fix(llm/messages): normalize Claude Code in-conversation system messages

[yalindogusahin]

Summary

Claude Code 2.1.157+ injects in-conversation system reminders inside the Anthropic messages array (messages[*].role == "system"), which is not part of the Anthropic Messages API spec — system is a top-level field, and messages[].role is limited to user/assistant. Strict spec implementations like vLLM's Anthropic endpoint reject the payload before request conversion (see vllm-project/vllm#44048), and the same payload also fails agentgateway's typed Role enum during messages → completions / bedrock translation.

This PR normalizes the request at the proxy entry point so every downstream conversion path sees a canonical Anthropic shape:

• drain messages[*].role == "system" entries from messages
• append their text content to the top-level system field, preserving any existing top-level system content first
• keep unknown roles rejected by the existing strict typed conversion

Changes

• crates/agentgateway/src/llm/types/messages.rs: add Request::normalize_system_messages() method handling both string and array ContentBlock forms; merges extracted content into a TextBlock::Array after existing top-level system. 6 new unit tests cover extraction, preservation order, array content blocks, no-op paths, and a regression test that exercises the actual completions::from_messages::translate call with a Claude Code 2.1.157+ shaped payload.
• crates/agentgateway/src/llm/mod.rs: call normalize_system_messages() inside LLM::process_messages_request after the request body is parsed and before process_request dispatch. Single chokepoint covers OpenAI, Bedrock, Vertex, and Anthropic-native passthrough.
• crates/agentgateway/src/llm/conversion/completions.rs, crates/agentgateway/src/llm/conversion/bedrock.rs, and RequestType::to_anthropic in messages.rs: defensive normalize_system_messages() calls in the direct library entry points so callers that bypass the proxy entry point (tests, embedders) get the same compatibility.

Trade-off (out of scope here)

The merge loses positional information: every in-conversation system block is appended after existing top-level system content. Preserving position would require provider-aware translation — OpenAI chat completions natively supports multiple system messages mid-conversation, Bedrock could use system blocks, Anthropic native cannot. A future opt-in flag (e.g. preserve_in_conversation_system) could expose this for backends that benefit from it. For now, normalizing into the standard Anthropic system field matches the same approach vLLM took in vllm-project/vllm#44048 and gives the least surprising baseline behavior.

Test Plan

• cargo build -p agentgateway (passes locally)
• cargo test -p agentgateway llm::types::messages::normalize_system_messages_tests — 6/6 passed locally
• CI on this PR

[Syraxius]

Yes actually I've tried two things before I decided to go with the merging into system (sacrificing the positional part).

1. Inject system messages wherever it was
2. Change the system messages into user messages

Both ways completely messed up the models behavior (I was testing with Qwen series) so I decided to go with the the merging.

I've inspected Claude's midway system messages - they're mainly to inject skills information, which is completely fine to be in system for now.

[yalindogusahin]

Thanks @Syraxius. Anything you'd change in this PR? Naming, normalization placement?