fix(native-chat): bound marker stripping by the run's image count

Anchoring decided which turn gets stripped but not how many markers may go,
so a turn attaching one image and mentioning a second marker in prose still
lost the user's literal text. Strip at most one marker per image in the
anchored source run, in document order across the turn's text blocks; markers
past the count stay verbatim.

Position, not ordinal, decides which markers are placeholders, so the rule
holds whether the producing CLI numbers markers per-turn or per-session.

mobile/src/session/mobile-native-chat-render-data.test.ts

@@ -133,6 +133,22 @@ describe('buildMobileNativeChatTransientData', () => {
     ])
   })
 
+  it('keeps a marker beyond the folded run’s image count as literal text', () => {
+    const data = build(
+      [
+        user('u1', '[Image: source: /tmp/a.png]'),
+        user('u2', '[Image #1] compare with the [Image #2] I mentioned earlier')
+      ],
+      null,
+      []
+    )
+
+    expect(data[0]?.blocks).toEqual([
+      { type: 'image-ref', path: '/tmp/a.png' },
+      { type: 'text', text: 'compare with the [Image #2] I mentioned earlier' }
+    ])
+  })
+
   it('renders a lone image marker turn (no caption) as an image-ref block', () => {
     const data = build([user('u1', '[Image: source: /tmp/a.png]')], null, [])
     expect(data[0]?.blocks).toEqual([{ type: 'image-ref', path: '/tmp/a.png' }])

src/shared/native-chat-image-transcript-markers.test.ts

@@ -259,6 +259,79 @@ describe('normalizeImageTranscriptMessages', () => {
     ])
   })
 
+  it('keeps markers beyond the anchored run’s image count as the user’s own words', () => {
+    const out = normalizeImageTranscriptMessages([
+      userText('s1', '[Image: source: /tmp/a.png]'),
+      userText('p1', '[Image #1] compare with the [Image #2] I mentioned earlier')
+    ])
+
+    expect(out).toHaveLength(1)
+    expect(out[0]!.blocks).toEqual([
+      { type: 'image-ref', path: '/tmp/a.png' },
+      { type: 'text', text: 'compare with the [Image #2] I mentioned earlier' }
+    ])
+  })
+
+  it('strips by document position, not by marker ordinal', () => {
+    const out = normalizeImageTranscriptMessages([
+      userText('s1', '[Image: source: /tmp/a.png]'),
+      userText('p1', '[Image #2] came first but [Image #1] came second')
+    ])
+
+    expect(out[0]!.blocks).toEqual([
+      { type: 'image-ref', path: '/tmp/a.png' },
+      { type: 'text', text: 'came first but [Image #1] came second' }
+    ])
+  })
+
+  it('strips a marker whose ordinal exceeds the run size when it is within the count', () => {
+    const out = normalizeImageTranscriptMessages([
+      userText('s1', '[Image: source: /tmp/a.png]'),
+      userText('p1', '[Image #7] what is this?')
+    ])
+
+    expect(out[0]!.blocks).toEqual([
+      { type: 'image-ref', path: '/tmp/a.png' },
+      { type: 'text', text: 'what is this?' }
+    ])
+  })
+
+  it('spends the image budget across text blocks in document order', () => {
+    const prompt: NativeChatMessage = {
+      ...userText('prompt', 'unused'),
+      blocks: [
+        { type: 'text', text: '[Image #1] look at' },
+        { type: 'image-ref', path: '/tmp/existing.png' },
+        { type: 'text', text: 'not [Image #2] though' }
+      ]
+    }
+    const out = normalizeImageTranscriptMessages([
+      userText('s1', '[Image: source: /tmp/a.png]'),
+      prompt
+    ])
+
+    expect(out[0]!.blocks).toEqual([
+      { type: 'image-ref', path: '/tmp/a.png' },
+      { type: 'text', text: 'look at' },
+      { type: 'image-ref', path: '/tmp/existing.png' },
+      { type: 'text', text: 'not [Image #2] though' }
+    ])
+  })
+
+  it('keeps the surplus marker on a two-image run that mentions a third', () => {
+    const out = normalizeImageTranscriptMessages([
+      userText('s1', '[Image: source: /tmp/a.png]'),
+      userText('s2', '[Image: source: /tmp/b.png]'),
+      userText('p1', '[Image #1] [Image #2] versus the [Image #3] from before')
+    ])
+
+    expect(out[0]!.blocks).toEqual([
+      { type: 'image-ref', path: '/tmp/a.png' },
+      { type: 'image-ref', path: '/tmp/b.png' },
+      { type: 'text', text: 'versus the [Image #3] from before' }
+    ])
+  })
+
   it('leaves assistant messages untouched', () => {
     const assistant: NativeChatMessage = {
       id: 'a',

