fix(gui): harden multimodal image attachments

packages/llm/token-meter/tests/token-meter.spec.ts

@@ -1,5 +1,6 @@
 import { describe, expect, expectTypeOf, it } from 'vitest'
 import { Context } from 'cordis'
+import { AttachmentId } from '@deepseek-ai/dsh-attachment'
 import { CallId } from '@deepseek-ai/dsh-llm'
 import type { ContentBlock, Message, TokenUsage } from '@deepseek-ai/dsh-llm'
 import SessionStore, { Session, SessionId, canonicalHeader } from '@deepseek-ai/dsh-session'
@@ -116,6 +117,16 @@ describe('TokenMeterService pricing', () => {
     const blocks: ContentBlock[] = [
       { type: 'text', text: 'abcd' },
       { type: 'reasoning', text: 'ab' },
+      {
+        type: 'image',
+        attachment: {
+          attachmentId: AttachmentId(`sha256:${'a'.repeat(64)}`),
+          mediaType: 'image/png',
+          bytes: 1,
+          width: 1024,
+          height: 513,
+        },
+      },
       { type: 'tool-call', id: CallId('c'), name: 'read', arguments: '{"x":1}' },
       {
         type: 'tool-result',
@@ -126,7 +137,7 @@ describe('TokenMeterService pricing', () => {
       { type: 'future-block', payload: 'abcd' } as unknown as ContentBlock,
     ]
     const estimated = service.estimateMessage({ role: 'assistant', content: blocks })
-    expect(estimated).toBeGreaterThan(30)
+    expect(estimated).toBe(813)
     expect(service.estimateMessage(textMessage('abcd'))).toBe(9)
   })