Attribution and repair search over declared counterfactual mechanisms.
A baseline and an outcome score are part of the question. Contributions
are allocations of that score, not probabilities that a node is a cause.

Counterfactual queries by abduction, action and prediction (Pearl).

A program's sample and observe sites are structural equations
(`foerster.model.mechanism`). Given evidence and interventions:

1. abduction — condition the program on the evidence (importance sampling
   by `gfi/generate`) and read each weighted factual trace's exogenous
   noise off its sites;
2. action — install the interventions;
3. prediction — run the program again in a world of its own with every
   other site fed its factual noise, aligned by address.

The result pairs each factual world with its counterfactual twin, under
the factual weight: a joint measure, so quantities that need both worlds
(probability of necessity, effect of treatment on the treated) are
expectations over it. A site that exists only in the counterfactual world
draws fresh noise and is listed as `:unaligned`: its answer is
interventional, not counterfactual.

Graphical identification in declared acyclic directed mixed graphs (ADMGs).
Directed edges are causal assumptions; bidirected edges represent latent
common causes. This is not a causal interpretation inferred from source.

`identify` implements Shpitser and Pearl's ID/IDC algorithms (2006).
It returns an observational functional or a nonidentification witness.
Identification assumes the graph, its latent-variable interpretation and
the positivity required by the functional. It does not fit mechanisms.
Expressions are data and can be evaluated on finite probability tables.

Fit and check declared, causally sufficient DAGs from complete numeric rows.
The default mechanisms are linear Gaussian with independent standardized
noise. Fitting is not causal discovery, and diagnostics do not establish
the graph or its counterfactual coupling. No external numerical runtime.