feat(diagram): named styles the model defines once and uses by name

The model may now write <mxStyle name="step" value="..."/> before its cells
and refer to the name inside a cell's style, the way draw.io resolves its
own named styles. style-classes.ts expands the names after validation (so
repaired cells get them too) and adds draw.io's defaults html=1 and
whiteSpace=wrap where they belong, so the prompts can tell the model never
to write them. Wired into display_diagram, append_diagram, create_new_diagram,
add_page, edit_diagram (defaults only) and the streaming preview.

Prompts (system prompt, tool description, MCP drawing guide) teach the
notation with a shared example, drop the draw.io defaults from every
example and tell the model how labels, line breaks and reserved names work.

Measured on the five start-screen examples, same model and prompt, old vs
new prompt: gpt-6-luna 19% fewer output tokens, Opus 5.5 19% fewer; flow
and architecture diagrams 15% to 46%, creative drawings about equal.

lib/utils.ts

@@ -2,6 +2,7 @@ import { type ClassValue, clsx } from "clsx"
 import * as pako from "pako"
 import { twMerge } from "tailwind-merge"
 import { hasCells } from "@/packages/mcp-server/src/pages.ts"
+import { readStyleClasses } from "@/packages/mcp-server/src/style-classes.ts"
 
 export function cn(...inputs: ClassValue[]) {
     return twMerge(clsx(inputs))
@@ -41,7 +42,9 @@ export function isRealDiagram(xml: string | undefined | null): boolean {
  * @returns true if XML appears complete, false if truncated or empty
  */
 export function isMxCellXmlComplete(xml: string | undefined | null): boolean {
-    const trimmed = xml?.trim() || ""
+    // Named style definitions before the cells are not cells: output cut off
+    // right after them is incomplete
+    const trimmed = readStyleClasses(xml || "").xml.trim()
     if (!trimmed) return false
 
     // Find position of last complete mxCell ending (either /> or </mxCell>)