diff --git a/packages/catalog/src/model-manager.ts b/packages/catalog/src/model-manager.ts
index ca7230aac..4c48908c8 100644
--- a/packages/catalog/src/model-manager.ts
+++ b/packages/catalog/src/model-manager.ts
@@ -39,7 +39,7 @@ export interface ModelManagerOptions<TApi extends Api = Api, TModelsDevPayload =
 	cacheTtlMs?: number;
 	/** When true, a successful dynamic fetch is the complete provider catalog and prunes static-only models. */
 	dynamicModelsAuthoritative?: boolean;
-	/** Cached model ids to ignore when the cache was written against a different static catalog fingerprint. */
+	/** Cached model ids whose presence forces refresh when the static or migration-policy fingerprint changes. */
 	dropCachedModelIdsOnStaticMismatch?: readonly string[];
 	/**
 	 * Trusted, provider-wide request headers (compile-time constants, never
@@ -199,10 +199,24 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 	const usableCachedModels = restoredCache.models.filter(model => !restoredCache.unresolvedModelIds.has(model.id));
 	const cacheHasUnresolvedHeaders = restoredCache.unresolvedModelIds.size > 0;
 	const dynamicModelsAuthoritative = options.dynamicModelsAuthoritative ?? false;
-	const staticFingerprint = fingerprintStatic(staticModels, dynamicModelsAuthoritative);
+	const cacheDropIds = options.dropCachedModelIdsOnStaticMismatch;
+	const staticCatalogFingerprint = fingerprintStatic(staticModels, dynamicModelsAuthoritative);
+	// Endpoint-migration policy is cache identity: adding an id must invalidate
+	// matching-static-catalog caches written by the prior resolver.
+	const staticFingerprint =
+		cacheDropIds && cacheDropIds.length > 0
+			? `${staticCatalogFingerprint}:drop:${Bun.hash(cacheDropIds.join("\0")).toString(36)}`
+			: staticCatalogFingerprint;
 	const cacheFingerprintMatches = cache?.staticFingerprint === staticFingerprint && staticFingerprint.length > 0;
+	const cacheNeedsModelMigration =
+		!cacheFingerprintMatches &&
+		cacheDropIds !== undefined &&
+		usableCachedModels.some(model => cacheDropIds.includes(model.id));
 	const hasUsableFreshCache =
-		(cache?.fresh ?? false) && !cacheHasUnresolvedHeaders && (!dynamicModelsAuthoritative || cacheFingerprintMatches);
+		(cache?.fresh ?? false) &&
+		!cacheHasUnresolvedHeaders &&
+		!cacheNeedsModelMigration &&
+		(!dynamicModelsAuthoritative || cacheFingerprintMatches);
 	const dynamicFetcher = options.fetchDynamicModels;
 	const hasDynamicFetcher = typeof dynamicFetcher === "function";
 	const hasAuthoritativeCache = ((cache?.authoritative ?? false) && hasUsableFreshCache) || !hasDynamicFetcher;
diff --git a/packages/catalog/test/github-copilot-model-limits.test.ts b/packages/catalog/test/github-copilot-model-limits.test.ts
index 992258afe..f0cccf5b4 100644
--- a/packages/catalog/test/github-copilot-model-limits.test.ts
+++ b/packages/catalog/test/github-copilot-model-limits.test.ts
@@ -334,10 +334,10 @@ describe("github copilot model limits mapping", () => {
 		expect(model?.api).toBe("openai-responses");
 	});
 	for (const migration of [
-		{ id: "mai-code-1-flash-picker", name: "MAI-Code-1-Flash", expectedApi: "openai-responses" },
-		{ id: "grok-4.5", name: "Grok 4.5", expectedApi: undefined },
+		{ id: "mai-code-1-flash-picker", name: "MAI-Code-1-Flash" },
+		{ id: "grok-4.5", name: "Grok 4.5" },
 	]) {
-		it(`invalidates a cached ${migration.name} completion route after the endpoint migration`, async () => {
+		it(`refreshes a cached ${migration.name} completion route after the endpoint migration`, async () => {
 			const tempDir = await fs.mkdtemp(path.join(os.tmpdir(), `pi-ai-copilot-${migration.id}-cache-`));
 			const cacheDbPath = path.join(tempDir, "models.db");
 			const cacheProviderId = `github-copilot-${migration.id}-cache-test`;
@@ -346,7 +346,6 @@ describe("github copilot model limits mapping", () => {
 					providerId: "github-copilot",
 					cacheProviderId,
 					cacheDbPath,
-					staticModels: [],
 					fetchDynamicModels: async () => [
 						{
 							id: migration.id,
@@ -365,7 +364,15 @@ describe("github copilot model limits mapping", () => {
 				await oldManager.refresh("online");
 
 				const fetchMock = vi.fn(async () => {
-					throw new Error("a fresh cache must avoid discovery");
+					return new Response(
+						JSON.stringify({
+							data: [{ id: migration.id, name: migration.name }],
+						}),
+						{
+							status: 200,
+							headers: { "Content-Type": "application/json" },
+						},
+					);
 				});
 				const manager = createModelManager({
 					...githubCopilotModelManagerOptions({ apiKey: "copilot-test-key", fetch: fetchMock }),
@@ -375,8 +382,8 @@ describe("github copilot model limits mapping", () => {
 				const { models } = await manager.refresh("online-if-uncached");
 				const model = models.find(candidate => candidate.id === migration.id);
 
-				expect(fetchMock).not.toHaveBeenCalled();
-				expect(model?.api).toBe(migration.expectedApi);
+				expect(fetchMock).toHaveBeenCalledTimes(1);
+				expect(model?.api).toBe("openai-responses");
 			} finally {
 				await fs.rm(tempDir, { recursive: true, force: true });
 			}