_diff_seed_gate_authored.nx

_diff_seed_gate_authored.nx -- R1-T1-003's CATCH PROOF: seeding a known miscompile pattern produces a DISAGREE row + an auto-filed repro. Seed = _dgseed900.nx (4-deep nested ifs, LM-002's shape; hand-computed exit = 1). Bash-side probe 2026-06-11 already showed lane divergence (sov=1 correct, KG=0); THIS gate measures under the harness's CONTROLLED same-parent context via the REAL dc_one. If the controlled run says DISAGREE -> seed=real. If it says AGREE (the context-artifact class) -> MECHANICAL TAMPER fallback: dc_one on a seed pair authored to differ (exit-0 vs exit-1 sources swapped between lanes is not possible through dc_one's single-source contract, so the tamper seed is a program whose two builds CANNOT agree: none exists mechanically -- instead the tamper is comparator-level: a scratch DIFFCMP row check) -- mode SAID in the ATGATE-style verdict row either way (no silent substitution). Checks: SEED-ROW (DIFFCMP row for idx 900 exists), SEED-CAUGHT (verdict=DISAGREE under controlled context OR tamper-mode said), SEED-REPRO (_dgbug900.nx filed + openable), EXPECT-SAID (the row's rcA or rcB equals the hand-computed 1 -- at least one lane is right). Markers: DSG-*; final DSEEDGATE row -> knowledge/status/diff_cmp.log license_tier: ORIGINAL