diff --git a/docs-site/src/content/docs/guides/combos.md b/docs-site/src/content/docs/guides/combos.md
index 6f51870ad3..b3bfd8f65b 100644
--- a/docs-site/src/content/docs/guides/combos.md
+++ b/docs-site/src/content/docs/guides/combos.md
@@ -83,7 +83,10 @@ The request then follows normal combo selection and failover.
 
 Explicit provider/combo selectors and configured combo aliases take precedence over this recall.
 Failed, incomplete, or cancelled responses do not replace the last successful selection. Recall is
-process-local and bounded to 256 lanes for 30 minutes; it does not store account credentials.
+process-local and bounded to 256 conversations for 30 minutes, and to 1 KiB per remembered model
+name and 64 KiB in total; expired entries are also cleaned up in the background. A response whose
+model name is too large to retain leaves the previous selection untouched rather than clearing it.
+Recall does not store account credentials.
 Without usable conversation identity or valid remembered state, normal compaction routing applies.
 A restart clears the remembered state.
 
diff --git a/docs-site/src/content/docs/ko/guides/combos.md b/docs-site/src/content/docs/ko/guides/combos.md
index b8d4087431..238127c1ce 100644
--- a/docs-site/src/content/docs/ko/guides/combos.md
+++ b/docs-site/src/content/docs/ko/guides/combos.md
@@ -66,7 +66,7 @@ alias는 클라이언트가 요청하는 공개 이름만 바꿉니다. 콤보
 
 클라이언트가 콤보를 바꾼 뒤 공급자 접두사 없는 모델 이름으로 압축을 요청하면, opencodex는 같은 대화에서 가장 최근에 응답을 성공적으로 마친 콤보를 기억해 사용할 수 있습니다. 모델 이름이 완료된 응답과 일치하고, 현재 설정에 해당 콤보와 대상이 남아 있어야 합니다. 압축 요청도 일반 콤보 선택과 페일오버를 따릅니다.
 
-명시한 공급자·콤보 선택자와 설정된 콤보 별칭이 기억한 값보다 우선합니다. 실패·미완료·취소된 응답은 마지막 성공 기록을 덮어쓰지 않습니다. 기록은 프로세스 안에서 최대 256개 대화, 30분 동안 유지하며 계정 자격증명을 저장하지 않습니다. 유효한 대화 식별자나 기록이 없으면 일반 압축 라우팅을 사용합니다. 재시작하면 기록은 사라집니다.
+명시한 공급자·콤보 선택자와 설정된 콤보 별칭이 기억한 값보다 우선합니다. 실패·미완료·취소된 응답은 마지막 성공 기록을 덮어쓰지 않습니다. 기록은 프로세스 안에서 최대 256개 대화, 30분 동안 유지하고, 모델 이름 하나당 1 KiB·전체 64 KiB로 제한하며, 만료된 기록은 배경에서도 정리합니다. 모델 이름이 너무 커서 보관할 수 없는 응답은 이전 선택을 지우지 않고 그대로 둡니다. 기록은 계정 자격증명을 저장하지 않습니다. 유효한 대화 식별자나 기록이 없으면 일반 압축 라우팅을 사용합니다. 재시작하면 기록은 사라집니다.
 
 ## 전략 선택
 
diff --git a/src/lib/state-store-registrations.ts b/src/lib/state-store-registrations.ts
index 13a22bfce0..849f145848 100644
--- a/src/lib/state-store-registrations.ts
+++ b/src/lib/state-store-registrations.ts
@@ -21,7 +21,7 @@ import {
 } from "../combos/failover";
 import { reconcileComboWarningMemos } from "../combos/request";
 import { reconcileComboRotationState } from "../combos/resolve";
-import { reconcileComboRecall } from "../server/responses/combo-session-recall";
+import { reconcileComboRecall, sweepExpiredComboRecall } from "../server/responses/combo-session-recall";
 import { listLiveComboTargetKeys } from "../combos/types";
 import {
   listLiveConfigOwnershipRoots,
@@ -112,7 +112,11 @@ export const STATE_STORE_REGISTRATIONS = [
   { name: "model-cache-history", reconcileGeneration: reconcileModelCacheGeneration },
   { name: "pool-rotation", reconcileGeneration: reconcilePoolRotationState },
   { name: "combo-rotation", reconcileGeneration: reconcileComboRotationState },
-  { name: "combo-session-recall", reconcileGeneration: reconcileComboRecall },
+  {
+    name: "combo-session-recall",
+    sweepExpired: sweepExpiredComboRecall,
+    reconcileGeneration: reconcileComboRecall,
+  },
   { name: "guardian-backoff", reconcileGeneration: reconcileGuardianBackoff },
   { name: "codex-reauth", reconcileGeneration: reconcileCodexReauthState },
   { name: "oauth-reauth", reconcileGeneration: reconcileOAuthReauthState },
diff --git a/src/server/responses/combo-session-recall.ts b/src/server/responses/combo-session-recall.ts
index 84dfd8d466..b3af31f1c4 100644
--- a/src/server/responses/combo-session-recall.ts
+++ b/src/server/responses/combo-session-recall.ts
@@ -8,14 +8,46 @@ interface ComboRecallEntry {
   target: Pick<OcxComboTarget, "provider" | "model">;
   responseModel: string;
   at: number;
+  /** UTF-8 size of `responseModel`, the only client-influenced field of unbounded length. */
+  bytes: number;
 }
 
 const RECALL_CAPACITY = 256;
 const RECALL_TTL_MS = 30 * 60 * 1000;
+/**
+ * A model id is provider-reported and arrives on the response, so nothing upstream of here
+ * bounds its length. Lane keys are already SHA-256 digests, so the model string is the only
+ * field that can grow, and 256 lanes alone do not bound the bytes they hold.
+ */
+const RECALL_MODEL_BYTES_MAX = 1024;
+const RECALL_TOTAL_BYTES_MAX = 64 * 1024;
 const recall = new Map<string, ComboRecallEntry>();
+let recallBytes = 0;
 let lastReconciledGeneration = 0;
 let liveOwners: Pick<GenerationContext, "comboIds" | "comboTargets" | "providerNames"> | undefined;
 
