test(ai): refresh typed catalog cases

[quick-ricon]

Summary

• replace the stale Anthropic abort-test model fixture with a current hydrated reasoning model
• keep the OpenCode max_tokens assertion on a current openai-completions model
• refresh the adaptive-thinking fast-model fixture exposed by the first hosted full-suite run
• restore the merged-main checks that failed after feat(coding-agent): add turn boundary renderer hook #44

Validation

• npm run check
• focused Vitest for all three touched AI test files: 48 passed, 35 credential-gated skips
• focused Biome check
• git diff --check

The full ./test.sh suite was not rerun locally; the fork owner explicitly waived that upstream contributor requirement for this focused typed-catalog fixture repair. PR CI runs the full repository test workflow.