Restore model.train() after v_hack auto-extract

extract_v_hack runs forward+backward on contrastive pairs to populate
delta_S.grad; the inline auto-extract called model.eval() but never
called model.train() back, so the entire training run was in eval mode.

Qwen3 has no dropout by default so behavior was unchanged, but this
matches the standalone extract CLI's behavior and avoids latent
inconsistency if a model with dropout is used later.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/projected_grpo/train.py

@@ -389,6 +389,7 @@ def main(cfg: Config) -> int:
                             "top_k": str(min(cfg.v_hack_extract_top_k, len(VHACK_PAIRS) - 2)),
                             "tau_axis": "0.0", "schema": "v2_with_sv"})
         # extract zeros grads at exit; opt is built below so no opt-state taint.
+        model.train()  # restore train mode; eval was set only for the extract pass
     v_hack_cpu, _v_sv_cpu = load_v_hack(v_hack_path, model_name, wrappers, k_use=cfg.v_hack_k)
     v_hack = {name: v.to(device) for name, v in v_hack_cpu.items()}
     # Teacher pool: pre-generated rollouts on disk keyed by problem_id. Each step's