Preserve independent images in multimodal prompts - #17

Adjacent, same-size image arguments can be merged into a temporal image chunk by the native Qwen backend. A PAW function trained on independent images then receives a different representation, despite having the correct prompt and weights.

Tokenize each image occurrence separately and evaluate the resulting chunks in order. Preserve all literal prompt text, whitespace, image order and repeated references; do not insert separators or wrap the prompt in another template. Text-only and single-image calls retain their existing behavior.

Validation:

• The complete hermetic test selection from .github/workflows/test.yml passes locally on Python 3.12.
• The adjacent-image regression fails against unpatched main and passes with this change. Existing tests cover exact prompt text, repeated/reordered images, text-only calls, context limits and cleanup on failures.
• Native Mac/Qwen3.5-0.8B diagnostic with unchanged weights, prompt and 220 reviewed image pairs: exact matches improved from 74/220 to 216/220. This is a training-fit/runtime discrepancy check, not held-out accuracy or gameplay validation.

Only the runtime fix and its tests are included. No version bump or package release.