diff --git a/PanTS-Demo/src/components/segmentation/RefineFlyout.tsx b/PanTS-Demo/src/components/segmentation/RefineFlyout.tsx
index 48944ea8..2b505d39 100644
--- a/PanTS-Demo/src/components/segmentation/RefineFlyout.tsx
+++ b/PanTS-Demo/src/components/segmentation/RefineFlyout.tsx
@@ -21,9 +21,20 @@ const FRIENDLY_ERRORS = new Set([
 	"The interactive model is busy. Wait a moment and try again.",
 ]);
 
+const BUSY_MESSAGE = "The interactive model is busy. Wait a moment and try again.";
+const UNAVAILABLE_MESSAGE = "The interactive model is unavailable right now. Try again in a moment.";
+
+// The request layer words a full model server as "Server is busy (too many
+// requests)..." (HTTP 429) and a stalled or down one as "Server error (5xx)...",
+// neither of which is in FRIENDLY_ERRORS above, so both used to surface as a
+// generic "Refining failed" with no hint that waiting would help.
 function refineErrorMessage(error: unknown): string {
 	if (error instanceof TypeError) return "Could not reach the model server. Check your connection and try again.";
-	if (error instanceof Error && FRIENDLY_ERRORS.has(error.message)) return error.message;
+	if (error instanceof Error) {
+		if (FRIENDLY_ERRORS.has(error.message)) return error.message;
+		if (/\b(busy|at capacity|too many requests)\b/i.test(error.message)) return BUSY_MESSAGE;
+		if (/^Server error \(5\d\d\)/.test(error.message)) return UNAVAILABLE_MESSAGE;
+	}
 	return "Refining failed. Try again.";
 }
 
diff --git a/PanTS-Demo/src/helpers/viewer/useInteractivePromptTool.test.tsx b/PanTS-Demo/src/helpers/viewer/useInteractivePromptTool.test.tsx
index 2d972e54..578e4bfd 100644
--- a/PanTS-Demo/src/helpers/viewer/useInteractivePromptTool.test.tsx
+++ b/PanTS-Demo/src/helpers/viewer/useInteractivePromptTool.test.tsx
@@ -203,6 +203,26 @@ describe("interactive prompt resource changes", () => {
 		);
 	});
 
+	it("releases the model session when the page is closed, and only once when it unmounts afterwards", async () => {
+		promptApi.releaseInteractivePromptSession.mockClear();
+		const { unmount } = renderHook(() => useInteractivePromptTool({
+			enabled: true, mode: "point", apiBase: "http://api.test", caseId: "case-a",
+			activeSegmentIndex: 12, res: "full",
+		}));
+
+		// Closing a tab never unmounts the component, so React cleanup alone would leave
+		// the model server's only lease held until the idle timeout.
+		window.dispatchEvent(new Event("pagehide"));
+		expect(promptApi.releaseInteractivePromptSession).toHaveBeenCalledTimes(1);
+		expect(promptApi.releaseInteractivePromptSession).toHaveBeenCalledWith("http://api.test", "case-a", 12, "full");
+
+		unmount();
+		// The unmount still releases (the real helper finds no token the second time and does nothing).
+		expect(promptApi.releaseInteractivePromptSession).toHaveBeenCalledTimes(2);
+		window.dispatchEvent(new Event("pagehide"));
+		expect(promptApi.releaseInteractivePromptSession).toHaveBeenCalledTimes(2);
+	});
+
 	it("identifies a region-grow fallback as not produced by nnInteractive", async () => {
 		promptApi.submitInteractiveSegmentPrompt.mockResolvedValue({
 			changed: 8, modelFallback: true, sessionActive: false,
diff --git a/PanTS-Demo/src/helpers/viewer/useInteractivePromptTool.ts b/PanTS-Demo/src/helpers/viewer/useInteractivePromptTool.ts
index 844904e5..94bf6410 100644
--- a/PanTS-Demo/src/helpers/viewer/useInteractivePromptTool.ts
+++ b/PanTS-Demo/src/helpers/viewer/useInteractivePromptTool.ts
@@ -141,11 +141,21 @@ export function useInteractivePromptTool({
 	// The four prompt-mode hooks are mounted together. Their shared session is
 	// released only when its case, target label, or resolution is abandoned;
 	// changing between point, box, lasso, and scribble keeps the same context.
+	//
+	// Closing the tab or leaving the site never runs this cleanup, and the model
+	// server may allow only one lease at a time, so an abandoned tab would keep
+	// every other user's prompts and Refine "busy" until the idle timeout. Release
+	// on `pagehide` too (releasing twice is harmless: the second call finds no token).
 	useEffect(() => {
 		if (caseId == null || activeSegmentIndex == null) return;
-		return () => {
+		const release = () => {
 			void releaseInteractivePromptSession(apiBase, caseId, activeSegmentIndex, res);
 		};
+		window.addEventListener("pagehide", release);
+		return () => {
+			window.removeEventListener("pagehide", release);
+			release();
+		};
 	}, [apiBase, caseId, activeSegmentIndex, res]);
 
