test: record real gateway tool loop cassettes (OpenAI + vLLM)

[ashwing]

Summary

Follows up on the review thread on #83 — replaces the hand-authored composed cassette (which merged with #83) with real recorded traffic, per @maralbahari's ask to keep cassettes a true source of data rather than stub values.

Recorded via the repo's own record_cassette.py against two backends — OpenAI (gpt-4o) and local vLLM (openai/gpt-oss-20b) — mirroring the dual-backend approach in #77. Every turn is real typed SSE driven through the streaming loop; the recorder masks auth headers, so no credentials are in the cassettes.

Three distinct loop paths, each a dispatch_loop_cassette_test case:

Path	Backends	What it exercises
web_search → codex namespace across turns	OpenAI + vLLM	gateway Continue → client RequiresClientAction; flat agentic_ns__… restored to {namespace, name}
multi-round gateway	OpenAI + vLLM	web_search on two consecutive rounds then a final message — the only coverage of the loop's Continue decision firing more than once
same-turn mixed	OpenAI only	web_search + namespace in one model output → gateway executes and the turn hands back in a single round (client-owned precedence)

The same-turn path is OpenAI-only by necessity: gpt-oss-20b does not emit both tools in one turn, so a real vLLM recording of it isn't capturable (documented on the test).

Test Plan

• cargo test --workspace — 334 pass, 0 fail (8 in dispatch_loop_cassette_test)
• cargo clippy --all-targets -- -D warnings — clean
• cargo fmt --check — clean
• Cassettes recorded against live OpenAI + vLLM; verified no API keys present in the recorded YAML