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

tests/test_budget.py

@@ -33,9 +33,7 @@ async def _noop() -> None:
 
 def _resp(total: int | None) -> ChatResponse:
     usage = UsageDetails(total_token_count=total) if total is not None else None
-    return ChatResponse(
-        messages=[Message(role="assistant", contents=["x"])], usage_details=usage
-    )
+    return ChatResponse(messages=[Message(role="assistant", contents=["x"])], usage_details=usage)
 
 
 async def test_meter_reads_total_token_count_and_accumulates() -> None:
@@ -69,6 +67,24 @@ async def test_strict_usage_none_hard_fails() -> None:
         await mw.process(_Ctx(_resp(None)), _noop)  # type: ignore[arg-type]
 
 
+async def test_budget_middleware_fires_on_real_agent_chat(make_client_factory) -> None:
+    """F8 (real-client half): registering BudgetMiddleware on a real Agent (layered chat client
+    that carries ChatMiddlewareLayer) and running an actual chat call short-circuits with
+    BudgetExceeded — the middleware<->client integration the prior suite never exercised (the
+    minimal-base stand-in silently no-ops the middleware)."""
+    from agent_framework import Agent
+
+    client = make_client_factory("ok", tokens=8)("proposer")  # 8 tokens/reply > cap 5
+    meter = TokenMeter(Budget(max_tokens=5, max_rounds=10))
+    agent = Agent(
+        client, "propose a measure", name="proposer", middleware=[BudgetMiddleware(meter)]
+    )
+    with pytest.raises(BudgetExceeded) as exc:
+        await agent.run("hi")
+    assert exc.value.kind == "tokens"
+    assert meter.tokens == 8  # charged from the synthetic UsageDetails via the middleware
+
+
 def test_no_word_count_token_proxy_in_src() -> None:
     # The meter is fed from real UsageDetails, never a len(text.split()) word-count proxy
     # (research 03 Rec 3 — the Fase 1 _word_tokens anti-pattern is retired). NOTE: a bare

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)