Arc: throttle self-drive and remove broken GPT review workflow

[connortessaro]

Summary\n- remove the broken Code Review (GPT) workflow that fails before running\n- default unattended Claude self-drive workers to Sonnet 4.6, with env override support\n- cover the new Claude model selection path in runtime tests\n\n## Verification\n- pnpm exec vitest run src/code-cockpit/runtime.test.ts -t "uses ARC_SELF_DRIVE_CLAUDE_MODEL for new strict-Claude self-drive workers"\n- pnpm exec vitest run src/code-cockpit/runtime.test.ts -t "prefers Claude when both engines are healthy"\n- pnpm exec vitest run src/code-cockpit/runtime.test.ts -t "routes Claude workers through the claude-cli backend with engine-specific model metadata"\n- pnpm exec vitest run src/code-cockpit/runtime.test.ts -t "uses the persisted thread id when sending a follow-up worker turn"\n- git diff --check\n\n## Notes\n- the full runtime test file still shows unrelated async cleanup/finalization flakiness when run as a whole in this local environment\n- the live repo workflow is also disabled manually so new PR pushes stop receiving the broken red code-review check immediately