diff --git a/open-sse/services/combo.ts b/open-sse/services/combo.ts
index 894c79033c..2b09178737 100644
--- a/open-sse/services/combo.ts
+++ b/open-sse/services/combo.ts
@@ -926,6 +926,9 @@ async function handleComboChatInner({
   if (activeNativeTurnPin) {
     orderedTargets = applyNativeCodexTurnPin(orderedTargets, activeNativeTurnPin);
     if (orderedTargets.length === 0) {
+      // #11371: quota-share ordering already reserved a winner slot; release it on
+      // this early exit (idempotent).
+      targetResolution.quotaShareRelease?.();
       return errorResponse(
         409,
         "The pinned native Codex turn target is no longer available; the turn cannot be moved to another provider"
@@ -952,7 +955,8 @@ async function handleComboChatInner({
     // Surface a recovery hint + auto-clear the session pin after enough consecutive
     // no-target failures (silent-stop fix). Threshold of 3 prevents a one-off account
     // wipe from destroying the prompt-cache pin benefit on the next request.
-    recordComboFailure(effectiveSessionId, combo.name);
+    // #11371: same early-exit release as the pinned-turn path above.
+    targetResolution.quotaShareRelease?.();
     return errorResponseWithComboDiagnostics(
       404,
       "Combo has no executable targets",
@@ -2841,6 +2845,9 @@ async function handleComboChatInner({
     return await dispatchWithCooldownRetry();
   } finally {
     quotaShareConcurrencyRelease?.();
+    // #11371: release the in-flight slot quota-share ordering reserved for its
+    // winner — the counter must not leak monotonically upward across requests.
+    targetResolution.quotaShareRelease?.();
     // G2: Clean up candidate registry to prevent unbounded memory growth.
     _unregisterExecutionCandidates(_registeredExecutionKeys);
   }
diff --git a/open-sse/services/combo/applyStrategyOrdering.ts b/open-sse/services/combo/applyStrategyOrdering.ts
index 38f07fbcdc..e077082c5e 100644
--- a/open-sse/services/combo/applyStrategyOrdering.ts
+++ b/open-sse/services/combo/applyStrategyOrdering.ts
@@ -20,6 +20,21 @@ import {
 } from "./targetSorters.ts";
 import type { ComboLike, ComboLogger, ResolvedComboTarget } from "./types.ts";
 
+/**
+ * Result of {@link applyStrategyOrdering}.
+ *
+ * `quotaShareRelease` carries the idempotent release for the in-flight slot that
+ * quota-share ordering reserves for its winner (#11371). It is non-null only when
+ * the `quota-share` strategy ran; every other strategy leaves it null. The caller
+ * MUST invoke it exactly once when the request settles — selection reserves the
+ * slot, so dropping the callback leaks the counter monotonically upward and
+ * degenerates P2C into "fewest lifetime dispatches".
+ */
+export interface ApplyStrategyOrderingResult {
+  orderedTargets: ResolvedComboTarget[];
+  quotaShareRelease: (() => void) | null;
+}
+
 export interface ApplyStrategyOrderingDeps {
   combo: ComboLike;
   config: Record<string, unknown>;
@@ -45,9 +60,10 @@ export async function applyStrategyOrdering(
   strategy: string,
   initialOrderedTargets: ResolvedComboTarget[],
   deps: ApplyStrategyOrderingDeps
-): Promise<ResolvedComboTarget[]> {
+): Promise<ApplyStrategyOrderingResult> {
   const { combo, config, body, log, apiKeyAllowedConnections, sessionKey } = deps;
   let orderedTargets = initialOrderedTargets;
+  let quotaShareRelease: (() => void) | null = null;
 
