Bump version to 0.3.0

New in 0.3.0:
- Delight Policy Gradient gating (soft sigmoid + hard top-K)
- SEPA-annealed PG→DG transition with lambda override
- Scale-only surprisal normalization for instruct models
- Per-generation top-K surprisal token logging
- Hard delight transform (binary sign-aware token mask)
- Cross-model campaign support (Nemotron, Kimi, Qwen 235B)
- trust_remote_code + tokenize=True for broader model compat
- All ty type checker errors resolved (CI green)

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Author: teilomillet

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "retrain"
-version = "0.2.1"
+version = "0.3.0"
 description = "RLVR training framework for LLMs"
 readme = "README.md"
 requires-python = ">=3.11"