+/** Every removal path goes through here so the byte counter can never drift from the map. */
+function deleteEntry(lane: string): boolean {
+  const entry = recall.get(lane);
+  if (!entry) return false;
+  recall.delete(lane);
+  recallBytes -= entry.bytes;
+  return true;
+}
+
+/**
+ * UTF-8 size of a remembered model id, or null when it is too large to retain.
+ *
+ * The code-unit test runs first and is the part that matters: a UTF-8 encoding is never smaller
+ * than the code-unit count, so an oversized string is rejected without encoding it, and the
+ * bound cannot be defeated by paying the allocation it exists to prevent.
+ */
+function boundedModelBytes(responseModel: string): number | null {
+  if (responseModel.length > RECALL_MODEL_BYTES_MAX) return null;
+  const bytes = Buffer.byteLength(responseModel, "utf8");
+  return bytes > RECALL_MODEL_BYTES_MAX ? null : bytes;
+}
+
 function ownsEntry(context: Pick<GenerationContext, "comboIds" | "comboTargets" | "providerNames">, entry: ComboRecallEntry): boolean {
   return context.comboIds.has(entry.comboId)
     && context.providerNames.has(entry.target.provider)
@@ -32,14 +64,29 @@ export function rememberComboForLane(
   if (!lane || !comboId || !responseModel.trim()) return;
   // Reject even a same-named recreated owner: its previous in-flight turn is obsolete.
   if (writerGeneration < Math.max(lastReconciledGeneration, captureConfigGeneration())) return;
-  const entry = { comboId, target: { provider: target.provider, model: target.model }, responseModel, at: Date.now() };
+  // An unretainable model id DECLINES the write; it must not clear the lane. Every other
+  // rejection above returns the same way, and clearing here would let a late completion erase
+  // a newer selection that this function has no ordering information to compare against.
+  const bytes = boundedModelBytes(responseModel);
+  if (bytes === null) return;
+  const entry = {
+    comboId,
+    target: { provider: target.provider, model: target.model },
+    responseModel,
+    at: Date.now(),
+    bytes,
+  };
   if (liveOwners && !ownsEntry(liveOwners, entry)) return;
-  recall.delete(lane);
+  deleteEntry(lane);
   recall.set(lane, entry);
-  while (recall.size > RECALL_CAPACITY) {
+  recallBytes += bytes;
+  // Insertion order is recency order, because every write re-inserts its lane at the back.
+  // Evicting from the front therefore drops the least recently written lane, never this one:
+  // a single entry is capped well below the aggregate budget, so it always fits.
+  while (recall.size > RECALL_CAPACITY || recallBytes > RECALL_TOTAL_BYTES_MAX) {
     const oldest = recall.keys().next().value;
-    if (oldest === undefined) break;
-    recall.delete(oldest);
+    if (oldest === undefined || oldest === lane) break;
+    deleteEntry(oldest);
   }
 }
 
@@ -57,12 +104,25 @@ export function recallComboForLane(
     || !Object.hasOwn(config.providers, entry.target.provider)
     || !provider || provider.disabled === true
     || !combo?.targets.some(target => targetKey(target) === targetKey(entry.target))) {
-    recall.delete(lane);
+    deleteEntry(lane);
     return undefined;
   }
   return entry.responseModel === model ? entry.comboId : undefined;
 }
 
