diff --git a/internal/editor/logic.go b/internal/editor/logic.go
index f1e2890..70efed4 100644
--- a/internal/editor/logic.go
+++ b/internal/editor/logic.go
@@ -125,6 +125,10 @@ type Logic struct {
 	// the stat result, and is not a content-arrival signal).
 	restoreScroll      ui.Dp
 	restoreScrollArmed bool
+	// restoreArmedAt is when the restore scroll was armed; bounds how long
+	// maybeApplyRestoreScroll may wait for a trustworthy MaxScroll (see
+	// there) before landing best-effort.
+	restoreArmedAt time.Time
 	// restoreScrollLine/restoreScrollSub/restorePinDeadline implement the
 	// restore line-pin (see refreshRestorePin in session.go): the
 	// line-derived scroll offset is only consistent with the all-estimate
diff --git a/internal/editor/session.go b/internal/editor/session.go
index b6c41a9..9009431 100644
--- a/internal/editor/session.go
+++ b/internal/editor/session.go
@@ -258,6 +258,9 @@ func (l *Logic) BeginRestore(s SessionState) {
 	// struct).
 	l.restoreScroll = ui.Dp(s.Scroll)
 	l.restoreScrollArmed = s.Scroll > 0
+	if l.restoreScrollArmed {
+		l.restoreArmedAt = time.Now()
+	}
 	// The app-local font scale lands immediately, BEFORE any layout: the
 	// restore's line-based offset math reads it through EffectiveLineHeight.
 	if as := s.AppFontScale; as > 0 {
@@ -326,6 +329,19 @@ func (l *Logic) maybeApplyRestoreScroll() bool {
 	if l.state.Page() != EditorPage || !l.restoreContentLanded {
 		return false
 	}
+	// The clamp below needs a trustworthy MaxScroll. For a chunked buffer it
+	// is measured from the line index (or the large pre-index estimate) and
+	// is valid as soon as the content has landed. For a string buffer it
+	// derives from the shaper's last baseline, which is still 0 until the
+	// first non-empty frame has shaped — applying now would clamp the
+	// offset to 0 and the scroll position would be silently lost. Wait for
+	// that frame (its feedback triggers the emitFrame that re-runs this);
+	// a deadline covers content that never produces a baseline (empty or
+	// newlines-only), which lands best-effort.
+	if l.state.Editor.ChunkedBuffer == nil && l.state.LastLineY <= 0 &&
+		time.Since(l.restoreArmedAt) < restoreScrollArmTimeout {
+		return false
+	}
 	// Re-derive the offset from the persisted logical line (see the
 	// SessionState doc): the snapshot's pixel offset mapped to its position
 	// only while the saving session's WrapIndex was valid. On relaunch the
@@ -371,6 +387,7 @@ func (l *Logic) maybeApplyRestoreScroll() bool {
 	}
 	l.state.ScrollOffset = l.restoreScroll
 	l.restoreScrollArmed = false
+	log.Printf("RESTORE scroll off=%.1f line=%d maxScroll=%.1f", l.restoreScroll, s.ScrollLine, l.state.MaxScroll)
 	// The snapshot has fully landed (cursor/selection/find with the
 	// content, scroll now): lift the save suppression.
 	l.session = SessionState{}
@@ -381,6 +398,13 @@ func (l *Logic) maybeApplyRestoreScroll() bool {
 // the restored window has either settled or the user has moved on.
 const restorePinTimeout = 2 * time.Second
 
+// restoreScrollArmTimeout bounds how long maybeApplyRestoreScroll may wait
+// for a trustworthy MaxScroll before landing the offset best-effort (see
+// the guard there). Shaping a real file completes in well under this; it
+// exists so degenerate content (empty, newlines-only) cannot leave the
+// restore armed forever.
+const restoreScrollArmTimeout = 2 * time.Second
+
 // releaseRestorePin clears the restore line-pin (user or search took over
 // the viewport). Must be called on the logic goroutine.
 func (l *Logic) releaseRestorePin() {