nx_paramplumb.nx

nx_paramplumb.nx -- FIND PARAMETERS THAT WERE ADDED BUT NEVER PLUMBED THROUGH. THE DEFECT CLASS, measured three times in one day and invisible to every gate each time: * nx_gsplat's viewport became a parameter (gs_render_aniso_at) -- but nx_gsplat_tile_lib's gt_render kept reading gs_w()/gs_h(), the MODULE DEFAULTS, in six places. The tiled path could only ever partition a 512x384 frame. Its gate was 12/12 GREEN the whole time, because every test ran at the default. * nx_skeleton's evaluator was gated GREEN at 32 bones against a 104-370 joint corpus. * a surface gate's mutant survived because its fixture never exercised the changed path. ****** A PARAMETER THAT WAS ADDED BUT NOT PLUMBED IS INVISIBLE TO EVERY GATE, BECAUSE EVERY TEST STILL PASSES AT THE DEFAULT -- THE PROOF IS TRUE AND SCOPED TO A SIZE NOBODY SHIPS. That is worse than an absent capability: an absent one is counted as missing, this one is counted as DONE. WHAT IT DETECTS -- the estate's own parameterisation idiom, so the signature is exact rather than a guess. When a module-scope default becomes a parameter, this tree writes: func X_at(..., <the new parameters>) -> i64 { ...the real body... } func X(...) -> i64 { return X_at(..., <the module defaults>) } // thin wrapper The wrapper is correct and keeps existing callers working. The DEFECT is a caller that had the quantity available and still called X. So: for every confirmed X / X_at pair, count the call sites of the DEFAULTED name X. Each one is a path through which the new capability is unreachable. WHY "_at" ALONE IS NOT THE TEST. "_at" is also this tree's index-accessor idiom (nx_artifact_store_path_at(s, rowid)), which is a different thing entirely. A pair only counts when BOTH func X( and func X_at( are declared, so an accessor with no defaulted twin is never reported. That exclusion is the difference between a detector and a grep. COVERAGE IS PART OF THE ANSWER. The walk reports files seen, files read and files it could not read, and those must sum. An absence claimed from a partial scan is not an absence -- this organ was written because a lane (mine) claimed one from a filtered grep earlier today and was wrong. NOTHING HERE IS RANKED BY A PICKED THRESHOLD: the report is ordered by defaulted-call count, which IS the blast radius, and every count is printed so the reader adjudicates rather than trusting a cutoff.