feat(forge): forge/train — drive unsloth LoRA training (genome loop, Phase 5 slice 1)

[joelteply]

The flywheel's training executor

Replaces the forge stub's "Phase 5+ train lands later" with the real train stage. This closes the train half of the coordination↔learning loop:

room work → recorder turns → dataset/from-turns JSONL → forge/train drives unsloth → LoRA genome

continuum orchestrates; unsloth (the engine) executes the train.

What

forge/train builds an unsloth train invocation from a from-turns dataset + base model + LoRA params (r/alpha/epochs — sane defaults, recipe-borne later), runs it off the main thread (tokio::process), and returns the result. dry_run maps to unsloth's --dry-run (resolve config + exit) so the seam is validated without GPU time. Missing-dataset / missing-unsloth fail loud.

Live-validated

forge/train --dry-run against the running unsloth resolved the full training config end-to-end:

model: unsloth/Qwen3-0.6B
data: { local_dataset: <from-turns jsonl>, format_type: chat }
training: { training_type: lora, output_dir: ~/.continuum/forge/lora, num_epochs: 1, ... }
lora: { lora_r: 16, lora_alpha: 16, target_modules: q/k/v/o/gate/up/down_proj, ... }

— proving continuum builds a valid unsloth train invocation from a persona's room data.

Tests (extend the forge test mod)

• forge_train_errors_on_missing_dataset — the missing-input guard
• forge_train_dry_run_resolves_against_live_unsloth — the live dry-run (skip-if-no-unsloth)
• 10/10 forge tests pass.

Remaining for #32

Export the LoRA (unsloth export) + page it into the genome, and map a ForgeRecipe → these train params.

See memory coordination-learning-flywheel.

🤖 Generated with Claude Code