diff --git a/frontend/annotator/src/adapters/react/AnnotatorCanvas.tsx b/frontend/annotator/src/adapters/react/AnnotatorCanvas.tsx
index f92baad8..df940d7a 100644
--- a/frontend/annotator/src/adapters/react/AnnotatorCanvas.tsx
+++ b/frontend/annotator/src/adapters/react/AnnotatorCanvas.tsx
@@ -489,13 +489,47 @@ export function AnnotatorCanvas({
     [store, activeClass, mint],
   );
 
-  /** The whole asset, centred. What `mod+0` answers and what a mount starts at. */
+  /**
+   * The whole asset, centred. What `mod+0` answers and what a mount starts at.
+   *
+   * **Its dependencies are the asset's three numbers, never the descriptor
+   * object**, and that distinction is the whole of a defect that cost people
+   * their place in the picture on every save.
+   *
+   * A fit is a function of the frame: an id and a size. The *object* carrying
+   * them is minted afresh by `documentFromWire` on every rebuild, and a host
+   * rebuilds its document for reasons that have nothing to do with the frame —
+   * `ui-core` refetches after a save so the kernel's own annotation ids replace
+   * its client-minted ones, which is a materially different payload and so a new
+   * document, a new descriptor, and — while this closed over `asset` — a new
+   * `fit` and a layout effect that ran again. Zoom into a detail, store your
+   * work, and the stage jumped back to the fitted view.
+   *
+   * Depending on the numbers makes the effect below fire when the *picture*
+   * changes and at no other time, which is what it always meant. It is not a
+   * throttle on an effect that was otherwise right: a document rebuild is not a
+   * reason to move a camera, and an equality check on the object could never
+   * have told the two apart.
+   *
+   * The primitives live in **this** list rather than in the effect's, because
+   * `react-hooks/exhaustive-deps` is an error in this package and reports an
+   * unnecessary dependency as loudly as a missing one — so the honest spelling
+   * is a callback whose identity already tracks the right thing.
+   */
+  const { id: assetId, width: assetWidth, height: assetHeight } = asset;
   const fit = useCallback(() => {
     const pane = paneRef.current;
     if (pane === null) return;
     const rect = pane.getBoundingClientRect();
-    applyViewport(fitToViewport(asset, rect.width, rect.height, FIT_PADDING_PX));
-  }, [asset, applyViewport]);
+    applyViewport(
+      fitToViewport(
+        { id: assetId, width: assetWidth, height: assetHeight },
+        rect.width,
+        rect.height,
+        FIT_PADDING_PX,
+      ),
+    );
+  }, [assetId, assetWidth, assetHeight, applyViewport]);
 
   // Before the first paint, so the asset does not flash at native scale first.
   useLayoutEffect(fit, [fit]);
diff --git a/frontend/app/e2e/annotate.spec.ts b/frontend/app/e2e/annotate.spec.ts
index 845ecfbc..ccf55fee 100644
--- a/frontend/app/e2e/annotate.spec.ts
+++ b/frontend/app/e2e/annotate.spec.ts
@@ -2831,3 +2831,63 @@ test("Save and stay teaches its chord in a tooltip now the keycap is gone", asyn
   // half that does not move.
   await expect(tip).toContainText(/Save and stay \((⌘|Ctrl)S\)/);
 });
+
+/**
+ * Saving must not move the camera.
+ *
+ * The viewport is `AnnotatorCanvas`'s own state and only a real browser has one:
+ * jsdom's `getBoundingClientRect` answers all zeros, so there is no fit to
+ * disturb, no wheel notch to apply and no pan to measure. A component test for
+ * this would pass with the bug fully present.
+ *
+ * Both halves are read at once off the `<svg>`'s box, which is `_frame.ts`'s own
+ * idiom: the element is laid out at the asset's native size inside the
+ * `translate(pan) scale(zoom)` wrapper, so its on-screen rect folds zoom, pan and
+ * the pane's origin into one measurement. The readout is asserted beside it
+ * because a zoom that survived while the pan did not would otherwise read as a
+ * pass.
+ */
+test("saving leaves the viewport exactly where it was", async ({ page }) => {
+  const sent: Request[] = [];
+  await openJob(page, sent);
+
+  const canvas = page.getByTestId("annotator-canvas");
+  const pane = (await page.getByTestId("annotator-pane").boundingBox())!;
+
+  // Off the fitted view in both dimensions: a wheel notch over a point that is
+  // not the pane's centre changes the zoom *and* the pan, and the secondary drag
+  // after it moves the pan again on its own.
+  await page.mouse.move(pane.x + pane.width * 0.35, pane.y + pane.height * 0.35);
+  await page.mouse.wheel(0, -600);
+  await page.mouse.down({ button: "right" });
+  await page.mouse.move(pane.x + pane.width * 0.55, pane.y + pane.height * 0.5, { steps: 8 });
+  await page.mouse.up({ button: "right" });
+
+  const zoomBefore = await page.getByTestId("zoom-readout").textContent();
+  expect(zoomBefore).not.toBe("100%");
+  const frameBefore = (await canvas.boundingBox())!;
+
+  // Something to actually store — a save with an empty plan sends no request and
+  // rebuilds nothing, so a clean frame could not reproduce this at all.
+  await drawOneUnsavedBox(page);
+  await page.getByTestId("save-and-stay").click();
+  await expect(page.getByTestId("save-state")).toContainText("Saved");
+  // The refetch the save triggers is what rebuilds the store; wait for the
+  // rebuilt document rather than for the button, or the assertion below can run
+  // in the window before the camera has been moved.
+  await expect(page.getByTestId("object-total")).toContainText("1 object");
+
+  expect(await page.getByTestId("zoom-readout").textContent()).toBe(zoomBefore);
+  const frameAfter = (await canvas.boundingBox())!;
+  expect({
+    x: Math.round(frameAfter.x),
+    y: Math.round(frameAfter.y),
+    width: Math.round(frameAfter.width),
+    height: Math.round(frameAfter.height),
+  }).toEqual({
+    x: Math.round(frameBefore.x),
+    y: Math.round(frameBefore.y),
+    width: Math.round(frameBefore.width),
+    height: Math.round(frameBefore.height),
+  });
+});