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_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)