feat(billing): 长上下文阶梯定价（边际费率，与缓存折扣/峰段可叠加）

## 影响面实测

审计里 1000 条请求：**883 条（88.3%）prompt 超过 128k**，中位数 246849，
p90 298179，最大 354251。这不是小众场景，是 agent 负载的主流形态。

按「模型 × 源」拆开：其中真正按 token 计价的只有 alittokenplan 的 10 次
（deepseek-v4.1-flash，约 350k prompt），其余走 $70 套餐（commandcode）或
免费源（localzen），记 0 是正确的，对阶梯无影响。

那 10 次当前少算 54.9%（$0.3148 → $0.4877）。绝对值不大，但比例严重，
而且这个偏差只会随上下文变长而变大。

## 边际而非混合平均

阶梯只对**超过阈值的部分**重新计价，不是把整段按高价算。理由是可叠加性：
缓存读取同时享受缓存折扣**和**长上下文阶梯价，这与供应商的计费方式一致。
混合平均必须二选一，选错就正好低估「缓存命中的长请求」——而这恰是缓存
存在的意义所在（本部署缓存命中率 88%）。

超出部分按缓存/新鲜的比例拆分，不重不漏。

## DSL 两种写法

供应商两种发布方式都支持：绝对费率（OpenAI 报 \$3.00/M past 200k）和
倍数（DeepSeek 报 2x past 128k）。两者都缺必须报错 —— 一个没有费率的阶梯
会编译成一条「看起来处理了长上下文、实际没处理」的规则，静默少算且无迹可寻。
两者都给时以绝对费率为准，因为那是供应商真正报出的数字。

## 判据

Lua 侧两条（边际性、与缓存折扣叠加）+ Go 侧一条（两种写法与拒绝规则）。
变异自证三次：整段按高价 / 缓存部分不叠加折扣 / 缓存部分不计阶梯，
分别命中 0.40 vs 0.272、0.165 vs 0.2325、0.09 vs 0.2325。

规则编辑器加了三个输入（阈值 + 绝对费率 + 倍数），三项都留空 = 无阶梯 ——
与「阶梯费率为 0」刻意区分，所以是三个独立输入而不是从 0 推断是否存在阶梯。

Co-Authored-By: ModelRouter <noreply@modelrouter.dev>

internal/lua/billing_test.go

@@ -1139,3 +1139,100 @@ func TestReasoningClampedToCompletion(t *testing.T) {
 	// number plausible enough to survive a glance at a dashboard.
 	approx(t, "reasoning clamped to completion", cost, 0.00195)
 }
+
+// TestLongContextStepIsMarginal: the threshold reprices only the tokens PAST
+// it, at a marginal rate.
+//
+// A flat prompt price under-bills exactly the traffic an agent workload
+// produces — on this deployment 88% of requests exceed 128k with a median of
+// 247k — and it under-bills silently, because the number stays plausible. The
+// step must apply to the excess only; charging the WHOLE prompt at the higher
+// rate is the other obvious implementation and it overcharges short requests by
+// the same factor it was meant to fix long ones.
+func TestLongContextStepIsMarginal(t *testing.T) {
+
+	// Short request: below the threshold, so the step must not touch it at all.
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 128000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 100000, "completion_tokens": 0, "time": 1750000000000,
+	})
+	got := stateOf(t, ps)["total"].(map[string]interface{})["cost"].(float64)
+	approx(t, "short request unaffected", got, 0.1)
+
+	// Long request: 200k prompt with a 128k threshold → 72k tokens at the step.
+	ps2, _ := billingVM(t)
+	if err := ps2.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD",
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 128000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	ps2.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 200000, "completion_tokens": 0, "time": 1750000000000,
+	})
+	got2 := stateOf(t, ps2)["total"].(map[string]interface{})["cost"].(float64)
+	// 128000*1e-6 + 72000*2e-6 = 0.128 + 0.144 = 0.272
+	approx(t, "long request marginal step", got2, 0.272)
+}
+
+// TestLongContextComposesWithCacheDiscount pins the interaction, which is the
+// whole reason for computing the step as a marginal rate.
+//
+// On this deployment the cache hit rate is 88% and long prompts dominate, so
+// "cached tokens beyond the threshold" is the common case, not an edge case. A
+// cache read already bills at prompt*discount; the long-context rate applies on
+// top of THAT, not on top of the full prompt price. Computing it any other way
+// silently changes the price of the majority of traffic.
+func TestLongContextComposesWithCacheDiscount(t *testing.T) {
+	ps, _ := billingVM(t)
+	if err := ps.SetState("billing", map[string]interface{}{
+		"prices": map[string]interface{}{
+			"currency": "USD", "cache_discount": 0.1,
+			"models": map[string]interface{}{
+				"m": map[string]interface{}{
+					"prompt": 1e-6, "completion": 0,
+					"long_context": map[string]interface{}{"above": 100000, "prompt": 2e-6},
+				},
+			},
+		},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	// 200k prompt, 150k of it cached, 50k fresh. Excess past 100k = 100k, split
+	// 25% fresh / 75% cached.
+	ps.Fire(StageRequestEnd, map[string]interface{}{
+		"model": "m", "source": "s", "key": "***aaaaaa", "ok": true,
+		"prompt_tokens": 200000, "completion_tokens": 0,
+		"cache_hit_tokens": 150000, "cache_reported": true,
+		"time": 1750000000000,
+	})
+	got := stateOf(t, ps)["total"].(map[string]interface{})["cost"].(float64)
+	// base:  fresh 50000*1e-6            = 0.05
+	//        cache 150000*1e-6*0.1        = 0.015
+	// step:  fresh excess 25000*(2-1)e-6 = 0.025
+	//        cache excess 75000*(2-0.1)e-6 = 0.1425
+	// total                             = 0.2325
+	approx(t, "step composes with cache discount", got, 0.2325)
+}