fix(vision): use native image input and retain uploaded originals

[ericleepi314]

Description

A web image upload was deleted immediately after attachment even though its temporary path remained in the prompt. A later Read or vision_analyze call therefore failed with “No such image file.” Vision-capable main models also received the optional analysis tool and could unnecessarily send the same image to a second model.

• Preserve accepted uploads at their original resolution in the session's readable artifact directory before cleaning up the temporary upload. Persisted metadata points to the retained file; rejected uploads and failed writes do not leave accepted attachments.
• Omit vision_analyze from native vision model requests, deferred-tool hints, ToolSearch results, and nano prompt tool lists. Keep the configured fallback available for text-only models, including fusion's base model.
• If a stale conversation calls the tool after switching to a vision-capable model, return image pixels to that model without invoking another LLM.

The caption visible inside the reported preview is already present in the uploaded screenshot, along with its 09:15 timestamp. The web renderer continues to display the uploaded image bytes and a separate message bubble.

Validation

• 707 targeted Python regression tests and 62 subtests passed across query assembly, tool discovery, image ingestion, vision analysis, nano mode, fusion, and gateway controls.
• After adding an original-file size-limit check, all 56 image attachment control tests passed, including the new check against silent truncation.
• The upload integration test removes the gateway temporary file, verifies the preserved original bytes, then reads it through the existing permission checks and vision loader.
• Captured model requests verify that native vision models receive the image content without a vision-analysis tool schema or deferred-tool hint; model switching and stale-call behavior are covered.

New Feature Submissions

• Explained the change and why it should be included.

All Submissions

Notify authors: @agentforce314, @ericleepi314