diff --git a/cmd/harnesscli/tui/components/spinner/model.go b/cmd/harnesscli/tui/components/spinner/model.go
index c8377dea..95618265 100644
--- a/cmd/harnesscli/tui/components/spinner/model.go
+++ b/cmd/harnesscli/tui/components/spinner/model.go
@@ -277,13 +277,53 @@ func shortenLabel(glyph, label, full string, width int) string {
 		// Not even a stub of a label fits; the hint alone is more useful.
 		return glyph + " " + CancelHint
 	}
-	runes := []rune(label)
-	if len(runes) > budget {
-		label = string(runes[:budget-1]) + "\u2026"
-	}
+	label = truncateToWidth(label, budget)
 	return glyph + " " + label + " " + CancelHint
 }
 
+// truncateToWidth shortens label to at most budget terminal columns, appending
+// an ellipsis when it cuts.
+//
+// It measures display columns rather than runes. Counting runes was the bug in
+// issue #1432: CJK and emoji occupy two columns each, so a rune-budgeted label
+// could be twice its allowance in columns, overrun the line, and get clipped
+// from the right — taking the cancel hint with it. ASCII hid this completely,
+// because there columns and runes are the same number.
+//
+// Rune-by-rune accumulation is deliberate and imperfect: it does not segment
+// grapheme clusters, so a combining mark or ZWJ emoji sequence can still be
+// split. That is a strict improvement over rune counting and a much smaller
+// change than full segmentation, which is out of scope until a real case
+// appears.
+func truncateToWidth(label string, budget int) string {
+	if budget <= 0 {
+		return ""
+	}
+	if lipgloss.Width(label) <= budget {
+		return label
+	}
+
+	const ellipsis = "\u2026"
+	room := budget - lipgloss.Width(ellipsis)
+	if room <= 0 {
+		return ellipsis
+	}
+
+	var (
+		out   []rune
+		width int
+	)
+	for _, r := range label {
+		w := lipgloss.Width(string(r))
+		if width+w > room {
+			break
+		}
+		out = append(out, r)
+		width += w
+	}
+	return string(out) + ellipsis
+}
+
 // CompletionLine returns the one-line completion summary shown after the spinner stops.
 //
 // Format: "✻ Worked for 5s" or "✻ Worked for 1m 30s"
diff --git a/cmd/harnesscli/tui/components/spinner/truthful_label_test.go b/cmd/harnesscli/tui/components/spinner/truthful_label_test.go
index b8b211d6..e453bf6f 100644
--- a/cmd/harnesscli/tui/components/spinner/truthful_label_test.go
+++ b/cmd/harnesscli/tui/components/spinner/truthful_label_test.go
@@ -116,3 +116,40 @@ func TestSpinnerHintSurvivesEvenWhenLabelCannotFit(t *testing.T) {
 		t.Errorf("cancel hint should outrank the label when space is scarce, got %q", view)
 	}
 }
+
+// TestSpinnerKeepsCancelHintWithWideRunes pins issue #1432: the hint survives
+// regardless of the label's script.
+//
+// shortenLabel budgeted in display columns but truncated by rune count. For
+// CJK and emoji, which occupy two columns per rune, the label overran its
+// budget and the final MaxWidth clamp ate the hint — the exact outcome
+// shortenLabel exists to prevent. Every existing test passed throughout,
+// because they were all ASCII, where columns and runes coincide.
+//
+// The assertion is on the hint's survival, not on the line's width: the width
+// was always correct, which is why this went unnoticed.
+func TestSpinnerKeepsCancelHintWithWideRunes(t *testing.T) {
+	for _, tc := range []struct {
+		name  string
+		label string
+	}{
+		{name: "cjk", label: "Waiting for 模型模型模型模型模型模型模型模型"},
+		{name: "emoji", label: "Running 🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀"},
+		// ASCII control: a fix that over-truncates everything to satisfy the
+		// hint assertion would show up here as a needlessly stunted label.
+		{name: "ascii control", label: "Waiting for some-long-model-name-v2"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			m := New(0).Start().SetAction(tc.label)
+			for _, width := range []int{40, 50} {
+				view := m.View(width)
+				if !strings.Contains(view, CancelHint) {
+					t.Errorf("width %d: cancel hint truncated, got %q", width, view)
+				}
+				if got := lipgloss.Width(view); got > width {
+					t.Errorf("width %d: line is %d columns wide: %q", width, got, view)
+				}
+			}
+		})
+	}
+}
diff --git a/docs/logs/engineering-log.md b/docs/logs/engineering-log.md
index 7dba2d52..2aa892d2 100644
--- a/docs/logs/engineering-log.md
+++ b/docs/logs/engineering-log.md
@@ -1,5 +1,44 @@
 # Engineering Log
 
+## 2026-09-08 — Issue #1432 wide labels destroyed the cancel hint
+
+- Symptom: `shortenLabel` guarantees the `(esc to interrupt)` hint survives at
+  narrow widths, sacrificing the label instead. For double-width characters it
+  did the opposite:
+  ```
+  width=40  "· Waiting for 模型模型模型… (esc to inte"
+  width=40  "· Running 🚀🚀🚀🚀🚀🚀🚀🚀🚀🚀… (esc to "
+  ```
+  The line width was always correct — the final `MaxWidth` clamp saw to that —
+  which is exactly why nothing caught it. The clamp was cutting the hint off
+  because the label had not been shortened enough.
+- Cause: the function mixed two units. The budget was computed in display
+  columns (`lipgloss.Width`) but the truncation was applied in runes
+  (`runes[:budget-1]`). CJK and emoji occupy two columns per rune, so a
+  rune-budgeted label could be twice its allowance in columns. For ASCII the two
+  units coincide, so every existing test passed.
+- Fix: `truncateToWidth` accumulates `lipgloss.Width` per rune until the budget
+  is reached, counting the ellipsis's own width. Deliberately not
+  grapheme-cluster aware — combining marks and ZWJ sequences can still be split.
+  That is a strict improvement over rune counting and a far smaller change than
+  full segmentation, which stays out of scope until a real case appears.
+- Tests: `TestSpinnerKeepsCancelHintWithWideRunes` over CJK, emoji and an ASCII
+  control at widths 40 and 50, asserting the hint's *survival* rather than the
+  line's width — the width was never wrong, and asserting it is what let the bug
+  hide. The ASCII control guards against a fix that satisfies the hint
+  assertion by over-truncating every label.
+- **Provenance, and the reason this entry matters more than the bug.** Found by
+  `gpt-6-astra` through the Surplus proxy — an untrusted external reviewer given
+  the packed source and no tools — then confirmed locally by rendering the cases
+  before believing it. Our own tests were green throughout and stayed green,
+  because they were ASCII-only; the code asserted a guarantee in a comment that
+  it did not keep. An outside reader with no ability to run anything caught what
+  the test suite structurally could not.
+  - Signal-to-noise from that review, recorded honestly: three findings, one
+    confirmed (this), one unconfirmed and overlapping with it (a claim that
+    `MaxWidth` may wrap rather than clip — not observed), and one false positive
+    (elapsed-time nondeterminism in snapshots, which set `startTime` explicitly).
+
 ## 2026-09-08 — Issue #1428 CLAUDE.md: tickets, real-run proof, delegation
 
 - Three practices this session kept proving necessary were absent from