@@ -683,6 +683,7 @@ import {
   isGateSemanticLabel,
   labelNames,
   latestMarkedComment,
+  markerMatches,
   prDeliversCard,
   proxyRearmPlan,
   resolveSweepRepo,
@@ -781,6 +782,7 @@ const SELF_TEST_BATTERIES = Object.freeze({
   '#18701: ONE thread set -- what the template STATES is what the queue guard READS': 16,
   '#18719: a RETRACTED claim leaves the pool — a withdrawn claim never governs': 36,
   '#18683: the card-comment read pages to a cap — past 100 is UNJUDGED, ⛔ never a truncated pool': 27,
+  '#18764: a DECORATED claim ENTERS the pool — ONE reading, and it is the sibling\'s': 24,
 });
 
 // DELETING an entry silences that battery's floor exactly as effectively as
@@ -792,8 +794,8 @@ const SELF_TEST_BATTERIES = Object.freeze({
 // one #17915 adds, by the one #17959 adds, by the one #18042 adds, and by the
 // one #18174 adds, and by the one #18141 adds, and by the one #17919 adds, and
 // by the one #16833 adds, and by the one #18456 adds, and by the one #18719
-// adds, and by the one #18683 adds.
-const SELF_TEST_BATTERY_FLOOR = 30;
+// adds, and by the one #18683 adds, and by the one #18764 adds.
+const SELF_TEST_BATTERY_FLOOR = 31;
 
 // The key an assertion is filed under when no battery is open. It is not a
 // declared battery, so it reds by the same set difference rather than silently
@@ -1756,7 +1758,7 @@ export function claimRetractions(commentRows) {
   });
   const out = new Map();
   for (const claim of indexed) {
-    if (!CLAIM_COMMENT_MARKER.test(String(claim.row?.body ?? ''))) continue;
+    if (!markerMatches(CLAIM_COMMENT_MARKER, String(claim.row?.body ?? ''))) continue;
     // Fail closed, twice: an unattributable claim cannot be matched against an
     // author, and an unattributable candidate cannot be the seat that wrote it.
     if (claim.author === null) continue;
@@ -1791,12 +1793,82 @@ export function claimRetractions(commentRows) {
  */
 export const CLAIM_SELECTION_RULE =
   'the GOVERNING claim — the NEWEST comment whose body carries a line beginning `Claim:`/`Claimed:` '
+  + '(read through the sibling reader\'s ONE reading, `markerMatches`: the bare marker first, then the '
+  + 'shared stripper per line, with a markdown LIST ITEM refused — so a decorated `**Claim:**` is the '
+  + 'same record as a bare one and ⛔ the two readers of this thread cannot answer differently) '
   + 'AND whose `Branch:` line parses at least one protocol-shaped branch (newest by `created_at`; an '
   + 'unreadable stamp or a tie falls back to thread order, later row wins). The pool is every claim '
   + 'comment sharing that `created_at`; when NO claim names a branch at all, every claim comment is '
   + 'the pool. ⛔ Not earliest, ⛔ not a session match, ⛔ not the one whose body mentions the key. '
   + `MEMBERSHIP comes first: ${CLAIM_RETRACTION_RULE}`;
 
+// ---------------------------------------------------------------------------
+// #18764 — a DECORATED claim ENTERS the pool. ONE reading, and it is the
+// sibling's.
+//
+// The pool and the retraction indexer each tested `CLAIM_COMMENT_MARKER`
+// against the RAW body. The constant anchors the bare word at line start and
+// tolerates leading whitespace and one `>` — nothing else — so a claim a seat
+// wrote as `**Claim:** …` or `` `Claim:` … `` was not SUPERSEDED here, it was
+// never a candidate: not listed, not rejected, not named anywhere in the
+// record. Meanwhile `claimGovernance`, imported from the same sibling, had
+// already been reading both markers through `markerMatches` since #18680, so
+// ONE FUNCTION held both answers at once — governance saw the bolded claim and
+// the pool beside it did not.
+//
+// Measured on the two live records the #18764 escalation named (read
+// 2026-09-17T22:02Z; ⚠️ both cards have since left that state and the reading
+// is stated with its time for that reason), replayed through `--pair-json`:
+//
+//   a bolded/backticked claim carrying its own `Branch:` and `Clause-②: no`
+//     → `claim.selected: none — no comment on this thread carries a line
+//       beginning `Claim:``, the limb read MISPLACED and `--pair` exited 4,
+//       prescribing a `Clause-②-correction:` for a line the seat had already
+//       written in the right place.
+//   a BARE older claim declaring `yes` beneath a DECORATED newer one declaring
+//     `no` → the OLD claim governed the declaration, the reading answered
+//     `DECLARED yes`, and the newer record was not even listed as rejected.
