fix(qwen35): emit images inline in tool-response content

[eligotts]

Summary

• Qwen35Renderer._render_tool silently dropped image / video parts in tool messages by pre-stringifying content via _render_content. Tool-response screenshots (the dominant image source in browser-agent traces) never made it into rendered tokens or mm_items, even though the Qwen3.5 chat template emits <|vision_start|><|image_pad|><|vision_end|> inside <tool_response> blocks just as in user messages.
• Unify _render_tool to handle both text-only and media-bearing tool content in a single method. Internal _content_has_media dispatch; text-only path keeps the prior single-encode emit_text("\n" + content + "\n") for BPE parity; media path reuses the buffer-and-flush pattern from emit_user_with_media with emit_image threaded through as a closure arg.
• Consecutive-tools envelope (<|im_start|>user>...<|im_end|>) logic stays in one place and remains content-agnostic (prev_is_tool / next_is_tool only). Consecutive tool messages mixing media + text-only render under one shared envelope — same as apply_chat_template.
• Qwen36Renderer inherits the fix via subclassing.

Other renderers

• Qwen3VLRenderer already had the equivalent abstraction (render_media_content reused across user and tool); verified live via parity probe on Qwen3-VL-4B (106 tokens, exact match). No change needed.
• KimiK25Renderer._render_tool_body has the same class of bug (silently drops via emit_image=None default in _emit_content); confirmed via live probe (0 vs 1 <|media_pad|> token). Out of scope here — separate follow-up PR.

Test plan

• New test_tool_response_image_byte_parity asserts byte-identity vs processor.apply_chat_template + processor(images=, text=) for three cases:
  • tool_response_with_image — single tool image
  • multi_turn_tool_response_images — multiple turns each carrying an image
  • consecutive_tools_mixed_media — consecutive tools mixing media + text-only (the wrapper-boundary case)
• Gated by isinstance(renderer, Qwen35Renderer) so other VLM renderers skip until they grow the feature.
• Passes locally on Qwen/Qwen3.5-0.8B, Qwen/Qwen3.5-9B, and Qwen/Qwen3.5-35B-A3B.
• Full renderer test suite: 964 passed, 75 skipped, 1 xfailed (pre-existing). No regressions.

Known gaps

• Test coverage is image-only. Video parts in tool content take the same NotImplementedError branch they do in emit_user_with_media, but I haven't validated <|vision_start|><|video_pad|><|vision_end|> parity end-to-end.

🤖 Generated with Claude Code