os.name != "nt":
+        return 8
+    spec = importlib.util.spec_from_file_location("cbm_run_test_wave", scheduler)
+    module = importlib.util.module_from_spec(spec)
+    # Register before exec: @dataclass resolves annotations through
+    # sys.modules[cls.__module__], which is None for an unregistered module.
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    per_pass = module.WINDOWS_HELPER_TIMEOUT_SECONDS + (
+        module.WINDOWS_DESCENDANT_PROBE_SECONDS
+        * module.WINDOWS_DESCENDANT_PROBE_ATTEMPTS
+    )
+    return per_pass * 2 + 10
+
+
 def process_state(pid: int) -> str:
     if os.name == "nt":
         handle = ctypes.windll.kernel32.OpenProcess(0x101000, False, pid)
@@ -480,11 +511,7 @@ try:
             raise SystemExit("FAIL: scheduler did not observe the forced leader exit")
         time.sleep(0.02)
     release.write_text("release\n", encoding="utf-8")
-    # Generous on purpose: the scheduler's refusal is the asserted state, and
-    # on Windows it now spends up to the descendant-probe budget (twice --
-    # once in the wave loop, once in the cleanup pass) before refusing. This
-    # bound only has to exceed that worst case; it never decides the verdict.
-    stdout, stderr = process.communicate(timeout=120)
+    stdout, stderr = process.communicate(timeout=scheduler_wait_budget())
 
     if os.name == "nt":
         # Assert the PROPERTY, not the wording. This used to require the phrase