feat(agent): paste images into the agent composer

[Zen0-99]

Summary

• Let users paste images (screenshot / mockup) into the agent composer. Each pasted image becomes a base64 data-URL attachment sent as a user-message image block.
• Types: added AgentImageAttachment, an image variant on AgentMessageBlock, and images? on AgentRequestBody.
• Ingest: sendAgentMessage(content, images?) builds user-message blocks (text + one image block per attachment) and POSTs images; rehydrateMessages now materialises image blocks so they survive reload.
• Server (chat.ts): images added to the request schema (allow-listed MIME, ~14MB base64 cap, max 10), prompt made optional with a >=1 of text/image guard, and appendMessage persists text + image blocks.
• UI (AgentPanel): onPaste captures image files (rejecting unsupported MIME / >10MB with a toast), a pending-attachment thumbnail tray with remove buttons, a soft warning when the selected model lacks visionInput, and inline thumbnail rendering of pasted images in the thread.
• Test: conversationRoundtrip asserts an image block persists and replays via appendMessage / buildMessageHistory.

Why

The model already receives images today via the render_snapshot tool funnel (screenshots are pushed as image content the drivers map to native multimodal). Pasting is a second, parallel funnel into the same user-message image-block channel, so no driver/provider work was needed: all three drivers already map {kind:'image'} to native vision content. The only genuinely new server code is the appendMessage augmentation in chat.ts; the rest is ingest + UI. This makes it trivial for a user to show the agent a design and ask about it.

Test

• bun test src/__tests__/agent/agentSlice.test.ts -> 19/19 pass.
• bun test src/__tests__/ai/conversationRoundtrip.test.ts -> image persistence/replay assertion passes (the only failures on Windows are a pre-existing EBUSY temp-dir cleanup lock in the shared test helper; passes on Linux CI).
• tsc --noEmit clean; ESLint clean on all changed files.
• Manual: bun run dev, open the agent panel, paste an image, send with a vision model (e.g. Claude), confirm the model references it; reload and confirm the thumbnail reappears.

[Zen0-99]

@DavidBabinec would you be open to collaborating? I'm a UI/UX designer so most of the work I do here is front end (there is one big one that is backend though).