v3.19.0

ruvnet/ruflov3.19.0Jul 4, 2026by ruvnet

AI Summary

Routes the neural train command through the native TrainingPipeline, introducing a backend flag and ensuring checkpoints contain trained weights.

Key Highlights

New 'neural train --backend auto|native|wasm' routing flag
Native pipeline supports real epochs, loss history, and early stopping
Checkpoints now carry actual trained weights instead of untrained ones
Graceful WASM fallback when native module is absent

New Features

Native TrainingPipeline integration for LoRA training
--backend flag for training routing
Native backend with loss history and early stopping

Full Release Notes

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