Who’s to Blame? Causal Responsibility Attribution for Macro Placement

When a routed design fails design-rule checking (DRC), engineers ask a blame question: which macro should I move? The machine-learning answer has been gradient saliency on a learned congestion surrogate. We argue that this is a category error: saliency explains what a model looked at, not what in the design caused the failure. We frame macro-placement debugging as a causal responsibility-attribution problem and evaluate thirteen attribution methods in five groups: causal (Halpern-Pearl responsibility and blame, probability of necessity, Shapley value, counterfactual effect), economic (Pigouvian externality price, LP-dual shadow price on routing capacity), gradient saliency (Grad-CAM, Integrated Gradients, Input$\times$Gradient), a proximity heuristic, and a random control. We evaluate all thirteen against an interventional oracle, a small but real analytical-placement plus negotiated-congestion-routing flow in which every $do(\mathrm{macro}\ m \to \mathrm{site}\ s)$ query is a full re-place and re-route. Across three synthetic blocks (1,260 interventions, 2,991 place-and-route runs) we measure diagnosis (rank agreement with the tool's own but-for effects), repair (real DRC reduction after three guided ECO moves on two disjoint candidate-site groups), and faithfulness (deletion curves). Economic and causal methods agree with the tool (mean Spearman $\rho$ of $+0.67$ and $+0.29$); saliency does not ($-0.05$, below random). Acting on causal rankings recovers 85-92% of the oracle repair upper bound and, uniquely, gives identical repair gains on both held-out site groups, whereas Grad-CAM drops from 21.7% to 6.4% and the Pigou price from 23.8% to 15.9%. The action set available to an ECO is the natural contrast set for actual causation, and we show it makes Halpern-Pearl responsibility a computable, actionable quantity for physical design. Code, the interventional dataset, and an untested OpenROAD adapter are released. For explainable AI more broadly, the study is a controlled faithfulness test in a domain where every counterfactual can be checked by re-running the real process.