src/shared/native-chat-image-transcript-markers.ts

@@ -27,20 +27,41 @@ export function isImageSourceUserTurn(message: NativeChatMessage): boolean {
   return message.role === 'user' && imageSourcePathFromText(soleText(message) ?? '') !== null
 }
 
-export function stripImagePromptMarker(text: string): string {
-  const stripped = text.replace(IMAGE_PROMPT_MARKERS, '')
+function countMarkers(text: string): number {
+  return text.match(IMAGE_PROMPT_MARKERS)?.length ?? 0
+}
+
+/** Strips the first `limit` markers in document order; the rest are the user's
+ *  own words, so they stay verbatim. */
+function stripImagePromptMarkersUpTo(text: string, limit: number): string {
+  if (limit <= 0) {
+    return text
+  }
+  let used = 0
+  const stripped = text.replace(IMAGE_PROMPT_MARKERS, (match) => {
+    if (used >= limit) {
+      return match
+    }
+    used += 1
+    return ''
+  })
   if (stripped === text) {
     return text
   }
   let result = IMAGE_PROMPT_MARKER_AT_START.test(text)
     ? stripped.replace(HORIZONTAL_WHITESPACE_START, '')
     : stripped
-  if (IMAGE_PROMPT_MARKER_AT_END.test(text)) {
+  // Why: a trailing marker left in place still owns the whitespace before it.
+  if (used === countMarkers(text) && IMAGE_PROMPT_MARKER_AT_END.test(text)) {
     result = result.replace(HORIZONTAL_WHITESPACE_END, '')
   }
   return result
 }
 
+export function stripImagePromptMarker(text: string): string {
+  return stripImagePromptMarkersUpTo(text, Number.POSITIVE_INFINITY)
+}
+
 export function normalizeNativeChatUserText(text: string): string {
   return stripImagePromptMarker(text).trim().replace(/\s+/g, ' ')
 }
@@ -78,10 +99,14 @@ export function nativeChatUserMessageMatchText(message: NativeChatMessage): stri
   return nativeChatUserTextMatchText(joined, message.blocks.some(isImageRefBlock)) || null
 }
 
+/** `limit` is the number of images the run actually carried. Only that many
+ *  markers can be placeholders; the rest are text the user typed. */
 function stripImagePromptMarkersFromTextBlocks(
-  blocks: readonly NativeChatBlock[]
+  blocks: readonly NativeChatBlock[],
+  limit: number
 ): NativeChatBlock[] {
   let sawText = false
+  let remaining = limit
   let next: NativeChatBlock[] | null = null
   for (let index = 0; index < blocks.length; index += 1) {
     const block = blocks[index]!
@@ -91,7 +116,8 @@ function stripImagePromptMarkersFromTextBlocks(
     }
     const isFirstText = !sawText
     sawText = true
-    const text = stripImagePromptMarker(block.text)
+    const text = stripImagePromptMarkersUpTo(block.text, remaining)
+    remaining -= Math.min(remaining, countMarkers(block.text))
     if (!text.trim() && (text !== block.text || isFirstText)) {
       next ??= blocks.slice(0, index)
       continue
@@ -173,7 +199,7 @@ export function normalizeImageTranscriptMessages(
           ...prompt,
           blocks: [
             ...imagePaths.map((path) => ({ type: 'image-ref' as const, path })),
-            ...stripImagePromptMarkersFromTextBlocks(prompt.blocks)
+            ...stripImagePromptMarkersFromTextBlocks(prompt.blocks, imagePaths.length)
           ]
         })
         index = nextIndex