+/**
+ * Periodic expiry. Without it a lane that is never read again and never touched by a config
+ * reconciliation holds its entry for the life of the process: the existing TTL is only
+ * evaluated on read or on generation change.
+ */
+export function sweepExpiredComboRecall(now: number): number {
+  let removed = 0;
+  for (const [lane, entry] of recall) {
+    if (now - entry.at >= RECALL_TTL_MS && deleteEntry(lane)) removed += 1;
+  }
+  return removed;
+}
+
 export function reconcileComboRecall(context: GenerationContext): number {
   if (context.generation <= lastReconciledGeneration) return 0;
   lastReconciledGeneration = context.generation;
@@ -74,8 +134,7 @@ export function reconcileComboRecall(context: GenerationContext): number {
   let removed = 0;
   for (const [lane, entry] of recall) {
     if (!ownsEntry(context, entry) || Date.now() - entry.at >= RECALL_TTL_MS) {
-      recall.delete(lane);
-      removed += 1;
+      if (deleteEntry(lane)) removed += 1;
     }
   }
   return removed;
@@ -84,6 +143,7 @@ export function reconcileComboRecall(context: GenerationContext): number {
 /** Test-only reset, alongside the combo rotation/cooldown resets. */
 export function clearComboRecallForTests(): void {
   recall.clear();
+  recallBytes = 0;
   lastReconciledGeneration = 0;
   liveOwners = undefined;
 }
diff --git a/src/server/responses/input-admission.ts b/src/server/responses/input-admission.ts
index ef8b81265e..ca11fd60eb 100644
--- a/src/server/responses/input-admission.ts
+++ b/src/server/responses/input-admission.ts
@@ -10,7 +10,13 @@
  * catches the pathological case and stays out of the way otherwise. Every uncertainty
  * resolves toward admitting.
  */
-import { nativeOpenAiContextWindow, nativeOpenAiMaxInputTokens, type NativeContextLimitsInput } from "../../codex/catalog/metadata";
+import {
+  nativeOpenAiContextWindow,
+  nativeOpenAiMaxInputTokens,
+  nativeOpenAiMaxOutputTokens,
+  type NativeContextLimitsInput,
+} from "../../codex/catalog/metadata";
+import { getModelMetadata } from "../../generated/model-metadata";
 import { estimateTokens } from "../../lib/token-estimate";
 import { isCanonicalOpenAiForwardProvider, OPENAI_CODEX_PROVIDER_ID } from "../../providers/openai-tiers";
 import { modelRecordValue } from "../../reasoning-effort";
@@ -54,6 +60,8 @@ export interface InputAdmissionResult {
   estimatedTokens: number;
   /** Resolved ceiling, or null when nothing could be resolved (=> always admitted). */
   ceiling: number | null;
+  /** Output space reserved by the combo preflight; absent on the loose direct gate. */
+  requiredOutputHeadroom?: number;
 }
 
 function positive(value: unknown): number | null {
@@ -135,14 +143,19 @@ export function estimateInputTokens(parsed: OcxParsedRequest, modelId: string):
  * reject a user-defined provider that merely shares a built-in name using limits that
  * belong to a different service.
  */
-export function resolveInputCeiling(
+interface ResolvedContextLimits {
+  /** The target's total context window: input and output share it. */
+  window: number | null;
+  /** Largest admissible input, which input-only caps may tighten below the window. */
+  ceiling: number | null;
+}
+
+function resolveContextLimits(
   provider: OcxProviderConfig,
   providerName: string,
   modelId: string,
-  // Operator cap for the canonical native provider. Passed in rather than read from a
-  // config here so this stays pure: no filesystem, no catalog, no registry scan.
   nativeContextCap?: NativeContextLimitsInput,
-): number | null {
+): ResolvedContextLimits {
   // `modelRecordValue`, not a bare lookup: the catalog resolves these same two maps that
   // way, so a `gpt-oss` entry covers `gpt-oss:120b`. Reading raw here made the gate fall
   // back to the provider-wide window and refuse turns the model can plainly hold.
