v0.14.0

Changelog

• bac3d6c Step 1: persistent warm state with authoritative startup reconciliation
• 3262abf Step 2A: bench subcommand - cold vs warm TTFT through mesh proxy
• c25dccd Step 3: Router Decomposition - Pure refactoring: split router.go into placement.go, health.go, queue.go and create interfaces.go
• 36c5f65 Step 4: Weighted Placement Scoring - Replace basic routing with multi-factor scoring, model pinning, and node cooldown
• 7bbcca6 Step 5: Predictive Prewarming - implement in-memory transition ring buffer, prediction worker, time-of-day pattern prewarming, and accuracy metrics
• 8b11fba Step 6: Correct Option 2 description in README.md to reference multi-runtime subnet probing and PROBE=1 flag
• c63a784 Step 6: Expand install.sh discovery scan to detect vLLM, TGI, and llama.cpp runtimes in addition to Ollama
• ade88eb Step 6: Move Quick Start section to the very top of README.md to optimize developer UX
• efb3ab5 Step 6: Update install.sh and README to support installing without probing, separated behind START=1/--start flags
• 7541568 Step 6: Update installer and Docker compose with Prometheus & Grafana telemetry provisioning
• 6ea186f Step 6: Update pages.yml concurrency to cancel in progress deployments
• 3a6cad9 Step 6: Update website landing page installation step with standard versus network-probe options
• 3bac673 Step 6: Upgrade Go toolchain to 1.25.11 to patch standard library vulnerabilities in govulncheck
• e8e07c0 Step 6: Upgrade install.sh to probe the entire local physical subnet for active Ollama nodes and register them in config.yaml
• 7351a3b WIP: persistent warm state before startup reconciliation fix
• 2faa290 feat(lifecycle): manual + scheduled model unload
• b699922 fix(ci): gofmt formatting + enforce LF line endings via .gitattributes
• cb54f84 fix(ui): Models header counted the whole catalog as warm models
• fab1371 fix: bug-sweep fixes from live + static review (analytics persistence, others)