Preflight unsafe world templates

[YSCJRH]

Summary

• Add pre-write safety preflight for create-world template payloads.
• Expand redline terms for real-person/digital twin, political persuasion, real-world prediction, hidden surveillance, and high-risk decision requests, including Chinese examples.
• Add regression tests proving unsafe create-world payloads are rejected before state/world artifact directories are written.

Test Plan

• python -m pytest backend/tests/test_cli.py::test_cli_create_world_blocks_unsafe_payload_before_writing_state -q
• python -m pytest backend/tests/test_cli.py -k "create_world or safety" -q
• python -m pytest backend/tests/test_cli.py -q
• python -m pytest backend/tests -q
• python scripts/check_no_secrets.py
• git diff --check

Safety Notes

• Treating this as protected-core / risk:safety because it changes create-world redline behavior.
• No Hosted GPT, BYOK, upload, auth, billing, database, object storage, quota, public API, plugin MCP contract, scenario DSL, claim labels, run trace shape, or artifact layout changes.

Fixes #368