1.5.6

docs/GO-LIVE-v1.5.0.md

@@ -36,3 +36,13 @@ Vor Go-Live mit autonomem Research zusätzlich verifizieren:
 4. `research_authoritative_sources >= 1` und `quality_gate_version=staging-v2` sind im Draft vorhanden.
 5. Ein absichtlich nicht belegter Versions-/Errorcode im Synthese-Test wird fail-closed abgewiesen.
 6. Ein Blog-/Forum-only Evidence-Set erzeugt bei aktiviertem Authority-Gate keinen Staging-Artikel.
+
+
+## v1.5.6 Structured-Output Zusatzgate
+
+Für einen realen Research→Staging-Smoke-Test zusätzlich verifizieren:
+
+1. Ein Windows-/Registry-lastiges Testziel erzeugt keine `invalid ... string escape code`-Fehler.
+2. Staging-Synthese und Claim-Verifikation bleiben bei nicht reparierbarem JSON fail-closed.
+3. Der resultierende Draft enthält weiterhin `human_review_required=true` und `auto_reply=false`.
+4. Source-Authority und Claim-Verifikation aus v1.5.5 bleiben bestanden; JSON-Robustheit darf diese Gates nicht umgehen.

docs/MIGRATION-v1.5.5-to-v1.5.6.md

@@ -0,0 +1,17 @@
+# Migration v1.5.5 → v1.5.6
+
+v1.5.6 is a drop-in structured-output hardening release. No storage migration is required. Existing goals, memories, synapses, research history and staging drafts remain intact.
+
+1. Build/publish the v1.5.6 images through the normal Gitea pipeline.
+2. Set `IMAGE_TAG=1.5.6`.
+3. Pull and recreate NeuroForge and its worker:
+
+```bash
+docker compose --profile research pull neuroforge neuroforge-worker
+docker compose --profile research up -d --force-recreate neuroforge neuroforge-worker
+```
+
+4. Do not delete volumes.
+5. Re-run the previously blocked goal. Existing evidence can be reused; a reset is not necessary.
+
+There are no new mandatory environment variables. Ollama JSON mode is selected internally only for structured-output calls.

docs/RELEASE-v1.5.6.md

@@ -0,0 +1,20 @@
+# GLPI NeuroForge Mega v1.5.6
+
+v1.5.6 hardens structured LLM output after a live BitLocker staging test exposed `invalid character \\ in string escape code` even after the syntax-repair pass. The failure was caused by local-model JSON strings containing literal Windows/registry/Markdown backslashes that were not valid JSON escapes.
+
+## Structured-output hardening
+
+- **Provider-native JSON mode:** Ollama structured calls for staging synthesis, syntax repair, claim verification and grounded rewrite now send `format: "json"`.
+- **Strict schema decoding:** structured objects reject unexpected fields instead of silently accepting arbitrary model-added keys; arbitrary prose surrounding an object is no longer stripped/accepted.
+- **Narrow deterministic escape repair:** when strict decoding fails, NeuroForge may preserve a literal backslash by JSON-escaping it only when it occurs inside a JSON string and is followed by a character that JSON does not define as an escape.
+- **Valid escapes are untouched:** `\\`, `\"`, `\n`, `\t`, `\r`, `\/` and valid `\uXXXX` escapes retain their JSON semantics.
+- **Fail-closed remainder:** malformed quoting, control characters, multiple JSON values, invalid code fences, unknown schema fields and other syntax defects are not broadly normalized. They either enter the single bounded syntax-repair pass or fail staging.
+- **Shared verifier protection:** the same strict decoder/escape handling protects claim-verifier JSON and grounded rewrite JSON, not just the first synthesis response.
+
+## Production-quality gates retained
+
+All v1.5.5 source-authority, authoritative-action, claim-coverage, contradiction and human-review-only promotion gates remain enabled. This release changes structured transport/parsing robustness; it does not weaken evidence validation.
+
+## Regression coverage
+
+Tests cover Windows paths such as `C:\Windows\System32`, registry-like paths, a model-emitted backslash before a Markdown backtick, unknown schema fields, Markdown fences, missing outer braces, one bounded repair call and Ollama native JSON mode.