fix(agent): keep parallel tool-result messages contiguous on OpenAI Chat (Databricks image fix)

[tlongwell-block]

What

Fix tool-result framing for parallel image-returning tool calls on the OpenAI Chat path so requests via Databricks model serving stop being rejected.

Why

Reported by Wes in #sprout-bugs: agents using Databricks-routed Claude (Opus 4.6/4.7) fail when asked to view multiple images in a thread. Error from Databricks is Anthropic-shaped, despite the agent talking OpenAI Chat:

llm: 400 Bad Request: tool_use ids were found without tool_result
blocks immediately after: toolu_bdrk_017tcizkcWrCFzQGQyci3DVr.
Each tool_use block must have a corresponding tool_result block
in the next message.

The toolu_bdrk_ prefix gives it away — Databricks (config.rs:119: OpenAiApi::Chat, // Databricks invocations is chat-shaped) translates OpenAI Chat back into Anthropic on the way to the model and the translation breaks on this pattern.

Root cause

openai_body emitted images as a role:"user" message immediately after each tool result's role:"tool" message. With one image-returning tool call that's fine. With two parallel calls — exactly the case in Wes's screenshot, two stacked sprout-mcp__view_image invocations — the wire becomes:

role:"tool"  (A text)
role:"user"  (A image)     ← splits the run
role:"tool"  (B text)      ← Databricks: where's B's tool_result?
role:"user"  (B image)

Databricks' OpenAI→Anthropic translator folds consecutive role:"tool" messages into one Anthropic user message of tool_result blocks. The intervening role:"user" (image) ends the run, so the second tool result lands in a separate user message, leaving tool_use B unpaired in the immediately-following user turn. Anthropic rejects.

Fix

In openai_body, defer image content into a pending_images accumulator while emitting role:"tool" messages. Flush as a single trailing role:"user" carrying every image from the batch before any non-ToolResult history item (or at end of history). Mirrors the existing Anthropic body's pending/flush pattern.

OpenAI Chat semantics unchanged: each tool's text result still lands in its own role:"tool" message in order; images still ride on a role:"user" message after their text results. Only the grouping changes — one user message per run of tool results instead of one per result.

Test

openai_parallel_image_tool_results_stay_contiguous constructs the two-parallel-images case and asserts both role:"tool" messages are adjacent with a single trailing role:"user" carrying both images. Confirmed failing on the prior implementation (7-message interleaved shape) before applying the fix.

All 45 sprout-agent lib tests pass. Clippy and fmt clean.

Scope

• Single file: crates/sprout-agent/src/llm.rs
• One function changed (openai_body); other paths (Anthropic native, OpenAI Responses) already handled this correctly and are untouched.