fix(providers): preserve OpenAI-compatible reasoning history

[code-yeongyu]

Summary

• repair OpenAI-compatible send/stream model handling so provider-prefixed routes send the stripped downstream model while internal events keep the original model
• persist non-streaming Thinking output blocks into CLI assistant events so reasoning history survives the next turn
• isolate two runtime tests that were failing from real global session state / invalid git metadata fixtures

Validation

• cargo fmt --all --check
• git diff --check
• cargo test -p api deepseek_v4 -- --nocapture
• cargo test -p api
• cargo test -p runtime
• cargo test -p rusty-claude-cli response_to_events_renders_collapsed_thinking_summary -- --nocapture
• cargo clippy -p api --all-targets (passes with pre-existing warnings)
• cargo test --workspace

Manual QA

• cargo run -p rusty-claude-cli -- --help rendered the CLI help surface successfully

Notes

• The first commit is the provider/CLI behavior fix.
• The second commit is test isolation only, keeping the behavior fix independently reviewable.

[1716775457damn]

Preserving reasoning history for OpenAI-compatible providers is important for multi-turn conversations. This + #3236 should fix the build break from #3235. Looks solid.