feat(ms-ai-architect): RX-HDR oppløst → permanent R7-ustempelbar audit-signal (premiss 10→2) [skip-docs]

Premiss-verifisering mot ground truth (simulert R7-stempler insertVerifiedFields
over 327 non-advisor-filer): RX-HDR-populasjonen kollapser fra påstått 10 til 2.
De 9 'base-felt forbi 500B'-filene er falske positiver — stempleren setter Verified
øverst, uavhengig av hvor Status/Category ligger. Kun 2 filer er ekte ustempelbare
(gpt5-gpt41-pricing-models, entra-agent-id-zero-trust): headere med fete **Status:**-
prosahaler (175B/199B) der begge stempler (Verified + Verified by, judge-v3.1) ikke
får plass i 500B. Feltomstokking løser dem ikke — verdi-slanking kreves (utenfor
RX-HDRs scope), så de folder inn i R7s eksisterende 'header-slanking kreves'-flagg-sti.

audit-corpus-headers.mjs: ny ren isStampable() kjører R7s ekte stempler med fast
representativ probe (dato + judge-v3.1); per-fil 'stampable' + aggregat 'unstampable'
+ FLAG-liste. Gjør populasjonen permanent så den ikke stille kan vokse (gap-disiplin).
+2 tester (dense→ustempelbar, clean→stampable). Suite 924→926 exit 0.

tests/kb-update/test-audit-corpus-headers.test.mjs

@@ -228,3 +228,34 @@ test('auditHeaders — footer conflict needs BOTH a header date and a norsk foot
   assert.equal(r.files['b.md'].footerConflict, null);
   assert.equal(r.aggregate.footerConflicts, 0);
 });
+
+// ── RX-HDR (dissolved → permanent audit signal): R7-unstampable detection ───
+// RX-HDR was replanned as a 10-file header-slanking unit. Premiss-verification against
+// ground truth (simulating the real R7 stamper insertVerifiedFields over all 327 non-
+// advisor files, 2026-07-17) collapsed the population to 1: only headers so byte-dense
+// that BOTH stamps (**Verified:** + **Verified by:**) cannot fit inside the 500-byte
+// window are truly "passing-men-ustempelbar". Field-reordering alone cannot fix such a
+// file (its meta lines carry fat prose tails — a VALUE-slim is required, out of RX-HDR's
+// scope), so the unit dissolved into R7's existing "header-slanking kreves" flag path.
+// This signal makes the population permanent so it can never silently grow: the audit
+// runs the exact stamper R7 uses (representative stamp date + judge-v3.1) and flags any
+// file it would reject. [[gap-discipline-must-close]] — the lint signal IS the mechanism.
+
+// A header so dense that inserting Verified + Verified by lands past 500B — the
+// gpt5-gpt41-pricing-models.md shape (a ~175B **Status:** prose tail). `---` sits at
+// byte 500, so the two stamps (~53B) cannot fit however the meta lines are ordered.
+const UNSTAMPABLE =
+  '# T\n\n**Last updated:** 2026-06\n**Status:** ' + 'GA '.repeat(140).trim() +
+  '\n**Category:** x\n**Type:** reference\n\n---\n\n## A\n\ntekst\n';
+
+test('auditHeaders — a header too dense for R7 to stamp within 500B is flagged unstampable', () => {
+  const r = auditHeaders(['u.md'], reader({ 'u.md': UNSTAMPABLE }));
+  assert.equal(r.files['u.md'].stampable, false, 'dense header is not R7-stampable');
+  assert.equal(r.aggregate.unstampable, 1);
+});
+
+test('auditHeaders — a conformant header is stampable and not counted unstampable', () => {
+  const r = auditHeaders(['f.md'], reader({ 'f.md': CLEAN }));
+  assert.equal(r.files['f.md'].stampable, true, 'clean header is R7-stampable');
+  assert.equal(r.aggregate.unstampable, 0);
+});