feat(safety-lora): Sprint 17 — safety-subspace LoRA (SaLoRA / SPLoRA) (v1.7.0)

[konjoinfinity]

Summary

• toki.safety_lora (new module) — three complementary safety-preserving LoRA techniques from the 2025-2026 literature, all validated on 1B-3B models (toki's target range)
• toki.finetune (extended) — LoRAConfig gains four new safety fields; LoRAFinetuner.train() returns LoRATrainResult and wires in the safety hooks
• CLI — python -m toki finetune --safety-lora-rank --safety-subspace --splora-audit

Research basis

Paper	arXiv	Technique	Implementation
SaLoRA	2501.01765	Frozen safety delta before task fine-tuning	load_safety_subspace + freeze_safety_adapter
SPLoRA (TACL)	2506.18931	Post-hoc E-DIEM audit of weight-update shifts	splora_audit + SploraAuditResult
Rank-1 MLP alignment	2507.17075	Rank-1 up_proj safety adapter; zero reasoning tax	safety_lora_rank field in LoRAConfig

New symbols

Symbol	Module	Description
SafetyLoRAConfig	safety_lora	Four safety fields with safe defaults
SploraAuditResult	safety_lora	Frozen dataclass: flagged_layers, max_ediem, passed, threshold
LoRATrainResult	safety_lora	training_loss + num_steps + optional SploraAuditResult
load_safety_subspace(path)	safety_lora	Load safety delta checkpoint (SaLoRA)
freeze_safety_adapter(model, delta)	safety_lora	Apply + freeze safety params; no-op when delta is None
splora_audit(model, base_state, threshold)	safety_lora	E-DIEM post-hoc safety audit (SPLoRA)

All torch/peft operations are behind try-import guards — raise ImportError("requires toki[hf]: pip install toki[hf]") cleanly when deps absent.

Backward compatibility

LoRAConfig with no safety fields set (all defaults) produces identical training behaviour to v1.6.0 — confirmed by test_finetune_extended.py::test_train_no_safety_fields_no_audit.

Test plan

• test_safety_lora.py — 24 tests: SafetyLoRAConfig, SploraAuditResult, LoRATrainResult, load/freeze/audit import guards, no-op paths, mock-tensor integration
• test_finetune_extended.py — 15 tests: LoRAConfig new fields, backward compat, LoRAFinetuner construction, train() with mocked torch
• test_main.py additions — 5 CLI tests: finetune config print, flag reflection, import guard
• 644 / 644 tests passing (600 → 644)
• cargo test green, cargo clippy -- -D warnings clean

Unblocks

• P3-2 (compliance certification report) — can now report on safety-preservation during fine-tuning
• P3-1 (dual-agent red-team loop) — full pipeline now includes both the evaluator reliability fix (Sprint 16) and the safety-subspace fine-tuning protection (Sprint 17)

https://claude.ai/code/session_01XCHiLCiVeL6WXQdsAcQTbx

---

Generated by Claude Code