test(core): add request-driven runner scenarios

[kitlangton]

Summary

• add a request-driven RunnerScenario test DSL around the real LLMClient boundary
• migrate V2 runner tests from shared mutable response queues to linear next() / respond interactions
• preserve explicit resume boundaries, raw stream failure coverage, concurrency behavior, and replay assertions
• remove the order-dependent legacy runner LLM harness

Verification

• bun run test test/session-runner.test.ts test/runner-scenario.test.ts test/session-runner-recorded.test.ts (119 passed)
• bun run test test/session-runner.test.ts --randomize --rerun-each 3 (315 passed)
• bun run test test/runner-scenario.test.ts --randomize --rerun-each 100 (1300 passed)
• bun typecheck in packages/core
• pre-push workspace typecheck (31 packages passed)

[kitlangton]

Closing this version. The request-driven API improved isolation, but the migration increased total test code instead of making the runner suite materially smaller. I’m replacing it with a much smaller scripted fixture and limiting interactive controls to timing-sensitive tests.