glpi-neuroforge-mega/RELEASE-NOTES-v1.6.0.md

GLPI NeuroForge Mega v1.6.0

Durable Master/Subagent Orchestration

• NeuroForge is the authoritative scheduler/master.
• Separate CPU and GPU subagents with capability/resource routing.
• Persistent priorities, dependencies, idempotency, retries/backoff and bounded queues.
• Heartbeat/lease-token fencing; expired leases cannot submit late results.
• Durable apply_wait phase for worker results that mutate master state.
• Admin status/job list/retry/cancel and read-only Control Center visibility.

Knowledge Graph Convergence

• Bounded ANN-assisted graph backfill for imported knowledge memories.
• Multi-neighbor semantic relinking instead of leaving imported chunks isolated.
• Adjacency index, multiple relation labels per edge and bounded multi-hop retrieval.
• Graph health metrics for isolated/linked/multi-linked nodes, degree and connected components.
• Memory-version plus vector-fingerprint fencing rejects stale relink results after delete/recreate races.

Deployment

• Local neuroforge-worker-cpu and neuroforge-worker-gpu services.
• docker-compose.subagent.yml for additional remote CPU/GPU hosts.
• Non-destructive upgrade from v1.5.9; existing volumes and knowledge remain valid.