@@ -168,13 +181,120 @@ export function resolveInputCeiling(
     : null;
   const nativeMaxInput = canonicalNativeBare ? positive(nativeOpenAiMaxInputTokens(modelId, nativeLimits)) : null;
 
-  const window = canonicalNativeBare ? native : configured;
+  const window = canonicalNativeBare ? (native ?? generatedNativeWindow(modelId, configured, nativeContextCap)) : configured;
   // modelMaxInputTokens is an input-only cap, so it can only tighten the window.
   const configuredMaxInput = positive(modelRecordValue(provider.modelMaxInputTokens, modelId));
   const limits = [window, configuredMaxInput, nativeMaxInput].filter((v): v is number => v !== null);
+  return { window, ceiling: limits.length === 0 ? null : Math.min(...limits) };
+}
+
+/**
+ * Generated-catalog keys, not routing provider names. `OPENAI_CODEX_PROVIDER_ID` is the string
+ * `"openai"` -- the canonical Codex forward route -- so using it to index the generated bundle
+ * would silently skip the native Codex rows and read the public API rows instead.
+ */
+const NATIVE_METADATA_CATALOGS = ["openai-codex", "openai"] as const;
+
+/**
+ * Static in-tree metadata for a canonical native slug the narrower override and pinned-native
+ * tables do not carry. Falling through to null made input admission completely blind for
+ * exactly those models, which is how a 128k target accepted a turn it could not finish.
+ *
+ * This deliberately covers slugs that are no longer offered in the picker: a retired slug is
+ * still dispatchable when an operator names it explicitly in a combo target, and that is the
+ * configuration where the gate was inert. This is a generated bundle compiled into the binary,
+ * not a live catalog read, so it adds no I/O. Explicit provider and operator caps may only
+ * narrow the result, never widen it.
+ */
+function generatedNativeWindow(
+  modelId: string,
+  configured: number | null,
+  nativeContextCap: NativeContextLimitsInput | undefined,
+): number | null {
+  let generated: number | null = null;
+  for (const catalog of NATIVE_METADATA_CATALOGS) {
+    generated = positive(getModelMetadata(catalog, modelId)?.contextWindow);
+    if (generated !== null) break;
+  }
+  if (generated === null) return null;
+  const cap = typeof nativeContextCap === "number"
+    ? positive(nativeContextCap)
+    : positive(nativeContextCap?.cap);
+  return Math.min(generated, configured ?? generated, cap ?? generated);
+}
+
+export function resolveInputCeiling(
+  provider: OcxProviderConfig,
+  providerName: string,
+  modelId: string,
+  // Operator cap for the canonical native provider. Passed in rather than read from a
+  // config here so this stays pure: no filesystem, no catalog, no registry scan.
+  nativeContextCap?: NativeContextLimitsInput,
+): number | null {
+  return resolveContextLimits(provider, providerName, modelId, nativeContextCap).ceiling;
+}
+
+/**
+ * Largest output the concrete target can emit. Used only to avoid reserving MORE than the
+ * target could ever produce when a client asks for a bigger allowance than the model has.
+ * Unknown stays unknown rather than inventing a capability.
+ */
+export function resolveOutputCeiling(
+  provider: OcxProviderConfig,
+  providerName: string,
+  modelId: string,
+): number | null {
+  const configured = positive(modelRecordValue(provider.modelMaxOutputTokens, modelId))
+    ?? positive(provider.defaultMaxOutputTokens);
+  const canonicalNativeBare = providerName === OPENAI_CODEX_PROVIDER_ID
+    && isCanonicalOpenAiForwardProvider(provider)
+    && !modelId.includes("/");
+  const native = canonicalNativeBare ? positive(nativeOpenAiMaxOutputTokens(modelId)) : null;
+  const limits = [configured, native].filter((v): v is number => v !== null);
   return limits.length === 0 ? null : Math.min(...limits);
 }
 
