formulas: families poisson/binomial/negative_binomial/cumulative compile to undefined observe-families (CompileError)

[aaronstevenwhite]

Summary

Several quivers.formulas response families compile to observe-family names that the DSL _FAMILY_REGISTRY does not define, so fit(..., family=...) fails at compile time. The quivers.formulas module docstring also advertises families (Poisson / NegBin / Cumulative / hurdle / mixture) that are not backed.

Affected families (src/quivers/formulas/family.py)

family key	observe_family	in DSL _FAMILY_REGISTRY?
poisson	Poisson	no
binomial	Binomial	no
negative_binomial	NegativeBinomial	no
cumulative	OrderedLogistic	no

The released DSL registry (quivers.dsl.compiler._get_family_registry()) backs only: Bernoulli, Categorical, Normal, Beta, Gamma, StudentT, the continuous loc-scale / positive / (0,1) families, Dirichlet, MVN, etc. No Poisson / Binomial / NegativeBinomial / OrderedLogistic.

Repro (the OrderedLogistic case is directly verifiable in the DSL, 0.13.0)

from quivers.dsl import loads
loads("""
object Resp : FinSet 10
program m : Resp -> Resp
    sample eta <- Normal(0.0, 1.0)
    let cutpoints = zeros(4)
    observe y : Resp <- OrderedLogistic(eta, cutpoints)
    return y
export m
""")

quivers.dsl.compiler.CompileError: line 6, col 4: undefined morphism or distribution family 'OrderedLogistic'

The formula path (fit(formula, data, family="cumulative")) routes through the same compiler, so it raises the same error. family="poisson" | "binomial" | "negative_binomial" fail identically (registry has no matching family).

Expected

Either (a) register the corresponding observe families so the declared formula families compile, or (b) remove the unsupported entries from families and trim the quivers.formulas.__init__ docstring so it only lists families that actually compile.

Notes

• OrderedLogistic/cumulative-link is the one most likely to be missed — ordinal responses are common. A workaround is cutpoints + Categorical(probs) with probs built from sigmoid(c_k - eta) differences, but a first-class OrderedLogistic family would be the clean fix.
• Still present on main (HEAD 4fefe49): poisson / negative_binomial / cumulative have no registered backing.