SSC-027: apr-cli: wire real training execution for --task classify

[noahgift]

SSC-027: End-to-End CLI Execution

Priority: P2
Blocked by: SSC-026 (training loop)
Blocks: —
Contract: References training-loop-v1.yaml

Problem

apr finetune --task classify currently only does plan mode (VRAM estimation, pipeline summary). It does not actually execute training. The last mile: wire the real ClassifyTrainer::train() invocation from the CLI.

What Exists

• CLI argument parsing for --task classify, --num-classes, --data, --method lora
• Config construction and pipeline summary display
• Plan-mode output (VRAM estimation, parameter count)
• ClassifyTrainer (after SSC-026) with full training loop

What's Missing

1. Actual training invocation in run_classify():

   • Load tokenizer (BPE from HuggingFace or ShellVocab)
   • Load base model weights (from_safetensors)
   • Create ClassifyPipeline with LoRA
   • Create ClassifyTrainer
   • Call trainer.train(corpus)
   • Save final adapter + classification head

2. Progress reporting:

   • Per-epoch: loss, accuracy, LR, time
   • ETA estimation
   • Final summary with best metrics

3. Model saving:

   • Save LoRA adapter as .apr format
   • Save classification head weights
   • Save training config for reproducibility
   • Generate model card

4. Integration test:

   • apr finetune --task classify --data test_corpus.jsonl --epochs 2 --batch-size 4
   • Verify model files created
   • Verify training completed without error

Acceptance Criteria

• apr finetune --task classify --data corpus.jsonl actually trains
• Progress output shows loss/accuracy per epoch
• Final adapter saved to output path
• --plan still works (no regression)
• Integration test passes
• Documentation updated with full usage examples

References

• CLI dispatch in crates/apr-cli/src/commands/finetune.rs
• Spec: bashrs/docs/specifications/shell-safety-inference.md Section 14

[noahgift]

Implemented in 1b09ed4. apr finetune --task classify now wires ClassifyTrainer::train() with full epoch metrics display, --checkpoint-format flag, and JSON output.