Tool‑using language agents must revise prepared actions when evidence changes or authority is revoked. Selective withdrawal means pausing affected actions, preserving those still valid, and resuming only after sufficient repair. We study supervised fine‑tuning in NAQD‑Env, a synthetic environment that provides explicit dependencies and a deterministic reference policy. Four LoRA checkpoints of Qwen2.5‑3B‑Instruct compare an original curriculum with an enriched one; the latter is evaluated under three system‑prompt framings that keep views and targets identical. Three instruction‑tuned baselines are added. All models are evaluated under three inference prompts on fresh validation cases, held‑out dependency families, and held‑out structural combinations. Results show that under the naqd prompt, SFT checkpoints achieve decision agreement of 0.77‑0.92 and withdrawal recall of 0.46‑0.86, whereas baselines reach only 0.45‑0.63 and at most 0.06 recall. Rule‑level diagnostics reveal that the original‑curriculum checkpoint wrongly withdraws 26 of 34 actions that still have alternative support, while enriched checkpoints withdraw only 0‑1. Conversely, all SFT checkpoints narrow prohibited combinations less often than the 7B and 8B baselines, a case absent from both curricula. Delegation experiments show more withdrawal but poorer preservation of independently authorized actions, and all four checkpoints omit valid event reports. Framing contrasts vary across metrics and splits, failing to establish a reproducible advantage or equivalence. Each arm was trained only once, and jointly changed curriculum properties, prior evaluation exposure, and baseline‑SFT quantization differences limit causal attribution. The contribution is a method for detecting rule‑specific regressions that aggregate accuracy can conceal, not evidence that fine‑tuning alone ensures agent safety. Review