Comprehensive-test slow-path registry
Source:R/comprehensive_slow_paths.R
EDI_COMPREHENSIVE_SLOW_PATHS.RdPerformance-based exclusions used by EDI's comprehensive test harness. These rules describe paths that are implemented but intentionally omitted from routine exhaustive execution because their observed runtime is too high. They do not change an inference class's public capabilities and must not be interpreted as "not implemented" declarations.
Format
A named list. `exact_operations` contains keys of the form `response_type||InferenceClass||operation`; every other element contains formula-, dataset-, and design-independent concrete inference-class names for the named slow-path family.
Details
Rule-name prefixes use the harness vocabulary: `boot` is ordinary nonparametric bootstrap, `bbt` is Bayesian bootstrap, `brt` is bootstrap randomization, `pboot`/`param_bootstrap` are parametric bootstrap, and `rand` is randomization inference. Suffixes identify the affected CI, p-value, or typed variant. A class can appear in more than one category.
`InferenceSuite$run_all_inference()` also omits matching class/method/type combinations when its `methods` argument is left at the default `NULL`. Supplying `methods` explicitly opts into the requested paths even when they appear in this registry. Because of that coupling, an `exact_operations` entry must never name the base `compute_estimate` operation itself (only a CI/p-value sub-method, e.g. `compute_rand_two_sided_pval`) – doing so doesn't just skip one slow sub-computation, it silently drops the whole class from `run_all_inference()`'s default output (found 2026-08-26: a pre-existing `"survival||InferenceSurvivalCoxPHRegr||compute_estimate"` entry, harmless while this registry was internal-only, started doing exactly that the moment `run_all_inference()` began consulting it – `compute_estimate()` for that class takes ~0.09s, nowhere near "too slow"; removed).