+/**
+ * Combo-only admission. A fallback must be able to satisfy the caller's declared output
+ * allowance inside its OWN context window. Otherwise it returns 200, emits a few hundred
+ * tokens, and terminates on `finish_reason: length` — which the Anthropic surface renders as
+ * "response exceeded the output token maximum" even though the real cause was the total
+ * window. By then the next target cannot be tried, because output has already committed.
+ *
+ * Two budgets are checked separately so the reserve is counted exactly once. `ceiling` is an
+ * input-only budget once `modelMaxInputTokens` tightens it below the window, so the output
+ * reserve belongs against `window`, not against `ceiling`.
+ *
+ * Direct and single-target requests keep the deliberately loose 2.5x pathological-input gate.
+ * This stricter rule applies only to synthetic combo children, where skipping one known-small
+ * target is safe and the ladder continues before any upstream bytes are sent. Unknown context
+ * stays fail-open, and a caller that declared no output allowance is unaffected.
+ */
+export function checkComboTargetInputAdmission(
+  parsed: OcxParsedRequest,
+  provider: OcxProviderConfig,
+  providerName: string,
+  modelId: string,
+  nativeContextCap?: NativeContextLimitsInput,
+): InputAdmissionResult {
+  const { window, ceiling } = resolveContextLimits(provider, providerName, modelId, nativeContextCap);
+  const requestedOutput = positive(parsed.options.maxOutputTokens);
+  if (window === null || ceiling === null || requestedOutput === null) {
+    return checkInputAdmission(parsed, provider, providerName, modelId, nativeContextCap);
+  }
+  const targetOutput = resolveOutputCeiling(provider, providerName, modelId);
+  const requiredOutputHeadroom = targetOutput === null
+    ? requestedOutput
+    : Math.min(requestedOutput, targetOutput);
+  const estimatedTokens = estimateInputTokens(parsed, modelId);
+  return {
+    admitted: estimatedTokens <= ceiling && estimatedTokens + requiredOutputHeadroom <= window,
+    estimatedTokens,
+    ceiling,
+    requiredOutputHeadroom,
+  };
+}
+
 /**
  * Fail-open when no ceiling is known; refuse only past `ceiling * ADMISSION_TOLERANCE`.
  *
diff --git a/src/server/responses/request-prepare.ts b/src/server/responses/request-prepare.ts
index 81ffeb013f..b655d8f16d 100644
--- a/src/server/responses/request-prepare.ts
+++ b/src/server/responses/request-prepare.ts
@@ -101,7 +101,7 @@ import {
   isCodexReserveHelperUnsupported,
   CODEX_RESERVE_HELPER_UNSUPPORTED_MESSAGE,
 } from "../../codex/loopback-target";
-import { checkInputAdmission } from "./input-admission";
+import { checkComboTargetInputAdmission, checkInputAdmission } from "./input-admission";
 import { nativeContextLimits } from "../../codex/catalog";
 import { streamingContextOverflowResponse } from "./context-overflow";
 import {
@@ -860,7 +860,12 @@ export async function prepareResponsesRequest(
   // refusing the turn that shrinks the context would deadlock the client against the very
   // limit this gate reports — it would be told to compact and then denied the compaction.
   if (parsed._compactionRequest !== true) {
-    const inputAdmission = checkInputAdmission(parsed, route.provider, route.providerName, parsed.modelId, nativeContextLimits(config));
+    // A combo child is the one caller that can afford a strict gate: skipping a target it
+    // cannot fit is safe before any upstream bytes are sent, and the ladder continues. A
+    // direct request has nowhere to go, so it keeps the loose pathological-input gate.
+    const inputAdmission = options.comboAttempt
+      ? checkComboTargetInputAdmission(parsed, route.provider, route.providerName, parsed.modelId, nativeContextLimits(config))
+      : checkInputAdmission(parsed, route.provider, route.providerName, parsed.modelId, nativeContextLimits(config));
     if (!inputAdmission.admitted) {
       // #1524: this is a LOCAL preflight refusal, not an upstream verdict. A policy or combo
       // fallback must be able to skip this candidate and try one whose context window fits,
@@ -876,9 +881,13 @@ export async function prepareResponsesRequest(
       return formatErrorResponse(
         413,
         "input_admission_refused",
-        `Estimated input (~${inputAdmission.estimatedTokens} tokens) is far past the context window `
-          + `of ${parsed.modelId} (${inputAdmission.ceiling} tokens). Start a new session or choose a `
-          + `model with a larger context window.`,
+        inputAdmission.requiredOutputHeadroom !== undefined
+          ? `Estimated input (~${inputAdmission.estimatedTokens} tokens) plus ${inputAdmission.requiredOutputHeadroom} `
+            + `tokens of requested output headroom cannot fit the context window of ${parsed.modelId} `
+            + `(${inputAdmission.ceiling} tokens).`
+          : `Estimated input (~${inputAdmission.estimatedTokens} tokens) is far past the context window `
+            + `of ${parsed.modelId} (${inputAdmission.ceiling} tokens). Start a new session or choose a `
+            + `model with a larger context window.`,
       );
     }
   }
diff --git a/structure/transports/responses.md b/structure/transports/responses.md
index 9af5e63100..98f518946e 100644
--- a/structure/transports/responses.md
+++ b/structure/transports/responses.md
@@ -178,6 +178,20 @@ explicit configured selectors before consulting bounded lane state. The existing
 reconciliation owns removal of obsolete targets and generation fencing; core imports no registration
 composition root or Lab code. Recall retains routing identity only, never account credentials.
 
