Release v0.4.3

⚠️ Breaking Changes

• Vision routing now only triggers when the latest user message contains a new (unseen) image; historic images no longer cause automatic switches to vision models.
• Some models previously marked as vision‑capable have been reverted; they will no longer be used for image input unless explicitly re‑verified.

New Features

• Add Images field and NormalizedImage struct to preserve media type and base64 data across normalization.
• Extend model registry with three 1M‑context vision‑capable models: qwen3.7-max, qwen3.7-plus, mimo-v2-omni.
• Flip Vision: true for mimo-v2.5-pro, mimo-v2.5, and minimax-m3 based on empirical verification.
• Update recommended vision scenarios: vision → kimi-k2.7-code, vision_complex → mimo-v2-omni, vision_long_context → qwen3.7-plus.

Bug Fixes

• Preserve image data in the normalize/denormalize bridge by checking for non‑empty source data.
• Fix request routing to trigger vision only on new images, eliminating false positives from stale history keywords.
• Correct over‑eager vision flag for unsupported models (minimax-m3, qwen3.7-max).
• Streamline request facts by removing unused fields from scenario logic.

Improvements

• Enhance end‑to‑end reliability of image handling for 32×32 PNG test cases.
• Refine model metadata updates to avoid speculative vision flag changes.

Documentation

• Update internal/config/model_registry.go to reflect new vision‑capable entries.
• Adjust scenario test expectations to match the revised routing behavior.

Chores

• Update internal imports and struct definitions to accommodate new Images field.
• Clean up unused imports and dead code in normalization and scenario packages.