fix(runtime): bound OpenAI Chat image tool results

[Dante-dan]

Summary

• prevent OpenAI Chat wires from lowering image Tool Results into textual base64
• preserve the matching assistant tool call and tool result with a bounded text fallback, while leaving user-message and native image paths unchanged
• apply the same projection rule to current and durable replay, with request-body regression coverage

Fixes #5277

Verification

• npm run lint — 3,569 files checked
• npm run format:check — 2,108 files checked
• npm run build — all workspaces passed
• npm run typecheck — all workspaces passed after the dependency build
• npx knip --workspace apps/desktop
• npx knip --workspace packages/ui
• node --test packages/runtime/dist/__tests__/ai-sdk-backend.test.js — 240 passed
• git diff --check

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Codex traced the protocol boundary, implemented the wire-aware Tool Result projection, and authored the request-body regression.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[hqhq1025]

Reviewed exact commit 32ac7ddf7338c2d1ffdfc63db9a51541d9a42899. I found no P0-P3 issues.

The change keeps image Tool Results off the OpenAI Chat wire before artifact hydration, emits a bounded string fallback, and preserves the assistant tool-call/result identity. The durable projection path applies the same rule to live continuation and restart replay, while user-message images and image-capable non-Chat protocols remain unchanged.

I verified the production OpenAI-compatible request body: the tool result is a string, the matching tool call ID remains present, the image bytes are absent, and the artifact reader is not invoked. The real Host/store path covers the live turn, a later turn, and reopen replay. Checks passed: build:test, full typecheck/lint/format, ASF headers, git diff --check, Runtime 3472 passed / 13 skipped, focused Runtime/Runtime Host image-path tests, hosted test, and a clean merge with current main (f109ccde). Runtime Host full tests were 1917 passed / 19 skipped / 1 failed; the sole failure is the unchanged managed-Bash sandbox integration because this runner rejects both unshare and bwrap user namespaces. I did not call a real external OpenAI-compatible provider or exercise native Windows/macOS.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.