+Retention is bounded on four axes: 256 lanes, 30 minutes, 1 KiB per remembered model id, and 64 KiB
+in aggregate. The model id is the only field of unbounded length — lane keys are already SHA-256
+digests — so the lane cap alone does not bound the bytes those lanes hold. The size test runs on code
+units before encoding, since a UTF-8 encoding is never smaller than its code-unit count and the bound
+must not pay the allocation it exists to prevent. Aggregate eviction drops the least recently written
+lane, which is the front of the map because every write re-inserts its own lane at the back.
+
+An unretainable model id declines the write rather than clearing the lane, matching how every other
+rejection in `rememberComboForLane` returns. Clearing would let a late completion erase a newer
+selection, and the publication path carries a config generation, not a request order, so it has no
+basis on which to decide that its own result is the newer one. The store is also swept periodically
+now: the TTL was previously evaluated only on read or on a generation change, so a lane never read
+again held its entry for the life of the process.
+
 > Decision record: [ADR-0038](../decisions/ADR-0038-responses-http-sse.md)
 
 A replayed compaction item carries an `encrypted_content` blob only its minting backend can decode,
@@ -702,3 +716,30 @@ What must not happen is a ladder that charges and then returns through a path th
 nor releases. That is not a lost send; it is a send the request never made, spending an allowance a
 later recovery in the same request then cannot have. `tests/lib/execution-budget-permits.test.ts`
 pins both ladder shapes against exactly that.
+
+## Combo output headroom
+
+A combo child is admitted against two budgets, not one. `resolveInputCeiling` in
+`src/server/responses/input-admission.ts` answers "how much input may this target take", which
+`modelMaxInputTokens` can tighten below the window. The context window itself is what input and
+output actually share. When the caller declared `max_output_tokens`,
+`checkComboTargetInputAdmission` requires both `estimated input <= ceiling` and
+`estimated input + min(declared output, target output ceiling) <= window`, so the output reserve
+is counted once rather than charged twice against an already-tightened input budget.
+
+The refusal is local: HTTP 413 `input_admission_refused` before any upstream bytes are sent, which
+existing combo policy already treats as a safe hop. That ordering is the whole point. A target whose
+total window cannot hold the turn plus the caller's allowance answers 200, emits a few hundred
+tokens and stops on `finish_reason: length`, which the Anthropic surface renders as an output-token
+error naming a limit the model never approached — and by then output has committed and no later
+target may be tried.
+
+Scope is deliberately narrow. Direct and single-target requests keep the loose 2.5x
+pathological-input gate, because they have nowhere to hop. Compaction turns stay exempt. Unknown
+context and a caller that declared no output allowance both remain fail-open, so this invents no
+limits for custom providers. Canonical native slugs that the narrower pinned table does not carry
+resolve their window from the generated in-tree bundle, which is what made the gate inert on the
+route where this was first observed; explicit provider and operator caps may only narrow it.
+
+Regression coverage: `tests/server/input-admission.test.ts` and
+`tests/helpers/combo-context-headroom-cases.ts`.
diff --git a/tests/helpers/combo-context-headroom-cases.ts b/tests/helpers/combo-context-headroom-cases.ts
new file mode 100644
index 0000000000..2d8769bee7
--- /dev/null
+++ b/tests/helpers/combo-context-headroom-cases.ts
@@ -0,0 +1,91 @@
+import { expect, test } from "bun:test";
+import type { OcxConfig, OcxProviderConfig } from "../../src/types";
+
+interface ComboHarness<Server> {
+  serve(handler: () => Response | Promise<Response>): Server;
+  baseUrl(server: Server): string;
+  chatSuccess(text: string, model?: string): Response;
+  provider(adapter: string, url: string, apiKey: string, extra?: Partial<OcxProviderConfig>): OcxProviderConfig;
+  comboConfig(providers: OcxConfig["providers"]): OcxConfig;
+  post(config: OcxConfig, raw?: Record<string, unknown>): Promise<Response>;
+}
+
+/** Roughly `tokens` worth of plain ASCII at the default 4 chars/token ratio. */
+function asciiTokens(tokens: number): string {
+  return "a".repeat(tokens * 4);
+}
+
+/** Register under the caller's isolated homes, mock state and server cleanup hooks. */
+export function registerComboContextHeadroomCases<Server>({
+  serve, baseUrl, chatSuccess, provider, comboConfig, post,
+}: ComboHarness<Server>): void {
+  test("a target that cannot hold input plus requested output is skipped before any bytes commit", async () => {
+    let smallHits = 0;
+    let largeHits = 0;
+    const small = serve(() => {
+      smallHits += 1;
+      return chatSuccess("MUST NOT RUN", "m1");
+    });
+    const large = serve(() => {
+      largeHits += 1;
+      return chatSuccess("large context target", "m2");
+    });
+    // ~10k input, and m1 can reach 3,200 output inside a 12,800 window, so the turn cannot
+    // finish there. m2 holds the same turn with the caller's full 6,400 allowance.
+    const response = await post(comboConfig({
+      a: provider("openai-chat", baseUrl(small), "key-a", {
+        modelContextWindows: { m1: 12_800 },
+        modelMaxOutputTokens: { m1: 3_200 },
+      }),
+      b: provider("openai-chat", baseUrl(large), "key-b", {
+        modelContextWindows: { m2: 100_000 },
+        modelMaxOutputTokens: { m2: 32_000 },
+      }),
+    }), { input: asciiTokens(10_000), max_output_tokens: 6_400 });
+    expect(response.status).toBe(200);
+    expect(smallHits).toBe(0);
+    expect(largeHits).toBe(1);
+    expect(await response.text()).toContain("large context target");
+  });
+
+  test("the same undersized target still serves a turn that declares no output allowance", async () => {
+    // The strict reserve is opt-in on the caller's declared allowance. Without one, the
+    // deliberately loose pathological-input gate still applies and nothing is skipped.
+    let smallHits = 0;
+    const small = serve(() => {
+      smallHits += 1;
+      return chatSuccess("small context target", "m1");
+    });
+    const response = await post(comboConfig({
+      a: provider("openai-chat", baseUrl(small), "key-a", {
+        modelContextWindows: { m1: 12_800 },
+        modelMaxOutputTokens: { m1: 3_200 },
+      }),
+    }), { input: asciiTokens(10_000) });
+    expect(response.status).toBe(200);
+    expect(smallHits).toBe(1);
+    expect(await response.text()).toContain("small context target");
+  });
+
+  test("provider-specific prompt-too-long 400 hops to a larger-context combo target", async () => {
+    let backupHits = 0;
+    const capped = serve(() => Response.json({ error: {
+      message: "Prompt 346030 > 262144 maximum context length",
+      type: "invalid_request_prompt_too_long",
+      code: "5059",
+      raw_status_code: 400,
+    } }, { status: 400 }));
+    const backup = serve(() => {
+      backupHits += 1;
+      return chatSuccess("larger context backup", "m2");
+    });
+    const response = await post(comboConfig({
+      a: provider("openai-chat", baseUrl(capped), "key-a"),
+      b: provider("openai-chat", baseUrl(backup), "key-b"),
+    }));
+    expect(response.status).toBe(200);
+    expect(backupHits).toBe(1);
+    expect(await response.text()).toContain("larger context backup");
+  });
+}
+
diff --git a/tests/oauth/state-store-sweeper.test.ts b/tests/oauth/state-store-sweeper.test.ts
index 5b691d28d6..4cfb17ee84 100644
--- a/tests/oauth/state-store-sweeper.test.ts
+++ b/tests/oauth/state-store-sweeper.test.ts
@@ -215,6 +215,72 @@ describe("state-store sweeper", () => {
     }
   });
 
