bug: parallel async tool execution races on shared ToolContext

[G9000]

Summary

execute_parallel (executor.py:187-203) runs async tools via asyncio.gather sharing the same ToolContext contextvar. Unlike sync tools (which get a contextvars.copy_context()), async tools all operate on the same ToolContext object.

Race Conditions

1. current_tool_call_id: Set at line 69/140 before each execution — whichever tool runs last wins. PendingMemoryOps created by concurrent tools may reference the wrong tool call ID.
2. memory_modified: _check_memory_modified (line 255-263) reads the flag and resets it to False. If tool A sets it and tool B's check clears it before tool A's result is processed, the memory refresh callback is skipped.

Fix

For parallel async execution, either:

• Scope current_tool_call_id per-call (pass it as an argument rather than via ToolContext)
• Create a context copy for each async tool invocation
• Track memory_modified per-result rather than via shared mutable state

Files

• apps/server/src/anima_server/services/agent/executor.py:187-203, 206-252

Severity

Medium — incorrect tool call ID attribution and missed memory refreshes under parallel tool execution.