server: support Responses assistant message phases

[inf9144]

Summary

Add OpenAI Responses assistant message phase support for Codex CLI 0.147 (commentary / final_answer) while keeping Chat Completions behavior unchanged.

• preserves Responses assistant phase metadata through historical replay
• scopes Qwen phase markers to the dedicated Codex template via llama.cpp:responses-phase=marker-v1
• consumes the marker as parser metadata so it cannot leak into visible content/reasoning
• emits phase on non-streaming and streaming Responses message items
• falls back structurally for unmarked/non-participating models (tool_calls => commentary, otherwise final_answer)
• keeps shared Qwen3-Coder/Qwen3.5 parser behavior opt-in only

Validation

Source-only commit was produced from codex-compat (a407cfa8db0e9771e559d9bda621340fc4d58fbd) by the isolated validation workflow after all gates passed in run #21 (32181883538):

• patch application + git diff --check
• focused llama-server, test-chat, test-qwen38-codex-template build
• test-chat
• focused Qwen3.8 Codex regression suite, including partial-marker leakage and response.output_item.added phase coverage
• exact 9-file source-delta audit
• validated source commit publish

Validated commit: 0ce5908aac568eb316c54f8393d1c705aeee944a.

Codex reference semantics were checked against exact Codex CLI 0.147 commit be6e8eac029b183056b7e4402879f15d2c85f61b.