feat(chat): 动作/旁白格式统一——兼容星号(＊…＊/＊…) + 人设强制 XML

模型常漂移成 ＊动作：弯起眼睛＊(星号)，前端/后端只认 <action>/<narration>，导致星号版
没被剥成灰斜体注记/没从 IM 剥除→裸出在气泡、上下文不连贯。
- parseContent.ts(新): 把 XML <action>/<narration>、全角＊…＊(带'动作/旁白/内心/神态/轻声'
  关键词或整段≤40字)、半角*词…*(仅关键词前缀)统一解析成 ContentPart,剥出正文进 BubbleNotes
  灰斜体注记(＊动作:…＊); 超长无关键词＊…＊保留为文本; <p>拆气泡不变
- sanitize.go: stripNarration 同样剥星号动作(与 XML 一致); splitBubbles 不变
- cyrene.yaml: 新增格式铁律——动作/旁白必须用 <action>/<narration>,禁用＊/Markdown等
验收: 后端 build/vet/test 全绿 + 前端 build 绿 + Node parseContent/splitBubble 实测(5组)

frontend/web/src/lib/parseContent.ts

@@ -1,9 +1,15 @@
-// Structured parser for persona action/narration XML tags (design/07 §1).
-// A single message `content` may mix plain text with `<action>…</action>` and
-// `<narration>…</narration>`; this splits it into an ordered parts array so the
-// UI can position each block (inline, or as an attachment note via `splitBubble`)
-// while the structured data stays available for downstream feeds (mind-map /
-// state) via `summarizeContent`.
+// Structured parser for persona action/narration markup (design/07 §1).
+// A single message `content` may mix plain text with an action/narration block in
+// any of three equivalent forms, so a model that drifts from the canonical XML
+// still peels into the same muted note instead of leaking markup into a bubble:
+//   1. <action>…</action> / <narration>…</narration>  — the canonical form;
+//   2. ＊…＊ full-width star  (＊动作：弯起眼睛＊);
+//   3. *…* half-width star, only when it clearly carries a role-play keyword
+//      (*动作：弯起眼睛*) — plain *emphasis* stays as text.
+// This splits `content` into an ordered parts array so the UI can position each
+// block (inline, or as an attachment note via `splitBubble`) while the structured
+// data stays available for downstream feeds (mind-map / state) via
+// `summarizeContent`.
 
 export type ContentPartType = "text" | "action" | "narration";
 
@@ -12,35 +18,90 @@ export interface ContentPart {
   text: string;
 }
 
-/** Matches a well-formed `<action>…</action>` or `<narration>…</narration>`.
- *  Non-greedy body (`[\s\S]*?`) stops at the first matching close tag, so the
- *  tags can span multiple lines. Unclosed or malformed tags simply do not match
- *  and fall through as plain text — a literal `<` in the body never breaks. */
-const TAG_RE = /<\s*(action|narration)\s*>([\s\S]*?)<\/\s*\1\s*>/g;
+/** Matches one action/narration block in any of its three accepted forms:
+ *  `<action>…</action>` / `<narration>…</narration>`, a `＊…＊` full-width star
+ *  block, or a `*…*` half-width star block that begins with a role-play keyword.
+ *  The alternation is ordered left-to-right (the leftmost token wins), and each
+ *  body is non-greedy (`[\s\S]*?`) so a block may span newlines while a malformed
+ *  or unclosed tag/star simply falls through as plain text. Capture groups:
+ *  1 = XML tag name, 2 = XML body, 3 = full-star body, 4 = half-star body. */
+const ROLE_RE =
+  /<\s*(action|narration)\s*>([\s\S]*?)<\/\s*\1\s*>|＊([\s\S]*?)＊|\*((?:动作|旁白|内心|神态|轻声)\s*[:：]?[\s\S]*?)\*/g;
+
+/** Leading role-play keyword of a star block, stripped before it becomes a note. */
+const ROLE_KEYWORD_RE = /^(动作|旁白|内心|神态|轻声)\s*[:：]?\s*/;
+
+/** Caps how long a keyword-less full-width star block may be and still count as a
+ *  role-play remark (the drift shape ＊弯起眼睛＊). A long keyword-less ＊…＊
+ *  paragraph stays as plain text rather than being mislabelled as an action note. */
+const STAR_SHORT_RUNES = 40;
+const NEWLINE_RE = /[\n\r]/;
 
