feat: default two structured re-asks and sharper AIMD cut

Round 5 left three residual failure classes; ladder exhaustion (3.2%
of structured calls with a single re-ask) and 429 leakage (5.8%, AIMD
oscillating above the sustainable point) are addressable: re-ask
default goes 1 to 2 (conservative vs instructor's 3) and the AIMD cut
factor drops to 0.5.

tests/unit/test_config.py

@@ -148,6 +148,13 @@ class TestResilienceKeys:
 
 
 class TestAssemblyGuards:
+    def test_structured_retries_default_two(self):
+        # M2.5 迭代 4: 缺省重问 1→2(instructor 缺省 3 的保守版;P6 阶梯死亡 3.2% 实证)
+        s = GatewaySettings.from_env("LLM", env=_env())
+        assert s.structured_max_retries == 2
+        s2 = GatewaySettings.from_env("LLM", env=_env(PGW_STRUCTURED_MAX_RETRIES="0"))
+        assert s2.structured_max_retries == 0
+
     def test_cache_requires_namespace_and_ttl(self):
         env = _env(PGW_CACHE_BACKEND="memory")
         with pytest.raises(ValueError, match="NAMESPACE"):

tests/unit/test_health_selector.py

@@ -101,7 +101,7 @@ class TestAdaptivePacer:
 
         pacer = AdaptivePacer(ceiling=32.0)
         pacer.on_backpressure("s1")
-        assert pacer.limit("s1") == pytest.approx(8.0 * 0.7)
+        assert pacer.limit("s1") == pytest.approx(8.0 * 0.5)
         before = pacer.limit("s1")
         pacer.on_success("s1")
         assert pacer.limit("s1") == pytest.approx(before + 1.0 / before)

tests/unit/test_retry.py

@@ -482,8 +482,8 @@ class TestAdaptivePacing:
             pacer=pacer,
         )
         await mw(_REQ)
-        # 429 削减一次(8→5.6),随后成功加性增长(5.6 + 1/5.6)
-        assert pacer.limit("a") == pytest.approx(8.0 * 0.7 + 1.0 / (8.0 * 0.7))
+        # 429 削减一次(8→4),随后成功加性增长(4 + 1/4)
+        assert pacer.limit("a") == pytest.approx(8.0 * 0.5 + 1.0 / (8.0 * 0.5))
 
     async def test_inflight_returns_to_zero_after_call(self):
         from polygateway.sources import AdaptivePacer