test(fase2): assert BudgetMiddleware short-circuits a real chat call + the debate

tests/test_vertical_slice_e2e.py

@@ -169,6 +169,36 @@ async def test_g_validated_proposal_derives_from_debate(
     assert isinstance(result.outcome, ValidatedProposal)
 
 
+async def test_h_tiny_budget_halts_via_debate_middleware(
+    docs_dir, make_client_factory, fresh_store, monkeypatch
+) -> None:
+    """F8 (debate-path half): a tiny budget must short-circuit in the DEBATE middleware, not the
+    later generate loop. Fence generation — if generate_via_llm is reached, the debate did NOT
+    short-circuit, so fail. Because generate_via_llm independently raises BudgetExceeded, a plain
+    pytest.raises would pass even with the debate middleware detached; the fence makes this test
+    load-bearing (detaching the middleware fails it)."""
+    from portfolio_optimiser import run as run_mod
+
+    async def _forbidden(*args, **kwargs):
+        raise AssertionError(
+            "generate_via_llm reached — debate did not short-circuit via middleware"
+        )
+
+    monkeypatch.setattr(run_mod, "generate_via_llm", _forbidden)
+    with pytest.raises(BudgetExceeded):
+        await run_project(
+            "FV42-GSV-E1",
+            "local",
+            docs_dir=docs_dir,
+            verdict_input=_VI,
+            client_factory=make_client_factory(
+                _VALID, tokens=100
+            ),  # 100 > cap 10, on the FIRST debate call
+            store=fresh_store,
+            max_tokens=10,
+        )
+
+
 async def test_f_malformed_contract_raises_before_any_chat(docs_dir, make_client_factory) -> None:
     calls = {"n": 0}
     base = make_client_factory(_VALID)