fix(telemetry): gpu_tracked_call must open lease before GPU body

Caught live: /v1/gpu/status showed calls_total=0 after successful chat
calls because gpu_tracked_call never called lease_start, so lease_end
found no active call and dropped the record. Adds offline regression
tests for the success and error paths with module-state isolation.

app.py

@@ -403,6 +403,7 @@ def gpu_tracked_call(kind: str, fn, *args, model: str = "", **kwargs):
     is not observable; the recorded queue_wait_seconds is an upper bound.
     """
     t_submit = time.time()
+    call_id = GPU_USAGE.lease_start(kind)
 
     @spaces.GPU(duration=120)
     def _tracked_inner():
@@ -418,14 +419,11 @@ def gpu_tracked_call(kind: str, fn, *args, model: str = "", **kwargs):
         result, error, t_body_start = _tracked_inner()
     except Exception as exc:
         # Lease-level failure (e.g. 429 quota exhaustion): body never ran.
-        GPU_USAGE.lease_end(
-            f"{kind}-{int(t_submit * 1000)}", model=model, kind=kind,
-            error=str(exc),
-        )
+        GPU_USAGE.lease_end(call_id, model=model, kind=kind, error=str(exc))
         raise
     queue_wait = max(0.0, t_body_start - t_submit)
     GPU_USAGE.lease_end(
-        f"{kind}-{int(t_submit * 1000)}", model=model, kind=kind,
+        call_id, model=model, kind=kind,
         queue_wait_s=queue_wait, error=error,
     )
     if error:

tests/unit/test_gpu_usage_tracker.py

@@ -116,4 +116,44 @@ def test_usage_snapshot_shape(tracker):
     snap: Dict[str, Any] = tracker.usage_snapshot()
     assert set(snap.keys()) == {"daily", "profile"}
     assert set(snap["daily"].keys()) >= {"date", "gpu_seconds_used", "quota_consumed_percent"}
-    assert set(snap["profile"].keys()) >= {"total_calls", "avg_queue_wait_seconds"}
+    assert set(snap["profile"].keys()) >= {"total_calls", "avg_queue_wait_seconds"}
+
+
+@pytest.fixture()
+def fresh_usage():
+    app = _load_app()
+    app.GPU_USAGE = app.GPUUsageTracker()
+    yield app
+    app.GPU_USAGE = app.GPUUsageTracker()  # restore module-level state
+
+
+def test_gpu_tracked_call_records_usage(fresh_usage):
+    """Regression: gpu_tracked_call must open a lease BEFORE invoking the
+    GPU body, otherwise lease_end finds no active call and usage stays 0
+    (caught live: successful chat calls were not recorded)."""
+    app = fresh_usage
+    result = app.gpu_tracked_call(
+        "chat",
+        lambda *a, **k: {"choices": [{"message": {"content": "ok"}}]},
+        [],
+        model="test.gguf",
+    )
+    assert result["choices"][0]["message"]["content"] == "ok"
+    snap = app.GPU_USAGE.usage_snapshot()
+    assert snap["daily"]["calls_total"] == 1
+    assert snap["daily"]["calls_ok"] == 1
+    assert snap["profile"]["total_calls"] == 1
+
+
+def test_gpu_tracked_call_records_error_path(fresh_usage):
+    app = fresh_usage
+    with pytest.raises(RuntimeError):
+        app.gpu_tracked_call(
+            "chat",
+            lambda *a, **k: (_ for _ in ()).throw(RuntimeError("boom")),
+            [],
+            model="test.gguf",
+        )
+    snap = app.GPU_USAGE.usage_snapshot()
+    assert snap["daily"]["calls_total"] == 1
+    assert snap["daily"]["calls_failed"] == 1