v7.1.4 — Preserve model output strings

Red-team judging now passes the complete output string to Jev unchanged, including nested JSON or serialized messages. Prompt-envelope normalization remains separate. Non-string response values reject ingestion; replay hashes bind to the complete response string.

Validation: 2,026 tests passed; typecheck, lint, format, build and release contracts passed. All 4,362 supplied outputs are preserved verbatim and match the Python skill across response hashes. No live Jev accuracy claim is made.

Published through next; stable latest remains 7.0.1. This supersedes 7.1.3 response-envelope extraction.