+//     ⭐ That is the expensive direction: not a missing reading but a WRONG
+//     value, reported with every appearance of having been read.
+//
+// ⛔ The repair is NOT a `\*\*` added to `CLAIM_COMMENT_MARKER`. Decoration is
+// an OPEN set (#18680 settled that), so admitting one spelling buys exactly
+// that spelling and replays this card on the next one — and the constant is
+// the PROTOCOL's spelling, which is why every reader imports it rather than
+// restating it. ⛔ Nor is it a second undecorator written here: this file
+// would then own a definition of "decorated" that the sibling could drift
+// from, which is the very failure the card names. What both raw tests do
+// instead is READ THROUGH `markerMatches` — the sibling's one reading, bare
+// test first (so the change is provably additive: no body that matched
+// yesterday stops matching) and the shared stripper after it, with the list
+// item refused there and the near-miss vocabulary
+// (`OWNERSHIP_MARKER_NEAR_MISS_FORMS`) kept where it lives. ⛔ That vocabulary
+// is the sibling's and is not re-declared here.
+//
+// ## The OTHER undecorator in this file stays where it is, and that is MEASURED
+//
+// `undecorateRetractionLine` looks like the same job one section up, and the
+// card asked whether the file should end up with one undecoration path. It
+// should not, and the two paths were run over one fixture set to find out:
+// 5 of 12 fixtures read DIFFERENTLY. `undecorateRetractionLine` strips `_` and
+// then every leading non-letter/non-digit character, so it reads
+// `__Claim:__`, `- Claim:`, `* Claim:`, `## Claim:` and `🚨 Claim:` as the
+// directive; `markerMatches` reads none of those, deliberately — the list item
+// and the heading are NAMED refusals with pins behind them, and the underscore
+// form is a named near miss.
+//
+// ⭐ The divergence is load-bearing in BOTH directions, which is why unifying
+// by hand would be a regression whichever way it went. The #18373 retraction
+// this file's own battery replays opens `🚨 **撤回上一条认领…`, and the leading
+// sigil is exactly what the prose channel must strip before the anchor can
+// OPEN the line — narrowing the retraction stripper to the shared one would
+// stop reading the measured specimen #18719 landed for. Widening the shared
+// one the other way would make `- Claim:` a claim, which H20 pins as not one.
+// ⇒ two strippers, two jobs, and the difference is stated here rather than
+// discovered again. ⛔ Not unified from here: it is filed, with the fixture
+// table, for the seat to route.
+// ---------------------------------------------------------------------------
+
 /**
  * WHICH claim comment this reading is built from, and which it is not — the
  * input half of the declaration limb (#18456).
@@ -1837,7 +1909,7 @@ export function claimCarrierSelection(commentRows) {
   // ⭐ The full claim listing KEEPS the retracted rows: the record names them
   // RETRACTED below. A pool that silently shrank would replace one invisible
   // fact with another.
-  const claims = commentRows.filter((row) => CLAIM_COMMENT_MARKER.test(String(row?.body ?? '')));
+  const claims = commentRows.filter((row) => markerMatches(CLAIM_COMMENT_MARKER, String(row?.body ?? '')));
   const live = claims.filter((row) => !retracted.has(row));
   const governing = governance.governing;
   const matched = governing ? live.filter((row) => (row?.created_at ?? null) === governing.createdAt) : [];
@@ -1898,13 +1970,17 @@ const CLAUSE2_CORRECTION_KEY_TEXT = 'Clause-②-correction';
  * because the older claim happened to agree.
  *
  * The predicate that separates them is `CLAIM_COMMENT_MARKER`, imported rather
- * than restated: a claim comment is one whose body carries a LINE BEGINNING
- * `Claim:` (or `Claimed:`, optionally blockquoted), and that one spelling is
- * the whole set — so a heading-style claim (`## Claim — …`) is not a claim
- * comment here, however complete the reasoning under it, and its thread reads
- * `absent`. ⛔ Widening the predicate is not this file's to do: it is the
- * sibling's constant precisely so the two readers cannot drift, and the remedy
- * for a thread that reads `absent` is a comment in the fixed spelling.
+ * than restated, and OFFERED through `markerMatches` — the sibling's one
+ * reading (#18764), imported for the same reason the constant is. A claim
+ * comment is one whose body carries a LINE BEGINNING `Claim:` (or `Claimed:`,
+ * optionally blockquoted), read bare first and then with the shared
+ * decoration stripped, so `**Claim:**` and `` `Claim:` `` are that same line
+ * and ⛔ not a second spelling this file admits on its own. A heading-style
+ * claim (`## Claim — …`) is still not a claim comment here, however complete
+ * the reasoning under it, and its thread reads `absent`. ⛔ Widening the
+ * predicate is not this file's to do: the constant AND the reading are the
+ * sibling's precisely so the two readers cannot drift, and the remedy for a
+ * thread that reads `absent` is a comment in the fixed spelling.
