1.5.4

docs/MIGRATION-v1.5.3-to-v1.5.4.md

@@ -0,0 +1,9 @@
+# Migration v1.5.3 → v1.5.4
+
+v1.5.4 is a drop-in NeuroForge staging-synthesis robustness update. Persistent volumes, existing goals, evidence, drafts, and secrets are retained.
+
+1. Build/publish the v1.5.4 images through the normal Gitea pipeline.
+2. Set `IMAGE_TAG=1.5.4`.
+3. Run `docker compose --profile research pull`.
+4. Recreate at least `neuroforge` and `neuroforge-worker`; no volume reset is required.
+5. Leave the existing failed goal active. Its next autonomous cycle can retry staging from the persisted relevant evidence.

docs/RELEASE-v1.5.4.md

@@ -0,0 +1,12 @@
+# GLPI NeuroForge Mega v1.5.4
+
+v1.5.4 hardens the LLM staging-synthesis boundary after a live Windows 11 DISM 0x800f081f test showed that a local Ollama model could return JSON wrapped in Markdown or malformed structured output, causing `invalid staging synthesis JSON: invalid character '`' looking for beginning of value`.
+
+## Changes
+
+- The staging prompt explicitly requires a JSON object whose first character is `{` and last character is `}` and forbids Markdown/code fences.
+- A single outer ` ```json ` fence is safely removed before decoding.
+- The narrow local-model defect of emitting JSON members without the outer object braces is normalized only when the output is recognizably JSON-like and contains the required `answer` member.
+- Any other syntax failure triggers exactly one syntax-only LLM repair pass. That repair is forbidden from adding, inferring, or correcting facts.
+- The repaired draft still has to pass the existing minimum-content and goal-relevance gates. If repair or validation fails, staging remains fail-closed and no article is published.
+- Regression tests cover fenced JSON, missing outer braces, and malformed structured output repaired on the second model call.