+  describe("bounded combo recall retention", () => {
+    const config: OcxConfig = {
+      port: 0, defaultProvider: "a",
+      providers: { a: { adapter: "openai-chat", baseUrl: "https://a.example/v1" } },
+      combos: { first: { targets: [{ provider: "a", model: "m1" }] } },
+    };
+    const remember = (lane: string, responseModel: string) =>
+      rememberComboForLane(lane, "first", { provider: "a", model: "m1" }, responseModel, captureConfigGeneration());
+    /** A distinct model id of exactly 1 KiB, the largest this store will retain. */
+    const fullModel = (index: number) => `${index}-`.padEnd(1024, "m");
+
+    test("an unretainable model id declines the write instead of clearing the lane", () => {
+      remember("lane", "kept-model");
+      // A model id is provider-reported and arrives on the response, so its length is not
+      // bounded upstream of here. Refusing to retain it must not also destroy what is there:
+      // this callback carries a config generation, not a request order, so it cannot know its
+      // own result is newer than the entry it would be erasing.
+      remember("lane", "x".repeat(1025));
+      expect(recallComboForLane(config, "lane", "kept-model")).toBe("first");
+
+      // Measured in UTF-8 bytes, not code units: 600 three-byte characters is 1,800 bytes.
+      remember("lane", "가".repeat(600));
+      expect(recallComboForLane(config, "lane", "kept-model")).toBe("first");
+
+      // And an oversized id never establishes a lane of its own.
+      remember("fresh", "x".repeat(4096));
+      expect(recallComboForLane(config, "fresh", "x".repeat(4096))).toBeUndefined();
+    });
+
+    test("the aggregate byte budget evicts the least recently written lane", () => {
+      // 64 KiB holds exactly 64 maximum-size entries, well inside the 256-lane cap, so this
+      // isolates the byte budget from the lane count.
+      for (let i = 0; i < 64; i += 1) remember(`lane-${i}`, fullModel(i));
+      expect(recallComboForLane(config, "lane-0", fullModel(0))).toBe("first");
+
+      remember("lane-64", fullModel(64));
+      expect(recallComboForLane(config, "lane-0", fullModel(0))).toBeUndefined();
+      expect(recallComboForLane(config, "lane-1", fullModel(1))).toBe("first");
+      expect(recallComboForLane(config, "lane-64", fullModel(64))).toBe("first");
+    });
+
+    test("a rewritten lane is charged once, not once per write", () => {
+      // Replacing a lane must release the old entry's bytes. If it did not, 64 rewrites of one
+      // lane would exhaust the whole budget and start evicting unrelated lanes.
+      remember("stable", "stable-model");
+      for (let i = 0; i < 64; i += 1) remember("churn", fullModel(i));
+      expect(recallComboForLane(config, "stable", "stable-model")).toBe("first");
+      expect(recallComboForLane(config, "churn", fullModel(63))).toBe("first");
+    });
+
+    test("a periodic tick expires a lane that is never read again and releases its bytes", () => {
+      registerStateStore(STATE_STORE_REGISTRATIONS.find(row => row.name === "combo-session-recall")!);
+      for (let i = 0; i < 64; i += 1) remember(`stale-${i}`, fullModel(i));
+
+      // Before this the TTL was only evaluated on read or on a generation change, so a lane
+      // nobody reads again held its entry for the life of the process.
+      expect(sweepExpired(Date.now() + 30 * 60 * 1_000)).toEqual({ storesVisited: 1, rowsRemoved: 64 });
+      expect(recallComboForLane(config, "stale-0", fullModel(0))).toBeUndefined();
+
+      // The budget is genuinely free again: a full refill keeps its own oldest lane, which
+      // could not happen if the swept entries had left their bytes behind.
+      for (let i = 0; i < 64; i += 1) remember(`fresh-${i}`, fullModel(i));
+      expect(recallComboForLane(config, "fresh-0", fullModel(0))).toBe("first");
+    });
+  });
+
   test("a sweeper tick expires continuation and Antigravity rows without store traffic", () => {
     rememberResponseState({ input: "old" }, { id: "resp_sweeper_ttl", output: [], status: "completed" });
     observeAntigravityReplay("gemini-3-pro", "session-old", [{
diff --git a/tests/server/input-admission.test.ts b/tests/server/input-admission.test.ts
index 9b8ac9c978..90afdbe92b 100644
--- a/tests/server/input-admission.test.ts
+++ b/tests/server/input-admission.test.ts
@@ -1,9 +1,11 @@
 import { describe, expect, test } from "bun:test";
 import {
   ADMISSION_TOLERANCE,
+  checkComboTargetInputAdmission,
   checkInputAdmission,
   estimateInputTokens,
   resolveInputCeiling,
+  resolveOutputCeiling,
 } from "../../src/server/responses/input-admission";
 import { modelRecordValue } from "../../src/reasoning-effort";
 import type { OcxMessage, OcxParsedRequest, OcxProviderConfig, OcxTool } from "../../src/types";
@@ -257,3 +259,86 @@ describe("checkInputAdmission", () => {
     expect(calls).toBe(0);
   });
 });
+
+describe("combo target input admission", () => {
+  const capped: OcxProviderConfig = {
+    adapter: "openai-chat",
+    baseUrl: "https://example.test/v1",
+    modelContextWindows: { m: 128_000 },
+    modelMaxOutputTokens: { m: 32_000 },
+  };
+
+  const withMaxOutput = (inputTokens: number, maxOutputTokens = 64_000): OcxParsedRequest => ({
+    ...request([userText(asciiTokens(inputTokens))]),
+    modelId: "m",
+    options: { maxOutputTokens },
+  });
+  // A separate builder, because passing `undefined` to the one above would silently take its
+  // default and the row below would assert the opposite of what it claims to cover.
+  const withoutMaxOutput = (inputTokens: number): OcxParsedRequest => ({
+    ...request([userText(asciiTokens(inputTokens))]),
+    modelId: "m",
+    options: {},
+  });
+
+  test("skips a target that cannot hold the turn plus its own output ceiling", () => {
+    // 100k input + 32k of reachable output does not fit 128k, so this target would have
