diff --git a/apps/desktop/src/renderer/maka-tokens.css b/apps/desktop/src/renderer/maka-tokens.css
index 03881e213d..3c24b9b58e 100644
--- a/apps/desktop/src/renderer/maka-tokens.css
+++ b/apps/desktop/src/renderer/maka-tokens.css
@@ -1410,8 +1410,11 @@
     font: var(--maka-text-body);
     background: var(--background);
     color: var(--foreground);
-    -webkit-font-smoothing: antialiased;
-    -moz-osx-font-smoothing: grayscale;
+    /* Astryx's reset puts `antialiased` on `:where(body)`, which strips the
+       stroke dilation macOS applies and thins 14px Han glyphs. Maka is
+       CJK-first, so the platform default wins here, for the whole document. */
+    -webkit-font-smoothing: auto;
+    -moz-osx-font-smoothing: auto;
     overflow: hidden;          /* main shell handles its own scrolling */
   }
 
diff --git a/packages/ui/src/__tests__/markdown-han-script.test.tsx b/packages/ui/src/__tests__/markdown-han-script.test.tsx
new file mode 100644
index 0000000000..1f76792d6e
--- /dev/null
+++ b/packages/ui/src/__tests__/markdown-han-script.test.tsx
@@ -0,0 +1,41 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+import assert from 'node:assert/strict';
+import { describe, it } from 'node:test';
+import { renderToStaticMarkup } from 'react-dom/server';
+import { hasHanProse, MarkdownBody } from '../markdown-body.js';
+
+describe('markdown Han-script marker', () => {
+  it('reads the prose, not the code', () => {
+    assert.equal(hasHanProse('按 `lastActiveAt` 排序。'), true);
+    assert.equal(hasHanProse('Sort by `lastActiveAt`.'), false);
+    assert.equal(hasHanProse('Sort by `排序键`.'), false);
+    assert.equal(hasHanProse('Example:\n\n```ts\n// 返回值为空时抛错\nreturn null;\n```\n'), false);
+    assert.equal(hasHanProse('说明：\n\n```ts\nreturn null;\n```\n'), true);
+  });
+
+  it('marks the markdown root so the emphasis rule can key on the script of the text', () => {
+    const han = renderToStaticMarkup(<MarkdownBody text="扁平化*只作用于按时间视图*。" density="compact" />);
+    assert.match(han, /<div[^>]*data-maka-contract="markdown"[^>]*data-maka-script="han"/);
+
+    const latin = renderToStaticMarkup(<MarkdownBody text="Flattening *only* applies to the time view." density="compact" />);
+    assert.doesNotMatch(latin, /data-maka-script/);
+  });
+});
diff --git a/packages/ui/src/markdown-body.tsx b/packages/ui/src/markdown-body.tsx
index e293b4b59e..2f3afdda3e 100644
--- a/packages/ui/src/markdown-body.tsx
+++ b/packages/ui/src/markdown-body.tsx
@@ -111,6 +111,17 @@ export function applyMermaidRenderBudget(source: string): string {
   return lines.join('\n');
 }
 
