fix(agent): fix OpenAI-compat request body serialization and max_tokens

[tlongwell-block]

Problem

Two issues prevented sprout-agent from working with newer OpenAI models (e.g. gpt-5.5):

1. max_tokens rejected by newer models: OpenAI's newer models (gpt-5.5, o-series) reject the legacy max_tokens parameter and require max_completion_tokens instead.

2. Request body silently dropped: The post() helper used reqwest's .json(body) which sets Content-Type AND serializes the body. Callers then added .header("content-type", "application/json") via the closure, causing reqwest 0.13 to silently drop the request body. OpenAI returned "you must provide a model parameter" because it received an empty body.

Fix

• Replace max_tokens with max_completion_tokens in both OpenAI request paths (openai_body() and summarize())
• Switch post() to manual serialization via serde_json::to_vec() + .body() + explicit Content-Type header
• Remove duplicate content-type headers from all callers

Testing

• ✅ All 17 sprout-agent unit tests pass
• ✅ Full workspace cargo test passes (2 flaky pair-relay timing tests fail identically on main)
• ✅ cargo clippy --workspace clean
• ✅ cargo fmt clean
• ✅ Live tested with gpt-5.5: simple prompt + tool use (MCP shell)
• ✅ Anthropic regression test: claude-haiku-4-5 still works correctly

Notes

• max_completion_tokens is the standard for all current OpenAI models. Some older OpenAI-compatible servers (vLLM, Ollama) may still only accept max_tokens — if that's a concern, a follow-up could add a config knob or send both.