500K Twist Output Non-Monotonicity

Amplifies the coexisting tightening/relaxing effects of Algorithm 2 discovered in small samples in document 16 to a 500K-scale observation. The same commit caused twists_of_ec_labels_500k.json to generate a diff of +1899/-53404 lines—yet the base-curve set of Algorithm 1 only shrinks. If the twist JSON changed solely because base curves were removed, the number of added lines should approach zero, not 1899. Therefore, the full-scale output proves that the semantic changes in Algorithm 2 were indeed triggered in the large-scale data domain. The document also demonstrates the consistent rigorous self-restraint of this research line: git diff counts "lines", and the JSON contains key lines, bracket/comma structural lines, twist integer lines, and entire deleted base blocks all at once. Thus, 1899 is merely added diff lines, not a proven unique new twist count—a complete entry-level census requires truly materialising the old/current JSONs and performing a parse set difference. The document also honestly records the coverage failure of the small-sample fixture: for the 12 surviving curves currently <150, $T_{\rm old}(E)=T_{\rm new}(E)$ holds curve by curve, meaning the branch coverage of the small positive fixture for this Algorithm 2 semantic difference yielded 0 observed deltas—this is exactly the direct reason why v0.3 must add synthetic semantic tests, and v0.4 requires a full-file entry census.