 /**
- * Split `content` into an ordered parts array preserving tag position.
- * Pure text with no tags yields a single `{type:"text"}` part.
+ * Split `content` into an ordered parts array preserving block position.
+ * Pure text with no block yields a single `{type:"text"}` part.
  */
 export function parseContent(content: string): ContentPart[] {
   const parts: ContentPart[] = [];
-  let lastIndex = 0;
+  let last = 0;
+  ROLE_RE.lastIndex = 0;
 
-  for (const m of content.matchAll(TAG_RE)) {
-    const index = m.index ?? content.length;
-    const before = content.slice(lastIndex, index);
+  let m: RegExpExecArray | null;
+  while ((m = ROLE_RE.exec(content)) !== null) {
+    const before = content.slice(last, m.index);
     if (before) parts.push({ type: "text", text: before });
-    parts.push({ type: m[1] === "action" ? "action" : "narration", text: m[2] });
-    lastIndex = index + m[0].length;
+    parts.push(blockToPart(m));
+    last = m.index! + m[0].length;
   }
 
-  const tail = content.slice(lastIndex);
+  const tail = content.slice(last);
   if (tail) parts.push({ type: "text", text: tail });
   if (parts.length === 0) parts.push({ type: "text", text: content });
 
   return parts;
 }
 
+/** Turn a single matching block into its content part, per which form it took. */
+function blockToPart(m: RegExpExecArray): ContentPart {
+  if (m[1]) {
+    // Canonical <action>/<narration>: body is used as-is (it has no keyword prefix).
+    return { type: m[1] === "action" ? "action" : "narration", text: m[2] };
+  }
+  if (m[3]) return classifyFullStar(m[3]) ?? { type: "text", text: m[0] };
+  return classifyHalfStar(m[4]);
+}
+
+/** A `＊…＊` block is role-play (returned as a part) whenever it carries a role-play
+ *  keyword or is a short stand-alone fragment; a long keyword-less paragraph wrapped
+ *  in stars returns `null`, so the caller keeps it as plain text star-for-star. */
+function classifyFullStar(body: string): ContentPart | null {
+  const kw = matchRoleKeyword(body);
+  if (kw) return { type: kw.kind, text: kw.rest };
+  const t = body.trim();
+  if (t && t.length <= STAR_SHORT_RUNES && !NEWLINE_RE.test(t)) {
+    return { type: "action", text: t };
+  }
+  return null;
+}
+
+/** A `*…*` block is role-play only because the regex already required a keyword,
+ *  so it is peeled; a keyword-less `*emphasis*` never makes it here and stays text. */
+function classifyHalfStar(body: string): ContentPart {
+  const kw = matchRoleKeyword(body);
+  if (kw) return { type: kw.kind, text: kw.rest };
+  return { type: "action", text: body.trim() };
+}
+
+/** Matches the leading `动作/旁白/内心/神态/轻声` keyword of a star block, mapping
+ *  it to an action/narration kind and returning the bare text after it. */
+function matchRoleKeyword(body: string): { kind: "action" | "narration"; rest: string } | null {
+  const m = ROLE_KEYWORD_RE.exec(body);
+  if (!m) return null;
+  const kind: "action" | "narration" =
+    m[1] === "旁白" || m[1] === "内心" ? "narration" : "action";
+  return { kind, rest: body.slice(m[0].length).trim() };
+}
+
 /** One extracted tag value. */
 export interface ContentTagItem {
   text: string;