claude-flow

The final piece of the #2549 saga (reported by @pacphi): neural train now trains through @ruvector/ruvllm's native TrainingPipeline, not only the WASM path.

New

• neural train --backend auto|native|wasm — auto (default) routes the LoRA training leg through the native pipeline when @ruvector/ruvllm resolves: real epochs, loss history, early stopping, EWC registration
• Checkpoints carry trained weights — the previous best-effort block saved a freshly-constructed adapter's untrained weights; the native path checkpoints the pipeline that actually trained (ruvllm-checkpoint v1 envelope on disk)
• --backend native fails loudly when the pipeline can't run; --backend wasm preserves prior behavior exactly; SONA/ReasoningBank persistence unchanged in all modes; graceful WASM fallback when the module is absent

The complete #2549 arc

1. 3.18.1 — neural status no longer misreports the native path (dead variable + cross-process global)
2. @ruvector/ruvllm 2.5.7 — saveCheckpoint(path) actually persists (upstream, RuVector#637)
3. 3.18.2 — version floors + mkdir fix + version-gated capability reporting
4. 3.19.0 — training itself routes through the native pipeline (this release)

E2E: auto → native backend, finalLoss 4.25e-3, trained checkpoint on disk. Tests 6/6.

PR: #2556