nx_browser_struct_col_lib.nx

nx_browser_struct_col_lib.nx -- BR2: THE CSS-ON STRUCTURAL-EXACTNESS COLUMN OF THE BROWSER CENSUS. WHY A LIB AND NOT A BLOCK INSIDE THE CENSUS. nx_browser_census.nx computes every number it prints inside main() and imports nothing testable, so a gate over it can only exec the promoted binary and a mutation harness reads NOT-REACHED across that fork. The judgement therefore lives here, inside the gate's own build closure, where the BR2 bar can actually be bitten. The census keeps ONE delegating declaration (bc_struct_col) so the compare-matrix symbol ruler still measures the organ the row names. WHAT IT ADDS. The census plane row was six tab fields: site class reading_recall_pm ratio_pct chrome_h nishi_h BR2 appends four, STRICTLY ADDITIVELY -- a six-field row still parses and its struct columns read UNMEASURED, never a silent zero: ... struct_matched_pm struct_dy struct_dx struct_bands_oracle THE FOURTH COLUMN IS THE DENOMINATOR, AND IT IS THE REASON THIS LIB EXISTS. MEASURED 2026-08-26 on saved bytes with the incumbent ruler: goog_raw.html vs its Chrome oracle returned "matched=2 (1000 permille)" over TWO oracle bands, and so_aug/x_aug returned "matched=0 (0 permille)" over ZERO oracle bands because those two tall captures are blank. Both readings are arithmetic on an empty or near-empty subject, and both arrive wearing a number that reads as a measurement -- a perfect score and a total failure, neither of them evidence. A struct column fed from that ruler without its denominator would have recorded a 1000 and two 0s in the census as facts. So the floor is DERIVED FROM THE BAR ITSELF, not picked: with N oracle bands, one band's outcome moves the reading by 1000/N permil. The bar is matched >= 900, i.e. 100 permil of headroom below perfection, so a reading is only capable of resolving the bar when 1000/N <= 100, i.e. N >= 10. Below that a single band flips the answer by more than the entire distance the rung is measuring. Rows under the floor are counted as UNDERPOWERED -- named, excluded from the mean and from the bar, never averaged in and never silently dropped. THE BAR IS DATA. matched-permil-min, mean-dy-max and the four bar sites come from knowledge/browser_br2_bar.conf, which quotes the browser.plan BR2 done-rule verbatim. A missing or unreadable conf yields BSC_UNOBSERVABLE, never a default: a bar nobody can read must not be reported as met, and must not be reported as failed either. THE PAIR HAS A NOISE FLOOR. The done-rule's second clause is "the census pair no longer favours CSS-OFF". A mean difference is a WIN only when it exceeds the census-variance band the browser.plan debt row already measured (~60 permil); inside that band the honest answer is TIED, and the CSS_ENABLED flip must not be taken on it. license_tier: ORIGINAL No hw writes (Rule 26).