Fix previous mm data copy

[cdreetz]

Summary:

dict(previous_multi_modal_data.mm_items) copies only the outer dictionary, but the per-modality values are still the exact same list objects, so calling .extend(...) on merged_items["image"] also mutates previous_multi_modal_data.mm_items["image"].

---

Note

Low Risk
Localized fix to multimodal merge logic in two renderers plus a regression test; no auth, API, or inference-path changes.

Overview
Fixes a shallow-copy bug in bridge_to_next_turn for Qwen3.5 and Qwen3-VL when merging prior-turn MultiModalData with the new turn.

Merging used dict(...) on mm_hashes, mm_placeholders, and mm_items, so per-modality lists stayed aliased to previous_multi_modal_data. Appending with .extend() mutated the caller’s prior snapshot—problematic when trainers keep per-step RenderedTokens for loss reconstruction.

The bridge now builds fresh outer dicts and copies each modality’s list (list(vals) per key) before extending with new-turn entries.

Adds test_multimodal_bridge_does_not_mutate_previous_mm_data across the multimodal matrix: prior lists unchanged after bridge, and bridged inner lists are not the same objects as the prior turn’s.

^{Reviewed by Cursor Bugbot for commit 2a8fba0. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix mutation of previous turn's multimodal data in bridge_to_next_turn

• Replaces shallow dict(...) copies of mm_hashes, mm_placeholders, and mm_items with dict comprehensions that copy each per-modality list in qwen35.py and qwen3_vl.py.
• Without this fix, merging new turn data would mutate the previous MultiModalData's lists due to shared list references.
• Adds a parametrized regression test in tests/test_multimodal.py that asserts previous turn mm data is neither mutated nor aliased after bridging.

^{Macroscope summarized 2a8fba0.}

[macroscopeapp]

Approvability

Verdict: Approved

Straightforward bug fix converting shallow copies to deep copies to prevent unintended mutation of multimodal data during turn bridging. The change is small, isolated, and includes a comprehensive regression test.

^{You can customize Macroscope's approvability policy. Learn more.}