   if (strategy === "lkgp") {
     try {
@@ -229,14 +245,18 @@ export async function applyStrategyOrdering(
     const qsModel =
       typeof body?.model === "string" ? body.model : (orderedTargets[0]?.modelStr ?? "");
     const qsMaxConcurrent = await resolveMaxConcurrentByConnection(orderedTargets);
-    orderedTargets = selectQuotaShareTarget(orderedTargets, combo.name, qsModel, Date.now(), {
+    const qsSelection = selectQuotaShareTarget(orderedTargets, combo.name, qsModel, Date.now(), {
       maxConcurrentByConnection: qsMaxConcurrent,
-    }).orderedTargets;
+    });
+    orderedTargets = qsSelection.orderedTargets;
+    // #11371: the reservation made inside selectQuotaShareTarget must outlive this
+    // call — hand the release to the host so it can fire it when the request settles.
+    quotaShareRelease = qsSelection.decrementInflight;
     log.info(
       "COMBO",
       `Quota-share ordering: ${orderedTargets[0]?.modelStr}${orderedTargets[0]?.connectionId ? ` (${orderedTargets[0].connectionId})` : ""} selected (DRR+P2C)`
     );
   }
 
-  return orderedTargets;
+  return { orderedTargets, quotaShareRelease };
 }
diff --git a/open-sse/services/combo/targetResolution.ts b/open-sse/services/combo/targetResolution.ts
index eeec177c86..5022f2caa7 100644
--- a/open-sse/services/combo/targetResolution.ts
+++ b/open-sse/services/combo/targetResolution.ts
@@ -122,6 +122,13 @@ export interface ResolvedComboTargetPipeline {
   /** Session-stickiness result — the attempt loop reads `.messageHash` on success/failure. */
   sticky: ApplyStickinessResult;
   preScreenMap: Map<string, PreScreenResult>;
+  /**
+   * Idempotent release for the in-flight slot quota-share ordering reserved for
+   * its winner (#11371). Null unless the `quota-share` strategy ran. The host MUST
+   * invoke it when the request settles; this pipeline already releases it on any
+   * earlyResponse it produces after selection.
+   */
+  quotaShareRelease: (() => void) | null;
 }
 
 export type ResolveComboTargetPipelineResult =
@@ -394,7 +401,11 @@ async function orderByStrategy(
   initialOrderedTargets: ResolvedComboTarget[]
 ): Promise<
   | { earlyResponse: Response }
-  | { orderedTargets: ResolvedComboTarget[]; autoUsedExplicitRouter: boolean }
+  | {
+      orderedTargets: ResolvedComboTarget[];
+      autoUsedExplicitRouter: boolean;
+      quotaShareRelease: (() => void) | null;
+    }
 > {
   const { strategy, body, combo, settings, config, log } = deps;
   if (strategy === "auto") {
@@ -413,17 +424,22 @@ async function orderByStrategy(
     return {
       orderedTargets: autoResult.orderedTargets,
       autoUsedExplicitRouter: autoResult.autoUsedExplicitRouter,
+      quotaShareRelease: null,
     };
   }
-  const orderedTargets = await applyStrategyOrdering(strategy, initialOrderedTargets, {
-    combo,
-    config,
-    body,
-    log,
-    apiKeyAllowedConnections: deps.apiKeyAllowedConnections,
-    sessionKey: deps.relayOptions?.sessionId,
-  });
-  return { orderedTargets, autoUsedExplicitRouter: false };
+  const { orderedTargets, quotaShareRelease } = await applyStrategyOrdering(
+    strategy,
+    initialOrderedTargets,
+    {
+      combo,
+      config,
+      body,
+      log,
+      apiKeyAllowedConnections: deps.apiKeyAllowedConnections,
+      sessionKey: deps.relayOptions?.sessionId,
+    }
+  );
+  return { orderedTargets, autoUsedExplicitRouter: false, quotaShareRelease };
 }
 
 /**
@@ -714,10 +730,15 @@ export async function resolveComboTargetPipeline(
 
   const ordering = await orderByStrategy(deps, orderedTargets);
   if ("earlyResponse" in ordering) return ordering;
-  const { autoUsedExplicitRouter } = ordering;
+  const { autoUsedExplicitRouter, quotaShareRelease } = ordering;
 
