fix(agents): swap edge order in singleRunStrategy for nodeSendToolResult

[rainbow7]

Problem

When an LLM returns both Text and ToolCall simultaneously in the same response (common behavior with DeepSeek, and some other models), the singleRunStrategy fails to execute the tool calls after receiving tool results. Instead, it treats the intermediate text as the final answer.

Root Cause

The edge order for nodeSendToolResult was:

edge(nodeSendToolResult forwardTo nodeFinish onTextMessage { true })     // first — matches Text
edge(nodeSendToolResult forwardTo nodeExecuteTool onToolCalls { true })  // second — never reached

Since resolveEdge evaluates edges in definition order (FIFO) and returns the first match, onTextMessage always wins when both Text and ToolCall are present.

Notably, the edges for nodeCallLLM (lines 35-36) already have the correct order with onToolCalls first. This inconsistency suggests the nodeSendToolResult order was unintentional.

Fix

Swap the two edges so onToolCalls is evaluated first (matching the nodeCallLLM pattern):

edge(nodeSendToolResult forwardTo nodeExecuteTool onToolCalls { true })  // first — execute tools
edge(nodeSendToolResult forwardTo nodeFinish onTextMessage { true })     // second — finish on pure text

Test

Added test_SingleRunStrategy_SecondRoundMixedResponse_ExecutesAllTools which reproduces the bug:

• First round: LLM returns a tool call (no text) → tool executes ✅
• Second round (after tool result): LLM returns Text + ToolCall simultaneously → before fix, tools were skipped; after fix, all tools execute ✅

This test fails without the fix and passes with it.