feat: add AIMD adaptive concurrency pacing per source

P6 round 2 showed routing convergence turns account-level 429s into
the binding constraint (62% throttle rate at full concurrency). Each
source now carries a local AIMD limit: multiplicative cut on 429,
additive growth on success. Over-limit picks queue via the existing
quota-wait poll instead of burning retry budget or tripping the
circuit-open verdict.

tests/unit/test_health_selector.py

@@ -72,3 +72,43 @@ class TestHealthAwareSelector:
     def test_missing_stats_defaults_to_zero_inflight(self):
         sel = HealthAwareSelector(rng=lambda: 0.0)
         assert [s.name for s in sel.order([_src("s1")], {})] == ["s1"]
+
+
+class TestAdaptivePacer:
+    """AIMD 自适应并发(M2.5 设计 §3.35): 429 乘性削减,成功加性增长。"""
+
+    def test_initial_and_bounds(self):
+        from polygateway.sources import AdaptivePacer
+
+        pacer = AdaptivePacer(ceiling=32.0)
+        assert pacer.limit("s1") == pytest.approx(8.0)
+        for _ in range(200):
+            pacer.on_backpressure("s1")
+        assert pacer.limit("s1") == pytest.approx(1.0)  # 下限 1
+        for _ in range(2000):
+            pacer.on_success("s1")
+        assert pacer.limit("s1") == pytest.approx(32.0)  # 上限 = ceiling
+
+    def test_cut_and_growth_math(self):
+        from polygateway.sources import AdaptivePacer
+
+        pacer = AdaptivePacer(ceiling=32.0)
+        pacer.on_backpressure("s1")
+        assert pacer.limit("s1") == pytest.approx(8.0 * 0.7)
+        before = pacer.limit("s1")
+        pacer.on_success("s1")
+        assert pacer.limit("s1") == pytest.approx(before + 1.0 / before)
+
+    def test_inflight_gate(self):
+        from polygateway.sources import AdaptivePacer
+
+        pacer = AdaptivePacer(ceiling=32.0)
+        for _ in range(200):
+            pacer.on_backpressure("s1")  # limit → 1
+        assert pacer.admit("s1") is True
+        pacer.enter("s1")
+        assert pacer.admit("s1") is False  # 在途 1 ≥ limit 1
+        pacer.leave("s1")
+        assert pacer.admit("s1") is True
+        pacer.leave("s1")  # 多余 leave 不下穿 0
+        assert pacer._inflight.get("s1", 0) == 0

tests/unit/test_retry.py

@@ -98,6 +98,7 @@ def _harness(
     global_limits=_NO_GLOBAL,
     rng=lambda: 0.0,
     selector=None,
+    pacer=None,
 ):
     clock = clock or FakeClock()
     limiter = InMemoryLimiter(
@@ -125,6 +126,7 @@ def _harness(
         now=clock,
         sleep=sleep,
         rng=rng,
+        pacer=pacer,
     )
     return mw, limiter, gate, transport, sleep, clock
 
@@ -443,3 +445,49 @@ class TestM25Orchestration:
             await mw(_REQ)
         g = gate._gates["a"]
         assert g.a0 + g.a1 == 0
+
+
+class TestAdaptivePacing:
+    """AIMD 接线(设计 §3.35): 429 收紧准入,paced 源等待而非烧预算。"""
+
+    async def test_paced_source_waits_without_consuming_budget(self):
+        from polygateway.sources import AdaptivePacer
+
+        pacer = AdaptivePacer(ceiling=32.0)
+        for _ in range(200):
+            pacer.on_backpressure("a")  # limit → 1
+        pacer.enter("a")  # 模拟一个在途占满名额
+        mw, _, _, transport, _, _ = _harness(
+            [_src("a")], [_ok()], quota_full="fail_fast", pacer=pacer
+        )
+        with pytest.raises(AllSourcesExhausted) as ei:
+            await mw(_REQ)
+        assert ei.value.reason == "quota_exhausted"  # 走配额等待通道,非 CircuitOpen
+        assert ei.value.per_source_reasons.get("a") == "adaptive_paced"
+        assert transport.calls == []  # 未发起尝试 → 不烧重试预算
+
+    async def test_429_cuts_limit_success_grows_it(self):
+        from polygateway.sources import AdaptivePacer
+
+        pacer = AdaptivePacer(ceiling=32.0)
+        mw, _, _, _, _, _ = _harness(
+            [_src("a")],
+            [TransientError("throttled", status_code=429), _ok()],
+            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))
+
+    async def test_inflight_returns_to_zero_after_call(self):
+        from polygateway.sources import AdaptivePacer
+
+        pacer = AdaptivePacer(ceiling=32.0)
+        mw, _, _, _, _, _ = _harness(
+            [_src("a"), _src("b")],
+            [TransientError("x"), _ok()],
+            pacer=pacer,
+        )
+        await mw(_REQ)
+        assert pacer._inflight.get("a", 0) == 0
+        assert pacer._inflight.get("b", 0) == 0