Declared unbound ragged plate observation loses its public predictive output

StanBlocks 1e0af724be4f rejects a missing censored observation inside a ragged plate even when its hygienic cell name is declared as an observation. The same check is present on devibe ec50705dc48e673ad356b15afc6c58cd44e03213.

using StanBlocks

data = (; t=[[0.2, 0.8], [0.3]], lo=[[0.05, 0.05], [0.05]], nsub=2)
base = @slic data begin
    p ~ plate(t, lo; outer=(nsub,)) do tt, ll
        y ~ censored(normal, tt, 1.0; lower=ll)
        tt
    end
end
model = StanBlocks.SlicModel(base.model, base.data, base.mod, (:p_y,))
StanBlocks.stan_code(model)

This errors with censored: observations must be data-qualified; censored priors are not supported. The plate's discovery probe validates p_y, but emitted sampling uses private ragged storage p_y__pl_mem_1, which does not inherit that observation status. Adding the private name to the declaration permits code generation, but stan_descriptor(model).outputs contains only p_y__pl_mem_1; there is no public p_y_gen with flat coordinates and ragged segments.

Expected: one public predictive carrier for the declared observation, with a flat draw and inclusive per-group ends, and no sampled censored prior. Bound ragged cell observations should retain their flat predictive and aggregate pointwise likelihood carriers.