   const continuity = await applyContinuityFilters(deps, ordering.orderedTargets);
-  if ("earlyResponse" in continuity) return continuity;
+  if ("earlyResponse" in continuity) {
+    // #11371: selection already reserved the winner's in-flight slot; a hard
+    // filter exhausting the pool must not leak it.
+    quotaShareRelease?.();
+    return continuity;
+  }
   orderedTargets = applyTaskAwareOrdering(deps, continuity.orderedTargets, autoUsedExplicitRouter);
   orderedTargets = await applyPromptCacheStage(
     deps,
@@ -741,5 +762,6 @@ export async function resolveComboTargetPipeline(
     getWeightedStepKeyForTarget,
     sticky: continuity.sticky,
     preScreenMap,
+    quotaShareRelease,
   };
 }
diff --git a/tests/unit/combo-apply-strategy-ordering-split.test.ts b/tests/unit/combo-apply-strategy-ordering-split.test.ts
index 12eebb96f6..90547c29a9 100644
--- a/tests/unit/combo-apply-strategy-ordering-split.test.ts
+++ b/tests/unit/combo-apply-strategy-ordering-split.test.ts
@@ -50,7 +50,7 @@ test("unknown strategy -> input order unchanged (same reference contents)", asyn
   const input = [target("openai", "gpt-4o"), target("anthropic", "claude-3")];
   const out = await applyStrategyOrdering("no-such-strategy", input, deps());
   assert.deepEqual(
-    out.map((t: { executionKey: string }) => t.executionKey),
+    out.orderedTargets.map((t: { executionKey: string }) => t.executionKey),
     ["openai>gpt-4o", "anthropic>claude-3"]
   );
 });
@@ -59,7 +59,7 @@ test("fill-first -> preserves priority order", async () => {
   const input = [target("a", "m1"), target("b", "m2"), target("c", "m3")];
   const out = await applyStrategyOrdering("fill-first", input, deps());
   assert.deepEqual(
-    out.map((t: { executionKey: string }) => t.executionKey),
+    out.orderedTargets.map((t: { executionKey: string }) => t.executionKey),
     ["a>m1", "b>m2", "c>m3"]
   );
 });
@@ -67,8 +67,8 @@ test("fill-first -> preserves priority order", async () => {
 test("random -> same multiset of targets (a permutation)", async () => {
   const input = [target("a", "m1"), target("b", "m2"), target("c", "m3")];
   const out = await applyStrategyOrdering("random", input, deps());
-  assert.equal(out.length, 3);
-  assert.deepEqual(keys(out), keys(input));
+  assert.equal(out.orderedTargets.length, 3);
+  assert.deepEqual(keys(out.orderedTargets), keys(input));
 });
 
 test("cost-optimized manifest routing logs through the canonical strategy path", async () => {
@@ -90,7 +90,7 @@ test("cost-optimized manifest routing logs through the canonical strategy path",
     log,
   } as never);
 
-  assert.equal(out.length, 2);
+  assert.equal(out.orderedTargets.length, 2);
   assert.equal(
     debugCalls.some((args) => args[1] === "manifest routing applied"),
     true,
diff --git a/tests/unit/prompt-cache-affinity.test.ts b/tests/unit/prompt-cache-affinity.test.ts
index 7e25456f02..eb980700f8 100644
--- a/tests/unit/prompt-cache-affinity.test.ts
+++ b/tests/unit/prompt-cache-affinity.test.ts
@@ -134,7 +134,7 @@ test("cache-optimized strategy routes a stable prompt key to the same account",
   };
   const first = await applyStrategyOrdering("cache-optimized", targets, deps);
   const second = await applyStrategyOrdering("cache-optimized", [...targets].reverse(), deps);
-  assert.equal(first[0].connectionId, second[0].connectionId);
+  assert.equal(first.orderedTargets[0].connectionId, second.orderedTargets[0].connectionId);
 });
 
 test("foreign OAuth session softly redirects cache affinity while the same session stays local", () => {
@@ -145,11 +145,18 @@ test("foreign OAuth session softly redirects cache affinity while the same sessi
   ];
   const fixture = Array.from({ length: 10_000 }, (_, index) => {
     const body = { prompt_cache_key: `occupied-cache-key-${index}` };
-    const baseline = applyPromptCacheAffinity(oauthTargets, body, true, "global", "session-a").targets;
+    const baseline = applyPromptCacheAffinity(
+      oauthTargets,
+      body,
+      true,
+      "global",
+      "session-a"
+    ).targets;
     const occupied = baseline[0];
     const alternative = baseline[1];
     const release = reserveOAuthSession(occupied.connectionId!, "session-a");
-    const foreignFirst = applyPromptCacheAffinity(oauthTargets, body, true, "global", "session-b").targets[0];
+    const foreignFirst = applyPromptCacheAffinity(oauthTargets, body, true, "global", "session-b")
+      .targets[0];
     release();
     return foreignFirst.connectionId === alternative.connectionId
       ? { body, occupied, alternative }
@@ -160,12 +167,14 @@ test("foreign OAuth session softly redirects cache affinity while the same sessi
   const release = reserveOAuthSession(occupied.connectionId!, "session-a");
 
