Pin the C++ span names to the ladder, and stop writing them twice

tests/ut/cpp/hierarchical/test_scheduler.cpp

@@ -39,6 +39,7 @@
 
 #include "call_config.h"
 #include "common/host_span.h"
+#include "common/host_span_names.h"
 #include "common/host_span_scope.h"
 #include "orchestrator.h"
 #include "ring.h"
@@ -1210,13 +1211,13 @@ TEST(WorkerManagerTest, DispatchAndCompletionEmitHostSpans) {
 
     worker.dispatch(WorkerDispatch{slot, 0});
     ASSERT_TRUE(endpoint_ptr->wait_submitted(1));
-    EXPECT_TRUE(captured_host_span("node.dispatch"));
+    EXPECT_TRUE(captured_host_span(simpler::host_trace::host_span_name(simpler::host_trace::HostSpan::Dispatch)));
 
     const WorkerDispatch submitted = endpoint_ptr->submitted().front();
     endpoint_ptr->emit(WorkerProgressKind::COMPLETED, submitted);
     worker.progress();
     ASSERT_EQ(completed.size(), 1u);
-    EXPECT_TRUE(captured_host_span("node.complete"));
+    EXPECT_TRUE(captured_host_span(simpler::host_trace::host_span_name(simpler::host_trace::HostSpan::Complete)));
 
     worker.stop();
     allocator.shutdown();
@@ -1474,8 +1475,8 @@ TEST(WorkerManagerTest, TwoFrameLeaseSlotsDoNotDefineFifoOrAcceptance) {
     ASSERT_TRUE(endpoint.poll_progress(progress));
     EXPECT_EQ(progress.kind, WorkerProgressKind::ACCEPTED);
     EXPECT_EQ(progress.dispatch.dispatch_id, 42u);
-    EXPECT_TRUE(captured_host_span("node.frame_submit"));
-    EXPECT_TRUE(captured_host_span("node.activate"));
+    EXPECT_TRUE(captured_host_span(simpler::host_trace::host_span_name(simpler::host_trace::HostSpan::FrameSubmit)));
+    EXPECT_TRUE(captured_host_span(simpler::host_trace::host_span_name(simpler::host_trace::HostSpan::Activate)));
     allocator.shutdown();
 }
 
@@ -2042,10 +2043,13 @@ TEST_F(ProgressSchedulerFixture, GroupSubmitReportsNoSingleWorkerAndNoSingleInde
     );
     orchestrator.close_run_submission(run);
 
-    ASSERT_TRUE(captured_host_span("node.submit"));
+    ASSERT_TRUE(captured_host_span(simpler::host_trace::host_span_name(simpler::host_trace::HostSpan::Submit)));
     std::ostringstream expected;
     expected << "run_id=" << run << " task_slot=" << group.task_slot << " group_index=-1 group_size=2 role=facade";
-    EXPECT_EQ(captured_host_span_attrs("node.submit"), expected.str());
+    EXPECT_EQ(
+        captured_host_span_attrs(simpler::host_trace::host_span_name(simpler::host_trace::HostSpan::Submit)),
+        expected.str()
+    );
 }
 
 TEST_F(ProgressSchedulerFixture, SuccessorStagesButActivatesOnlyAfterFifoPromotion) {
@@ -2515,11 +2519,14 @@ TEST_F(SchedulerFixture, NextLevelSubmitEmitsAPairableHostSpan) {
 
     auto submitted = orch.submit_next_level(C(0x42), single_tensor_args(0xCAFE, TensorArgType::OUTPUT), cfg, 0);
 
-    ASSERT_TRUE(captured_host_span("node.submit"));
+    ASSERT_TRUE(captured_host_span(simpler::host_trace::host_span_name(simpler::host_trace::HostSpan::Submit)));
     std::ostringstream expected;
     expected << "run_id=" << run_id << " task_slot=" << submitted.task_slot
              << " group_index=0 group_size=1 worker_id=0 role=facade";
-    EXPECT_EQ(captured_host_span_attrs("node.submit"), expected.str());
+    EXPECT_EQ(
+        captured_host_span_attrs(simpler::host_trace::host_span_name(simpler::host_trace::HostSpan::Submit)),
+        expected.str()
+    );
 
     mock_worker.wait_running();
     mock_worker.complete();

tests/ut/py/test_worker/test_host_worker.py

@@ -19,6 +19,7 @@
 import inspect
 import multiprocessing.shared_memory as shared_memory_mod
 import struct
+import subprocess
 import sys
 import threading
 import time
@@ -65,6 +66,7 @@
     _mailbox_store_i32,
     _pack_py_callable_payload,
 )
+from simpler.worker_level import WorkerLevel
 
 from ._harness import chip_callable, fake_chip_l3, requires_sim_binaries
 
@@ -2755,7 +2757,8 @@ def bad_graph(*_args):
         with pytest.raises(ValueError, match="bad graph"):
             worker._submit_l3_locked(bad_graph, None, cast(Any, object()))
 
-        assert emitted == [("node.graph_build", 1, 0, 0, 100, 175, "run_id=1 role=facade")]
+        expected_name = f"{worker._host_span_prefix}.graph_build"
+        assert emitted == [(expected_name, 1, 0, 0, 100, 175, "run_id=1 role=facade")]
 
     def test_unsettled_graph_cancellation_abandons_the_handle_before_close(self):
         worker, events = self._submission_failure_worker(failures=2)
@@ -7655,3 +7658,25 @@ def test_unregister_removes_only_after_last_digest_ref(self):
             shm.unlink()
             payload_shm.close()
             payload_shm.unlink()
+
+
+def test_the_cpp_pre_bind_level_word_is_the_ladder_word_for_l3():
+    """`host_span_names.h` hand-writes the L3 word a third time, as its pre-bind default.
+
+    `WorkerLevel` is the source of truth and `strace_timing.py`'s copy is pinned to
+    it by its own test, but the C++ default is pinned to nothing — so a level word
+    could be renamed in Python while C++ keeps emitting the old one until a Worker
+    binds the prefix. Every span emitted before that binding would carry the stale
+    word, and no test would notice.
+
+    Read in a child process on purpose: the prefix freezes on first bind, so any
+    test in this process that constructed a Worker would leave the *bound* word
+    here instead of the default. Passing an empty word binds nothing (the setter
+    returns early) while the binding still reports what is currently in effect.
+    """
+    source = "from _task_interface import _set_host_span_level_prefix as bind; print(bind(''), end='')"
+    completed = subprocess.run([sys.executable, "-c", source], capture_output=True, text=True, check=True, timeout=120)
+
+    assert completed.stdout == WorkerLevel.node.name, (
+        f"C++ pre-bind level word is {completed.stdout!r}, ladder says {WorkerLevel.node.name!r}"
+    )