+    // answered 200, emitted a few hundred tokens and stopped on finish_reason: length.
+    const result = checkComboTargetInputAdmission(withMaxOutput(100_000), capped, "custom", "m");
+    expect(result.admitted).toBe(false);
+    expect(result.ceiling).toBe(128_000);
+    expect(result.requiredOutputHeadroom).toBe(32_000);
+  });
+
+  test("reserves no more than the target can actually emit", () => {
+    // The caller asked for 64k, but this model tops out at 32k, so reserving the caller's
+    // number would skip a target that fits.
+    const result = checkComboTargetInputAdmission(withMaxOutput(90_000), capped, "custom", "m");
+    expect(result.admitted).toBe(true);
+    expect(result.requiredOutputHeadroom).toBe(32_000);
+  });
+
+  test("an input-only cap is not charged the output reserve twice", () => {
+    // modelMaxInputTokens tightens the admissible INPUT; the output reserve belongs against
+    // the window. Charging both against the tightened number would refuse a turn that fits.
+    const inputCapped: OcxProviderConfig = { ...capped, modelMaxInputTokens: { m: 90_000 } };
+    const fits = checkComboTargetInputAdmission(withMaxOutput(85_000), inputCapped, "custom", "m");
+    expect(fits.admitted).toBe(true);
+    expect(fits.ceiling).toBe(90_000);
+    // The input cap itself still refuses on its own terms.
+    expect(checkComboTargetInputAdmission(withMaxOutput(95_000), inputCapped, "custom", "m").admitted).toBe(false);
+  });
+
+  test("unknown context stays fail-open", () => {
+    const unknown: OcxProviderConfig = { adapter: "openai-chat", baseUrl: "https://example.test/v1" };
+    const result = checkComboTargetInputAdmission(withMaxOutput(2_000_000), unknown, "custom", "m");
+    expect(result.admitted).toBe(true);
+    expect(result.ceiling).toBeNull();
+  });
+
+  test("no declared output allowance keeps the loose direct contract", () => {
+    const result = checkComboTargetInputAdmission(withoutMaxOutput(150_000), capped, "custom", "m");
+    expect(result.admitted).toBe(true); // still inside the existing 2.5x pathological gate
+    expect(result.requiredOutputHeadroom).toBeUndefined();
+  });
+
+  test("a canonical native slug missing from the override table resolves from generated metadata", () => {
+    // Spark carries 128k/32k in the generated bundle but is absent from the narrower pinned
+    // native table, which left the gate completely blind on exactly this route. It is retired
+    // from the picker and still dispatchable when an operator names it in a combo target.
+    expect(resolveInputCeiling(CANONICAL_NATIVE, "openai", "gpt-5.3-codex-spark")).toBe(128_000);
+    expect(resolveOutputCeiling(CANONICAL_NATIVE, "openai", "gpt-5.3-codex-spark")).toBe(32_000);
+    // The native Codex catalog is consulted first, and it is keyed "openai-codex" — which is NOT
+    // the routing provider id, because that one is the string "openai". `gpt-5-codex-mini` exists
+    // only in the native catalog, so resolving it proves the right key is being read.
+    expect(resolveInputCeiling(CANONICAL_NATIVE, "openai", "gpt-5-codex-mini")).toBe(272_000);
+    // A slug the override table does know keeps its own pinned window.
+    expect(resolveInputCeiling(CANONICAL_NATIVE, "openai", "gpt-5.6-sol")).toBe(272_000);
+    // An operator cap may only narrow the generated value, never widen it.
+    expect(resolveInputCeiling(CANONICAL_NATIVE, "openai", "gpt-5.3-codex-spark", 64_000)).toBe(64_000);
+    // A provider merely named openai still inherits nothing.
+    const impostor: OcxProviderConfig = { adapter: "openai-chat", baseUrl: "https://impostor.test/v1", authMode: "key" };
+    expect(resolveInputCeiling(impostor, "openai", "gpt-5.3-codex-spark")).toBeNull();
+    expect(resolveOutputCeiling(impostor, "openai", "gpt-5.3-codex-spark")).toBeNull();
+  });
+});
diff --git a/tests/server/server-combo-failover-e2e.test.ts b/tests/server/server-combo-failover-e2e.test.ts
index 78d288f069..f636e3480f 100644
--- a/tests/server/server-combo-failover-e2e.test.ts
+++ b/tests/server/server-combo-failover-e2e.test.ts
@@ -1,5 +1,6 @@
 import { registerComboForcedEffortCases } from "../helpers/combo-forced-effort-cases";
 import { registerComboContextOverflowCases } from "../helpers/combo-context-overflow-cases";
+import { registerComboContextHeadroomCases } from "../helpers/combo-context-headroom-cases";
 import { sessionLaneIdFromRequest } from "../../src/server/request-log-conversation";
 import { afterEach, beforeEach, describe, expect, mock, setDefaultTimeout, test } from "bun:test";
 import { logsFromApiBody } from "../helpers/logs-api";
@@ -2091,26 +2092,7 @@ describe("server combo failover 030 activation matrix", () => {
     serve, baseUrl, chatSuccess, chatStream, provider, comboConfig, post, collectSse,
   });
 
-  test("provider-specific prompt-too-long 400 hops to a larger-context combo target", async () => {
-    let backupHits = 0;
-    const capped = serve(() => Response.json({ error: {
-      message: "Prompt 346030 > 262144 maximum context length",
-      type: "invalid_request_prompt_too_long",
-      code: "5059",
-      raw_status_code: 400,
-    } }, { status: 400 }));
-    const backup = serve(() => {
-      backupHits += 1;
-      return chatSuccess("larger context backup", "m2");
-    });
-    const response = await post(comboConfig({
-      a: provider("openai-chat", baseUrl(capped), "key-a"),
-      b: provider("openai-chat", baseUrl(backup), "key-b"),
-    }));
-    expect(response.status).toBe(200);
-    expect(backupHits).toBe(1);
-    expect(await response.text()).toContain("larger context backup");
-  });
+  registerComboContextHeadroomCases({ serve, baseUrl, chatSuccess, provider, comboConfig, post });
 
   test("429 Retry-After 120 keeps A cooling at 60 seconds and restores it at 120", async () => {
     const t0 = Date.parse("2026-07-18T00:00:00.000Z");