fix(renderers): bridge merge must not mutate prior multi_modal_data

[hallerite]

Summary

MultimodalRenderer.bridge_to_next_turn merged the carried-forward multimodal sidecar with dict(previous_multi_modal_data.mm_*) — which copies only the outer dict. The per-modality lists stayed shared with the caller, so appending this turn's items via .setdefault(modality, []).extend(...) mutated the caller's previous_multi_modal_data in place. A reused prior RenderedTokens / MultiModalData would silently grow extra placeholders / items / hashes after a bridge.

Fix: copy the per-modality lists ({k: list(v) for k, v in ….items()}) before extending.

Scope

Cursor Bugbot flagged this on the Inkling PR (#103) for KimiK25Renderer; the identical pattern was in Qwen3VLRenderer and Qwen35Renderer, so all three are fixed here. (Inkling itself was fixed in #103.)

Test

Extends the shared test_multimodal_bridge_extends_and_carries_mm_data with a (2b) assertion that the prior turn's sidecar counts are unchanged after bridging — it now runs green for all five multimodal checkpoints (Qwen3-VL, Qwen3.5, Qwen3.6, Kimi K2.5, Kimi K2.6). Full tests/test_multimodal.py: 85 passed, 3 skipped.

Comment-and-copy only; no token-stream or bridged output change.

🤖 Generated with Claude Code

---

Note

Low Risk
Targeted copy-before-extend fix with no intended change to token streams; regression covered by extended multimodal bridge test.

Overview
Fixes in-place mutation of the caller's previous_multi_modal_data when multimodal renderers merge carried-forward vision sidecars during bridge_to_next_turn.

Previously, merging used dict(previous_multi_modal_data.mm_*), which only shallow-copied the outer map—per-modality lists stayed aliased, so .setdefault(modality, []).extend(...) could append placeholders, items, and hashes to a reused prior RenderedTokens / MultiModalData. The change deep-copies each modality list ({k: list(v) for k, v in ...}) in Kimi K2.5, Qwen3.5, and Qwen3-VL before extending; bridged token output is unchanged.

Adds test (2b) in test_multimodal_bridge_extends_and_carries_mm_data asserting prior sidecar counts stay (1, 1, 1) after bridging.

^{Reviewed by Cursor Bugbot for commit a5f0278. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix bridge_to_next_turn mutation of caller-supplied previous_multi_modal_data

The bridge_to_next_turn method in KimiK25Renderer, Qwen35Renderer, and Qwen3VLRenderer was shallow-copying only the outer dict when merging prior multimodal sidecar data, causing .extend() calls to mutate the caller's original per-modality lists. Each renderer now uses {k: list(v) for k, v in ....items()} to copy the inner lists, isolating the merged copy from the caller's data. A regression test in test_multimodal.py now asserts that per-modality counts in previous_multi_modal_data remain unchanged after bridging.

^{Macroscope summarized a5f0278.}

[macroscopeapp]

Approvability

Verdict: Approved

Straightforward bug fix that replaces shallow dict copies with deep copies of inner lists to prevent mutation of caller data. The change is consistent across 3 renderer files, includes test coverage, and the author is fixing their own code.

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