  *
  * @param {{ body?: string, created_at?: string }[]|null} commentRows — the REST
  *   comment rows, or `null` when the thread could NOT be read.
@@ -8108,6 +8184,82 @@ export async function selfTest() {
   })());
   t('⛔ CONTROL: the diagnosis KEY is unchanged, so every sentence already keyed to `comments` still finds it', readDiagnosisKey('comments', 770017) === 'comments:770017');
 
+  // -- #18764: a DECORATED claim ENTERS the pool -----------------------------
+  //
+  // The two live records the escalation named, replayed OFFLINE and ⛔ never
+  // re-graded: they are another repository's cards and another seat's work.
+  // Ids, stamps and the load-bearing LINES are the real ones, read from the
+  // REST rows at 2026-09-17T22:02Z — ⚠️ both cards have since left the state
+  // they were read in, which is why the reading carries its time. The bodies
+  // below ADD the `Branch:` / `Clause-②` lines the real comments did not carry:
+  // without them the thread reads `claim-branch-unparsed` on BOTH sides of this
+  // change (measured), and the defect this battery pins is the one that only
+  // shows once a claim is otherwise complete.
+  battery('#18764: a DECORATED claim ENTERS the pool — ONE reading, and it is the sibling\'s');
+  const D64 = (id, at, body, login = 'os-sales') => ({ id, created_at: at, user: { login }, body });
+  const D64_BOLD = D64(5721120402, '2026-09-17T20:57:09Z', [
+    '**Claim:** card objectui#9660, by the `domain:spec` @ objectui execution seat, session `session_01UanLVj6xvbS6puBCewLr8L`.',
+    'Branch: `claude/issue-9660-named-test-invocation`',
+    'Clause-②: no',
+  ].join('\n'));
+  const D64_TICK = D64(5720184809, '2026-09-17T19:41:36Z', [
+    '`Claim:` card objectui#9717, by the `domain:spec` @ objectui execution seat, session `session_01UanLVj6xvbS6puBCewLr8L`.',
+    'Branch: `claude/issue-9717-doc-component-types`',
+    'Clause-②: no',
+  ].join('\n'));
+  // The older BARE claim, declaring the OPPOSITE value — so "the wrong carrier
+  // governs" is a WRONG VALUE here and not merely a missing one.
+  const D64_BARE_OLD = D64(5700000001, '2026-09-17T18:00:00Z', [
+    'Claim: the older BARE claim',
+    'Branch: `claude/issue-9660-older-bare`',
+    'Clause-②: yes',
+  ].join('\n'));
+  const D64_POOL = (rows) => claimCarrierSelection(rows).pool.map((r) => r.id).join(',');
+  const D64_RECORD = (rows) => {
+    const rec = pairInputRecord({ pr: 9999, card: 9660, cardComments: rows, headSha: 'offline' });
+    return { selected: [rec['claim.selected']].flat().join('\n'), rejected: [rec['claim.rejected']].flat().join('\n') };
+  };
+
+  t('⭐ a BOLD claim ENTERS the pool — the state it could not reach at all before', D64_POOL([D64_BOLD]) === String(D64_BOLD.id), D64_POOL([D64_BOLD]));
+  t('…and GOVERNS: the branch resolved is the one IT names', (claimCarrierSelection([D64_BOLD]).governing?.branches ?? []).join() === 'claude/issue-9660-named-test-invocation');
+  t('…and the declaration limb reads the `Clause-②` line OFF IT', cardDeclaration([D64_BOLD]).state === 'declared' && cardDeclaration([D64_BOLD]).value === 'no');
+  t('⛔ CONTROL — the CONSTANT is NOT widened: the raw marker still refuses that same body', CLAIM_COMMENT_MARKER.test(D64_BOLD.body) === false);
+  t('⭐ the BACKTICKED spelling is the same record, by the same reading', D64_POOL([D64_TICK]) === String(D64_TICK.id));
+  t('…with its own `Branch:` line resolved', (claimCarrierSelection([D64_TICK]).governing?.branches ?? []).join() === 'claude/issue-9717-doc-component-types');
+  t('…and its own declaration read', cardDeclaration([D64_TICK]).value === 'no');
+  t('⛔ CONTROL: the raw marker refuses the backticked body too', CLAIM_COMMENT_MARKER.test(D64_TICK.body) === false);
+  t('⭐ a DECORATED NEWER claim SUPERSEDES a BARE older one — the pool is the newest, not the readable one', D64_POOL([D64_BARE_OLD, D64_BOLD]) === String(D64_BOLD.id), D64_POOL([D64_BARE_OLD, D64_BOLD]));