 	const executeSubmit = useCallback(async (_pane: CinePane, payload: any) => {
diff --git a/PanTS-Demo/src/test/refineFlyoutErrors.test.tsx b/PanTS-Demo/src/test/refineFlyoutErrors.test.tsx
new file mode 100644
index 00000000..a57bf468
--- /dev/null
+++ b/PanTS-Demo/src/test/refineFlyoutErrors.test.tsx
@@ -0,0 +1,51 @@
+/**
+ * What a person sees when "Refine with model" fails. The request layer words a full
+ * model server as HTTP 429 ("Server is busy...") and a down or stalled one as
+ * "Server error (5xx)...", and neither used to match the flyout's known messages, so
+ * every one of them read as a generic "Refining failed" with no hint to wait.
+ */
+import { fireEvent, render, screen } from "@testing-library/react";
+import { describe, expect, it, vi } from "vitest";
+import RefineFlyout from "../components/segmentation/RefineFlyout";
+
+const BUSY = "The interactive model is busy. Wait a moment and try again.";
+const UNAVAILABLE = "The interactive model is unavailable right now. Try again in a moment.";
+
+async function failWith(error: unknown) {
+	const refine = vi.fn().mockRejectedValue(error);
+	render(<RefineFlyout classLabel="Liver" refine={refine} />);
+	fireEvent.click(screen.getByRole("button", { name: /Refine Liver/ }));
+	return (await screen.findByRole("alert")).textContent;
+}
+
+describe("RefineFlyout failure messages", () => {
+	it.each([
+		["a 429 from the request layer", "Server is busy (too many requests). Please wait a moment and try again."],
+		["the backend's capacity wording", "Interactive segmentation is at capacity - please retry in a few seconds."],
+		["a stalled model server", "Interactive segmentation timed out - the server is busy. Please retry."],
+	])("says the model is busy for %s", async (_name, message) => {
+		expect(await failWith(new Error(message))).toBe(BUSY);
+	});
+
+	it.each([
+		["503", "Server error (503). Please try again later. If this keeps happening, try reloading or contact support."],
+		["502", "Server error (502). Please try again later. If this keeps happening, try reloading or contact support."],
+	])("says the model is unavailable for a %s", async (_code, message) => {
+		expect(await failWith(new Error(message))).toBe(UNAVAILABLE);
+	});
+
+	it("keeps the specific messages it already knew", async () => {
+		const known = "This class has no voxels to refine yet. Draw or segment it first.";
+		expect(await failWith(new Error(known))).toBe(known);
+	});
+
+	it("keeps the network wording for a failed fetch", async () => {
+		expect(await failWith(new TypeError("Failed to fetch"))).toBe(
+			"Could not reach the model server. Check your connection and try again.",
+		);
+	});
+
+	it("still falls back to the generic message for anything else", async () => {
+		expect(await failWith(new Error("Some unexpected failure"))).toBe("Refining failed. Try again.");
+	});
+});