release: v3.19.2 — stale codebase-graph freshness gate

Bumps manifest + marketplace + pyproject to 3.19.2 and adds release notes for
the freshness gate (b9c56c8): find_cached_graph now re-analyses when the
codebase changed since the graph was built, so /cortex-visualize stops serving
a stale AST. Pairs with the 3.19.1 deadlock guard.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: cdeust

.claude-plugin/marketplace.json

@@ -6,14 +6,14 @@
   },
   "metadata": {
     "description": "Persistent memory and cognitive profiling plugins for Claude Code",
-    "version": "3.19.1"
+    "version": "3.19.2"
   },
   "plugins": [
     {
       "name": "cortex",
       "source": "./",
       "description": "Persistent memory and cognitive profiling for Claude Code — thermodynamic memory with heat/decay, intent-aware retrieval, biological plasticity, codebase intelligence, and cognitive profiling. 49 MCP tools with enriched schemas. PostgreSQL + pgvector in CLI mode; automatic SQLite fallback in Cowork/sandboxed mode. v3.17.0 — autonomous per-project wiki: SessionStart auto-spawns a 6-hour consolidate cycle; a headless `claude -p` worker drains the curation-gap queue, calls codebase-intelligence MCP tools to ground each section in the real call graph, and authors missing anchor pages (architecture / services / api / data-flow / operations / decisions / PRD) per project from the source tree. 15 canonical scopes × 13 file sections; per-project dashboards under `wiki/_dashboards/`. Mermaid diagrams have a 🔍 lens with zoom + pan. Workflow graph with caller-qualified CALLS chains rendering full method-to-method dependencies (native tree-sitter, no AP required). Side panel humanized for non-technical users. Ingests codebase analysis (ai-automatised-pipeline) and PRDs (prd-spec-generator) into wiki + memory + knowledge graph. Docker image available.",
-      "version": "3.19.1",
+      "version": "3.19.2",
       "author": {
         "name": "Clement Deust",
         "email": "admin@ai-architect.tools"

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "cortex",
   "description": "Persistent memory for Claude Code — remembers across sessions automatically. Install and forget. Scientific retrieval backed by 41 published papers.",
-  "version": "3.19.1",
+  "version": "3.19.2",
   "author": {
     "name": "Clement Deust",
     "email": "admin@ai-architect.tools"

docs/release-notes/v3.19.2.md

@@ -0,0 +1,42 @@
+# v3.19.2 — stale codebase-graph gate
+
+Follow-up to the 3.19.1 ingest deadlock fix. The deadlock guard stopped the
+hang; this stops the *stale graph* it could leave behind.
+
+## What it fixes
+
+`/cortex-visualize` (and `ingest_codebase`) could serve an **out-of-date AST**:
+the cached code-graph was reused as long as the graph directory merely *existed*,
+with no check that the codebase hadn't changed since it was built. A graph built
+before your last edits was reused indefinitely — the "graph shows an old version
+of the analysis" report.
+
+## The change
+
+New `graph_is_fresh(project_path, graph_path)`: compares the graph artefact's own
+build mtime against the newest source file under the project, via a bounded,
+ignore-pruned walk that early-exits on the first newer file (vendored/build dirs
+— `.git`, `node_modules`, `target`, `.venv`, `search_index`, … — are skipped so
+they can't trip false staleness). `find_cached_graph` now skips a stale (or
+non-materialised) graph, so ingestion falls through to a fresh `analyze_codebase`.
+It errs fresh on uncertainty (absent project root / unreadable graph) to avoid
+needless re-analysis. Scope: ingestion only — the visualizer's roster resolver is
+untouched.
+
+## Recovering an already-bad cache
+
+If a prior run left a *half-built* graph (recent, so the freshness gate treats it
+as current), force one clean rebuild:
+
+```
+ingest_codebase(project_path="<repo>", force_reindex=true)
+# or: rm -rf ~/.cache/cortex/code-graphs/<project>-*/
+```
+
+With the 3.19.1 deadlock guard, that re-ingest completes instead of hanging.
+
+## Still in progress
+
+The deadlock guard is a safety net; the architectural root (the pooled MCP reader
+bound to a throwaway per-call event loop) and symbol-read producer/consumer
+overlap are tracked in `tasks/ingest-ecosystem-fix-plan.md`.

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "neuro-cortex-memory"
-version = "3.19.1"
+version = "3.19.2"
 description = "Scientifically-grounded memory system based on computational neuroscience research"
 readme = "README.md"
 license = "MIT"