   assert.equal(
-    applyPromptCacheAffinity(oauthTargets, body, true, "global", "session-a").targets[0].connectionId,
+    applyPromptCacheAffinity(oauthTargets, body, true, "global", "session-a").targets[0]
+      .connectionId,
     occupied.connectionId,
     "the owning session keeps its cache-local account"
   );
   assert.equal(
-    applyPromptCacheAffinity(oauthTargets, body, true, "global", "session-b").targets[0].connectionId,
+    applyPromptCacheAffinity(oauthTargets, body, true, "global", "session-b").targets[0]
+      .connectionId,
     alternative.connectionId,
     "a foreign session prefers the free OAuth account"
   );
diff --git a/tests/unit/quota-share-strategy.test.ts b/tests/unit/quota-share-strategy.test.ts
index 7d1e59f940..99d5c0fcbb 100644
--- a/tests/unit/quota-share-strategy.test.ts
+++ b/tests/unit/quota-share-strategy.test.ts
@@ -21,6 +21,7 @@ import {
   _clearDrrStateForTest,
   _getDrrDeficitForTest,
 } from "../../open-sse/services/combo/quotaShareStrategy.ts";
+import { applyStrategyOrdering } from "../../open-sse/services/combo/applyStrategyOrdering.ts";
 import {
   incrementInflight,
   decrementInflight,
@@ -490,3 +491,48 @@ describe("activation: qtSd/ combos use strategy 'quota-share'", () => {
     );
   });
 });
+
+// ─── #11371: release must travel out of the ordering path ───────────────────
+
+describe("applyStrategyOrdering threads the in-flight release (#11371)", () => {
+  const noopLog = { info() {}, warn() {}, error() {}, debug() {} };
+
+  test("quota-share ordering returns a release that restores the counter to 0", async () => {
+    const targets = [makeTarget("ek-rel-a", "conn-rel-a"), makeTarget("ek-rel-b", "conn-rel-b")];
+    const { orderedTargets, quotaShareRelease } = await applyStrategyOrdering(
+      "quota-share",
+      targets,
+      {
+        combo: { id: "c-rel", name: "qtSd/rel" },
+        config: {},
+        body: { model: "anthropic/claude-sonnet-4-5" },
+        log: noopLog,
+        apiKeyAllowedConnections: null,
+      } as never
+    );
+
+    assert.ok(orderedTargets.length > 0);
+    assert.ok(quotaShareRelease, "quota-share ordering must hand back a release callback");
+    const winnerConn = orderedTargets[0].connectionId ?? "";
+    assert.ok(winnerConn, "winner must carry a connectionId for the reservation");
+    // Selection reserved exactly one slot on the winner's connection.
+    assert.equal(getInflight(winnerConn, NOW), 1, "selection must reserve one in-flight slot");
+    // The real-caller contract: release once when the request settles.
+    quotaShareRelease!();
+    assert.equal(getInflight(winnerConn, NOW), 0, "release must return the counter to 0");
+    // Idempotent: a second call must not push the counter negative.
+    quotaShareRelease!();
+    assert.equal(getInflight(winnerConn, NOW), 0, "double release floors at 0");
+  });
+
+  test("non-quota-share strategies leave quotaShareRelease null", async () => {
+    const result = await applyStrategyOrdering("fill-first", [makeTarget("ek-null", "conn-null")], {
+      combo: { id: "c-ff", name: "plain" },
+      config: {},
+      body: {},
+      log: noopLog,
+      apiKeyAllowedConnections: null,
+    } as never);
+    assert.equal(result.quotaShareRelease, null);
+  });
+});