fix: keep keyword billing refusal reasons

src/tin_lite/keyword_plan_activities.py

@@ -312,7 +312,12 @@ async def _paid(
                     # Raw SDK/provider errors, headers and credentials never enter evidence;
                     # a model call keeps only a fixed label for why it failed.
                     result = (
-                        {"status": "unavailable", "reason": "spending_limit"}
+                        {
+                            "status": "unavailable",
+                            "reason": "spending_limit"
+                            if exc.code in {"spending_limit", "run_limit", "project_limit"}
+                            else f"billing_refused:{exc.code}",
+                        }
                         if isinstance(exc, BillingError)
                         else {
                             "status": "unknown",
@@ -1420,6 +1425,9 @@ def model_failure_cause(failure: dict) -> str:
     """Say in plain words why a model step stopped, from its stage and fixed reason label."""
     stage = STAGE_NAMES.get(failure.get("stage"), "keyword")
     reason = failure.get("reason") or ""
+    if reason.startswith("billing_refused:"):
+        code = reason.removeprefix("billing_refused:")
+        return f"a billing check refused the {stage} model call ({code})"
     if reason == "provider_timeout":
         seconds = MODEL_TIMEOUT_SECONDS.get(failure.get("stage"))
         limit = f"within {seconds // 60} minutes" if seconds else "in time"

tests/test_keyword_plan.py

@@ -555,6 +555,44 @@ async def test_unknown_attempt_never_repurchased_and_fingerprint_is_enforced():
     assert result["status"] == "unknown" and call.await_count == 0
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "code",
+    [
+        "spending_limit",
+        "run_limit",
+        "project_limit",
+        "unpriced_model",
+        "unmetered_operation",
+        "operation_already_attempted",
+        "usage_pending",
+        "insufficient_funds",
+        "spending_stopped",
+    ],
+)
+async def test_keyword_billing_refusal_keeps_its_cause_and_does_not_buy_again(code):
+    from tin_lite.billing_contracts import BillingError
+
+    activities, db, _, _, model = await fixture()
+    run_id = str(db.run.id)
+    model.generate.side_effect = BillingError(code, "Private diagnostic must not escape")
+    for _ in range(2):
+        with pytest.raises(ApplicationError):
+            await activities._model(run_id, "review", {})
+    is_limit = code in {"spending_limit", "run_limit", "project_limit"}
+    reason = "spending_limit" if is_limit else f"billing_refused:{code}"
+    receipt = db.effects[activities.key(run_id, "review")].result
+    assert receipt["status"] == "unavailable" and receipt["reason"] == reason
+    assert model.generate.await_count == 1
+    assert "Private diagnostic" not in json.dumps(receipt)
+    await activities.keyword_failure(run_id)
+    message = db.project_failure.await_args.kwargs["error_message"]
+    assert ("spending limit" in message) is is_limit
+    if not is_limit:
+        assert code in message and "could not be confirmed" not in message
+    assert "Private diagnostic" not in message
+
+
 @pytest.mark.asyncio
 async def test_budget_reserves_review_before_research_and_recovers_prepare_gap():
     activities, db, _, provider, _ = await fixture(budget=5)