fix(core): update recorded prompt cache key

[kitlangton]

What

Restores the recorded SessionRunnerLLM test by updating its OpenAI Chat cassette for the prompt cache key now sent by the runner.

Before / After

Before: 6f3a3600b904c6a14c01e7e675acf435de0150c1 began forwarding LLMRequest.promptCacheKey as prompt_cache_key. The session runner supplies its session ID, but the existing cassette was recorded without that field. Replay rejected the request as a mismatch and the session drain waited until Bun's 5-second test timeout.

After: the cassette includes prompt_cache_key: "ses_runner_recorded", matching the production request shape and allowing the recorded response to complete the drain.

How

• Updates packages/core/test/fixtures/recordings/session-runner/openai-chat-streams-text.json with the request field introduced by the culprit commit.
• Fixes the fixture rather than reverting prompt-cache forwarding because that behavior is intentional and independently covered in packages/ai.

Scope

This PR only fixes the recorded transport regression. The independent LocationServiceMap timeout and TUI CI races are being handled separately.

Testing

• cd packages/core && bun typecheck passes.
• cd packages/core && bun run test test/session-runner-recorded.test.ts passes: 2 tests, 0 failures.
• cd packages/core && bun run test passes the recorded runner test. The full suite still reports the independent LocationServiceMap > keeps flush open while later hot reload runs timeout and the documented local-only PluginSupervisor config > logs invalid packages and continues loading failure caused by global user plugins.
• The repository pre-push typecheck passes all 33 configured package tasks.