fix: sanitize pending/running tool parts to prevent orphaned tool_use

[BYK]

Problem

When a session errors mid-tool-execution (e.g. context overflow), the tool part stays in pending/running state. On the next transform(), the OpenCode SDK generates a tool_use block for this part but has no output to generate a matching tool_result, causing Anthropic to reject with:

messages.685: tool_use ids were found without tool_result blocks immediately after: toolu_017XWb5fnou7kqTt3CjSfiwc

Root Cause

In the SDK's ToolState discriminated union:

• ToolStatePending / ToolStateRunning → only generates tool_use (no output → no tool_result)
• ToolStateCompleted / ToolStateError → generates both tool_use + tool_result

Lore's gradient transform had no sanitization step for incomplete tool parts. The trailing drop loop preserved messages with ANY tool part regardless of status, which was correct for preventing the prefill error but didn't address the orphaned tool_use issue.

Fix

Add sanitizeToolParts() in gradient.ts that converts pending/running tool parts to error state after transformInner(). This:

• Generates both tool_use + tool_result(is_error=true), eliminating the orphan
• Preserves conversation history — the model sees the tool was attempted but failed
• Is a no-op (returns same array reference) when all tools are in terminal state (common case)
• Runs at all layers (0-4), covering every code path

Changes

• src/gradient.ts: Add sanitizeToolParts() function and integrate into transform()
• src/index.ts: Update trailing drop loop comment to note sanitization happens upstream
• test/gradient.test.ts: 6 new test cases covering pending/running/mixed/no-op scenarios