fix(vision): make image analysis streaming opt-in

[qiuYliangM]

Summary

• Preserve the existing non-streaming request shape unless [vision_model].stream = true is explicitly configured.
• Add a bounded total request budget and an option to disable the existing transient-error retry policy for local process endpoints.
• Reuse shared retry, Retry-After, error sanitization, and bounded error-body handling instead of introducing a second retry implementation.
• Parse UTF-8 SSE chunks safely, cap response buffering, fall back to ordinary JSON when a compatible endpoint ignores streaming, reject empty analyses, and mark incomplete usable output as truncated.
• Document the three compatibility controls.

Scope

This revision deliberately omits app-specific system prompts, default prompts, temperature overrides, and output-token overrides. Those policies belong in the application; the foundation change only supplies reusable transport and failure semantics.

Testing

• cargo fmt --all -- --check
• cargo test -p codewhale-tui vision::tools::tests -- --nocapture (19 passed)
• Affected-crate all-target Clippy passed with the target branch existing lint categories allowed.
• cargo test --workspace --all-features --locked (10,145 passed; one unrelated target-branch test failed: runtime_api::tests::skill_lifecycle_uninstall_removes_installed_skill. The same test failed again in isolation before reaching the uninstall step.)

The unmodified strict Clippy command is also blocked by 32 existing target-branch warnings in unrelated files. No warning points to this change.

Checklist

• Updated docs or comments as needed
• Added or updated tests where relevant
• Preserved the default request wire shape
• Harvested/co-authored credit uses a GitHub numeric noreply address
• No UI behavior changed

Known risks

• Streaming assumes the OpenAI-compatible data: event shape. Endpoints that ignore streaming and return ordinary JSON are handled explicitly.
• Downstream app integration that constructed the former app-specific configuration fields must keep those policies in the app and update its config construction.

No-Issue: hardens the existing vision tool for slow and local OpenAI-compatible endpoints without expanding foundation policy.

[github-actions]

Thanks @qiuYliangM for taking the time to contribute.

This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered.

Please read CONTRIBUTING.md for the expected contribution shape. A maintainer can grant recurring PR access by commenting /lgtm on a pull request.