fix: accept output_text, reasoning, and input_image in /v1/responses mapper (mlx-node#50)

<!-- CURSOR_SUMMARY -->
> [!NOTE]
> **Medium Risk**
> Medium risk because it changes request/history mapping and VLM
session/cache handling (image placeholders, M-RoPE offsets, and delta
continuations), which can affect multi-turn correctness and replayed
history integrity.
> 
> **Overview**
> **Expands the `/v1/responses` request mapper and stored-history codec
to properly handle client replay and images.** `mapRequest` now accepts
replayed assistant `output_text`, top-level `reasoning` items, and
`refusal/summary_text` parts; it also decodes `input_image` base64 data
URLs into `ChatMessage.images` and coalesces interleaved assistant
`reasoning`/`message`/`function_call` items into a single assistant
turn.
> 
> **Adds strict text/image ordering guards** in both the OpenAI and
Anthropic mappers to reject any text-after-image interleaving within a
single user turn (previously the pipeline could silently reorder
content).
> 
> **Makes images survive persistence and VLM templating, and fixes
Qwen3.5 VLM session correctness.** Stored `inputJson` is now written via
`stringifyStoredInputMessages` (base64 sentinel) and revived during
`reconstructMessagesFromChain` so `Uint8Array` images round-trip;
tokenizer sanitization/serialization now preserves user images so Jinja
templates can emit inline vision markers; Qwen3.5/Qwen3.5-MoE update
image placeholder expansion (per-image counts), robust M-RoPE position
indexing for multiple image runs, and session-delta handling so
text-only deltas can continue on image-bearing sessions without clearing
`cached_image_key`/rope deltas. Extensive new unit tests cover these
behaviors.
> 
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit
9c33817. Bugbot is set up for automated
code reviews on this repo. Configure
[here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: Brooooooklyn

__test__/server/anthropic-request-mapper.test.ts

@@ -263,6 +263,10 @@ describe('mapAnthropicRequest', () => {
   });
 
   it('rejects trailing image followed by text after a tool_result prefix', () => {
+    // The text-after-image guard fires during the main loop (before the
+    // trailing-mixed check) because the serializer cannot preserve text
+    // that follows an image in any turn, not just after a tool_result
+    // prefix. Either rejection is equivalent for the caller.
     expect(() =>
       mapAnthropicRequest({
         model: 'claude-3-5-sonnet-20241022',
@@ -278,7 +282,7 @@ describe('mapAnthropicRequest', () => {
           },
         ],
       }),
-    ).toThrow(/mixing trailing text and image blocks after a tool_result prefix/i);
+    ).toThrow(/text block after an image block in the same user turn/i);
   });
 
   it('rejects multiple trailing image blocks after a tool_result prefix', () => {
@@ -857,4 +861,112 @@ describe('mapAnthropicRequest', () => {
 
     expect(messages).toEqual([{ role: 'tool', content: 'Result: success', toolCallId: 'call_789' }]);
   });
+
+  describe('text/image ordering in pure user turns', () => {
+    // Existing tool_result-prefix rejection at line ~160 catches mixed
+    // trailing content after tool_result, but a PURE user turn (no
+    // tool_result blocks) was previously silently concatenating all text
+    // and stacking all images, reordering the caller's content. These
+    // tests pin the uniform rejection for both call patterns.
+    const png = 'iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg==';
+
+    it('rejects [image, text] in a pure user turn', () => {
+      // Image-first-then-text gets reordered to text-first-then-image by
+      // the flat ChatMessage + Jinja serializer pipeline. Reject rather
+      // than silently rewrite the caller's intent.
+      expect(() =>
+        mapAnthropicRequest({
+          model: 'claude-3-5-sonnet-20241022',
+          max_tokens: 1024,
+          messages: [
+            {
+              role: 'user',
+              content: [
+                { type: 'image', source: { type: 'base64', media_type: 'image/png', data: png } },
+                { type: 'text', text: 'describe this' },
+              ],
+            },
+          ],
+        }),
+      ).toThrow(/text block after an image block in the same user turn/i);
+    });
+
+    it('rejects [text, image, text] in a pure user turn', () => {
+      expect(() =>
+        mapAnthropicRequest({
+          model: 'claude-3-5-sonnet-20241022',
+          max_tokens: 1024,
+          messages: [
+            {
+              role: 'user',
+              content: [
+                { type: 'text', text: 'before' },
+                { type: 'image', source: { type: 'base64', media_type: 'image/png', data: png } },
+                { type: 'text', text: 'after' },
+              ],
+            },
+          ],
+        }),
+      ).toThrow(/text block after an image block/i);
+    });
+
+    it('accepts [text, image] in a pure user turn (representable by the flat ChatMessage)', () => {
+      const { messages } = mapAnthropicRequest({
+        model: 'claude-3-5-sonnet-20241022',
+        max_tokens: 1024,
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'text', text: 'what colour is this?' },
+              { type: 'image', source: { type: 'base64', media_type: 'image/png', data: png } },
+            ],
+          },
+        ],
+      });
+      expect(messages).toHaveLength(1);
+      expect(messages[0].role).toBe('user');
+      expect(messages[0].content).toBe('what colour is this?');
+      expect(messages[0].images).toHaveLength(1);
+    });
+
+    it('accepts [text, text, image] (all text parts before the image)', () => {
+      const { messages } = mapAnthropicRequest({
+        model: 'claude-3-5-sonnet-20241022',
+        max_tokens: 1024,
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'text', text: 'part one. ' },
+              { type: 'text', text: 'part two.' },
+              { type: 'image', source: { type: 'base64', media_type: 'image/png', data: png } },
+            ],
+          },
+        ],
+      });
+      expect(messages).toHaveLength(1);
+      expect(messages[0].content).toBe('part one. part two.');
+      expect(messages[0].images).toHaveLength(1);
+    });
+
+    it('accepts multiple images with no text (no ordering ambiguity)', () => {
+      const { messages } = mapAnthropicRequest({
+        model: 'claude-3-5-sonnet-20241022',
+        max_tokens: 1024,
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'image', source: { type: 'base64', media_type: 'image/png', data: png } },
+              { type: 'image', source: { type: 'base64', media_type: 'image/png', data: png } },
+            ],
+          },
+        ],
+      });
+      expect(messages).toHaveLength(1);
+      expect(messages[0].content).toBe('');
+      expect(messages[0].images).toHaveLength(2);
+    });
+  });
 });