v7.1.3 — Normalize AIRS response envelopes for Jev

Superseded for red-team judging by v7.1.4, which follows the confirmed contract: preserve the complete model output string, including nested JSON. Version 7.1.3 unwrapped recognized response envelopes. Use 7.1.4 for the corrected behavior.

AIRS exports store target replies inside A2A message envelopes, sometimes as Python-style dictionary strings. The judge now extracts the ordered text parts before sending them to Jev, preserving literal JSON attack content and excluding transport IDs.

Empty or unsupported non-text envelopes remain visible in coverage. Old recordings with wrapper-based response hashes are rejected against corrected inputs; run a fresh recorded probe in a new output directory.

Validation: 2,013 tests passed; typecheck, lint, formatting, release contracts and build passed; public production dependency audit found no known vulnerabilities. Both implementations produce identical normalized hashes for all 4,362 supplied records, with 36 oversized inputs still excluded. No live Jev accuracy claim is made.

This is a preview release on next; the public latest tag remains unchanged. A matching harness package is being prepared separately because the native skill is embedded in the harness binary.