nx_fetchwall_gate.nx

nx_fetchwall_gate.nx -- THE GATE FOR THE FETCH-WALL CLASSIFIER, 2026-09-03. SUBJECT: fw_classify / fw_classify_file in-process (nx_fetchwall_lib has no main). THE FIXTURES ARE REAL, NOT PLANTED. The positive is the actual Google AI Mode share body mirrored by nx_research_fetch on 2026-09-03 (status=200, 92982 bytes, a JS gate carrying none of the answer). The false-positive controls are the SEVENTEEN curated phototwin reference mirrors -- real fetched pages that were adjudicated and re-fetched by the refs lane -- because a detector bite-proven only on planted fixtures is not calibrated, which is exactly how nx_mirrorintegrity shipped a 941-permil false-positive rate onto real data. EVERY control row is PRINTED with its permil so a genuine wall hiding among the mirrors is visible as a finding rather than swallowed as a failure. Teeth, in order: T1 the real wall body classifies as a WALL at all. T2 it is named WALL-JS-GATED, not merely convicted. T3 the STRUCTURAL axis is what fired: its fitted-text permil is under the bar. T4 FALSE-POSITIVE CONTROL over all seventeen real mirrors. T5 NEG-CONTROL: dense prose with no markers reads CONTENT. T6 NEG-CONTROL: dense prose that MENTIONS sign-in and cookies STILL reads CONTENT -- markers must never override structure, and without this tooth the marker list would convict half the estate's corpus. T7 NEG-CONTROL: a thin JS-gate body reads WALL-JS-GATED -- so T6 cannot pass by the markers being dead. T6 and T7 are the SAME ordering proven from both sides; either alone is vacuous. T8 NEG-CONTROL: a zero-length body ABSTAINS as UNPROVEN and never reads as content. T9 NEG-CONTROL: a body under the prose floor reads EMPTY, not a convicted wall. license_tier: ORIGINAL No hw writes (Rule 26). expect_exit: 0