devswha/patina v8.1.0

v8.1.0 — Scorer prompt-injection hardening, KO research completion

on GitHub

8.1.0 — 2026-09-01

Scorer prompt-injection hardening and the completed Korean performance research program.

Semver rationale: minor — a shipped-security fix plus new research modules; no CLI or API contract changes.

Fixed

• Meaning-scorer prompts (AI-likeness, MPS, fidelity) now fence untrusted source/rewrite text with the fenceReferenceText data-fence contract instead of plain header interpolation, closing a prompt-injection surface on every language and request (#709). Validated live: fenced MPS calls return unchanged-shape JSON.
• assertIndependentJudge in the rewrite A/B harness rejects same-provider-family pairs regardless of transport shape (HTTP+HTTP same-family pairs were previously skipped).

Added

• Deterministic Korean research infrastructure landed from the completed preregistered program (#708): src/features/korean-diagnosis.js, korean-invariants.js, korean-structure-fingerprint.js, the 120-row confirmatory corpus and mutation fixtures, and their unit suites. The recorded verdict is NOT PROMOTED — production stays iterative-baseline; the treatment remains behind PATINA_KO_DIAGNOSIS_RESEARCH.