+// Whether the prose (not code) of a markdown source is written in Han script.
+// The document `lang` is the UI locale, which says nothing about what the
+// model wrote, so the CSS that styles Han-specific runs (emphasis has no
+// italic in Han faces) keys on this instead.
+export function hasHanProse(source: string): boolean {
+  const prose = source
+    .replace(/^( {0,3})(`{3,}|~{3,})[^\n]*\n[\s\S]*?\n {0,3}\2[ \t]*$/gm, '')
+    .replace(/`[^`\n]*`/g, '');
+  return /\p{Script=Han}/u.test(prose);
+}
+
 const MARKDOWN_COMPONENTS = {
   default: {
     ...BASE_MARKDOWN_COMPONENTS,
@@ -152,6 +163,7 @@ export function MarkdownBody(props: {
   return (
     <div
       data-maka-contract="markdown"
+      data-maka-script={hasHanProse(props.text) ? 'han' : undefined}
       // Migration-only identity wrapper. `display: contents` gives the
       // contract harness a stable declared subtree without adding a layout
       // box or interfering with Astryx's document root.
diff --git a/packages/ui/src/styles.css b/packages/ui/src/styles.css
index 659d9415f7..eee885b2be 100644
--- a/packages/ui/src/styles.css
+++ b/packages/ui/src/styles.css
@@ -409,16 +409,43 @@
 [data-maka-contract="markdown"] .astryx-markdown[data-density="compact"] .astryx-list {
   gap: var(--md-gap-list);
 }
-/* Heading scale: size says "this is a heading", weight and colour say which
-   level. Astryx's document ladder (20/18/16/14) is a page scale and an agent
-   turn emits `##` every few lines, so the transcript keeps two size steps —
-   but it keeps TWO, not one. The previous rule flattened h2-h6 onto a single
-   size, which left h2 and h3 identical in size, weight AND colour: three
-   levels of structure rendered as three identical bold lines.
-   Cursor (1.43/1.32/1.21), Claude Code (1.14/1.07/1.0) and Codex
-   (1.71/1.43/1.21) all flatten here relative to their own document styles, and
-   opencode goes further — 17/15/13 with h4+ dropped to a muted colour. Nobody
-   flattens to zero. Two steps plus colour sits inside that range.
+/* Prose reads differently from chrome. Han glyphs fill the whole em box, so
+   the 20px body leading that suits a control row leaves CJK paragraphs with
+   no air between lines; 1.6 is the common CJK reading ratio. Inherited from
+   the document root: headings and fenced code declare their own leading and
+   keep it, inline code follows the line it sits in, which is what keeps that
+   line's box even. `text-autospace` puts the conventional hair space between
+   Han and Latin or digit runs when the model did not type one. */
+[data-maka-contract="markdown"] .astryx-markdown {
+  line-height: 1.6;
+  text-autospace: normal;
+}
+/* Han faces have no italic, so `*emphasis*` in Chinese renders as a synthetic
+   shear of PingFang. Chinese typography marks emphasis with a dot under each
+   character instead. Keyed on the script of the turn's own prose (MarkdownBody
+   sets `data-maka-script`), not on the document `lang`: that is the UI locale,
+   and a model answers in whatever language it was asked in. Latin prose keeps
+   its italic. */
+[data-maka-contract="markdown"][data-maka-script="han"] .astryx-markdown em {
+  font-style: normal;
+  text-emphasis: filled dot;
+  text-emphasis-position: under right;
+}
+/* Heading scale: size says "this is a heading", colour says which level.
+   Astryx's document ladder (20/18/16/14) is a page scale and an agent turn
+   emits `##` and `###` every few lines, so the transcript keeps ONE size step
+   above body — 16 for h1/h2 — and everything below at body size. Chat-first
+   clients sit exactly here: Claude Code (1.07/1.0), opencode (15/13 with h4+
+   muted), Lody (16/14); only the document-leaning ones (Cursor 1.32/1.21,
+   Codex 1.43/1.21) go bigger, and a transcript at those sizes reads as a
+   stack of slabs rather than an answer.
+   h3 cannot be a body-size bold line in CJK, though: PingFang's Regular and
+   Semibold are close enough that it is indistinguishable from a bold
+   sentence, and a Chinese turn lost a whole level of structure. Size would
+   bring the slabs back, so h3 steps down in ink instead and shares the muted
+   tier with h4-h6 — the same axis opencode uses one level lower. `####` is
+   rare in a turn, so h3 and h4 sharing a style costs less than h3 and
+   `<strong>` sharing one.
 
    RISK, not just a choice: these are the only rules here that are typography
    rather than spacing, and Astryx's density RFC (facebook/astryx#839) draws
@@ -451,14 +478,6 @@
 }
 [data-maka-contract="markdown"] .astryx-markdown[data-density="compact"] .astryx-markdown-heading:not([data-level="1"], [data-level="2"]) {
   font: var(--maka-text-heading-4);
-}
-/* Below h3 the size ladder is spent — h4-h6 already sit at body size, and
-   growing them back would re-create the slabs the flattening exists to avoid.
-   Colour is the one axis left that costs no vertical space, so the deepest
-   levels step down in ink instead. opencode makes the same call at the same
-   point (13px h4-h6 dropped to muted); it is why a turn can carry four levels
-   of structure inside two type sizes. */
-[data-maka-contract="markdown"] .astryx-markdown[data-density="compact"] .astryx-markdown-heading:not([data-level="1"], [data-level="2"], [data-level="3"]) {
   color: var(--muted-foreground);
 }
 
diff --git a/packages/ui/stories/markdown.stories.tsx b/packages/ui/stories/markdown.stories.tsx
index b29b096e51..0a4f3b0c69 100644
--- a/packages/ui/stories/markdown.stories.tsx
+++ b/packages/ui/stories/markdown.stories.tsx
@@ -88,7 +88,7 @@ export const TranscriptTurn: Story = {
               '',
               '#### 边界情况',
               '',
-              '按项目视图仍然保留项目折叠，扁平化只作用于按时间视图。',
+              '按项目视图仍然保留项目折叠，扁平化*只作用于按时间视图*，按`lastActiveAt`排序，PR #4580 的 3 个 executor 不受影响。',
               '',
               '> 注意：新增状态只需补一个行图标。',
               '',