### What happened?

When WebSearch Interception is used for **Bedrock** Anthropic Claude models (which have no native web search), the **agentic-loop** path emits `web_search_tool_result` blocks whose `tool_use_id` is the model's original client tool id (`toolu_...`), not a `srvtoolu_`-prefixed server-tool id. Single-turn works, but on **multi-turn** requests (round 2+) the client replays the prior assistant turn containing that block, and the Anthropic Messages schema (which Bedrock inherits via the AnthropicBedrock shape) rejects it:

``` messages.1.content.0.web_search_tool_result.tool_use_id: String should match pattern '^srvtoolu_[a-zA-Z0-9_]+' ```

This makes multi-turn web search (e.g. Claude Code's normal usage) fail with a 400 on the second turn.

### Root cause

In `litellm/integrations/websearch_interception/handler.py` there are two paths that emit native `web_search_tool_result` blocks, and they handle `tool_use_id` inconsistently:

**Short-circuit path (correct)** — generates a `srvtoolu_` id and pairs it with a `server_tool_use` block:

```python # handler.py (v1.90.0) ~line 190 tool_use_id = f"srvtoolu_{uuid.uuid4().hex}" ... content.append({"type": "server_tool_use…