+  t('…and the older record is LISTED, as SUPERSEDED rather than dropped', says(D64_RECORD([D64_BARE_OLD, D64_BOLD]).rejected, 'a SUPERSEDED claim') && says(D64_RECORD([D64_BARE_OLD, D64_BOLD]).rejected, String(D64_BARE_OLD.id)));
+  t('⭐ …and the VALUE the limb reads is the newer one', cardDeclaration([D64_BARE_OLD, D64_BOLD]).value === 'no');
+  t('⛔ CONTROL: the older record declares the OPPOSITE, so selecting the wrong carrier is a WRONG value, ⛔ not a missing one', cardDeclaration([D64_BARE_OLD]).value === 'yes');
+  t('the input record NAMES the decorated row it selected — by id and by date', says(D64_RECORD([D64_BOLD]).selected, String(D64_BOLD.id)) && says(D64_RECORD([D64_BOLD]).selected, '2026-09-17T20:57:09Z'));
+  t('⛔ CONTROL: the same thread read the other way says nobody claimed at all', says(D64_RECORD([D64(1, '2026-09-17T20:57:09Z', 'no claim on this line')]).selected, 'no comment on this thread carries'));
+
+  // The retraction index reads the SAME predicate, so a decorated claim is
+  // retractable by its own author — ⛔ never a record that can be written but
+  // never withdrawn.
+  const D64_RELEASE = D64(5721120999, '2026-09-17T21:30:00Z', 'Release: session `session_01UanLVj6xvbS6puBCewLr8L` — 去向 `pm:queue`');
+  t('⭐ the retraction index SEES a decorated claim — it is retractable by its own author', claimRetractions([D64_BOLD, D64_RELEASE]).has(D64_BOLD));
+  t('…and the pool then says every claim on the thread is RETRACTED, ⛔ not that none was written', says(D64_RECORD([D64_BOLD, D64_RELEASE]).selected, 'RETRACTED'));
+  t('⛔ CONTROL: a DIFFERENT author\'s release retracts nothing, decorated or not', claimRetractions([D64_BOLD, { ...D64_RELEASE, user: { login: 'os-other' } }]).size === 0);
+
+  // The refusals are the SIBLING's and are pinned here as still-refused: this
+  // file admits no spelling of its own, so a form the sibling names as a NEAR
+  // MISS must not become a claim by arriving through this door.
+  t('⛔ a markdown LIST ITEM is still not a claim — the shape the shared reading refuses to undecorate through', D64_POOL([D64(2, '2026-09-17T20:00:00Z', '- Claim: seat.\nBranch: `claude/issue-1-x`')]) === '');
+  t('⛔ a HEADING-style claim is still not one', D64_POOL([D64(3, '2026-09-17T20:00:00Z', '## Claim: seat.\nBranch: `claude/issue-1-x`')]) === '');
+  t('⛔ UNDERSCORE emphasis is still a NAMED near miss, ⛔ not a claim', markerMatches(CLAIM_COMMENT_MARKER, '__Claim:__ seat.') === false);
+  t('⛔ and the `Clause-②-correction:` comment does not enter the pool through the new door either', D64_POOL([FIXED_CORRECTION('no')]) === '' && markerMatches(CLAIM_COMMENT_MARKER, FIXED_CORRECTION('no').body) === false);
+  t('⛔ provably ADDITIVE: a bare claim this file already read reads exactly as before', claimCarrierSelection([CLAIM('Clause-②: no')]).pool.length === 1);
+
+  // ⭐ The file keeps TWO undecoration paths, and that is a measurement rather
+  // than an oversight: `undecorateRetractionLine` strips `_` and then every
+  // leading non-letter/non-digit character, which the shared stripper does not.
+  // 5 of 12 fixtures read differently. Both directions are load-bearing, so the
+  // pair below pins the divergence instead of quietly closing it.
+  t('⭐ the RETRACTION path still reads the sigil-led #18373 line — narrowing it to the shared reading would unland #18719', claimRetractions(RTX_18373).has(RTX_18373.find((r) => r.id === RTX_WITHDRAWN)));
+  t('⭐ …while the SHARED reading does not strip a leading sigil, so the two paths are ⛔ NOT interchangeable', markerMatches(CLAIM_COMMENT_MARKER, '🚨 Claim: PM loop round 1') === false);
+
   // -- The floor: every declared battery RAN, and ran its cases (#13489) -----
   //
   // Evaluated after every battery has had its chance and BEFORE the verdict, so
@@ -8176,7 +8328,9 @@ export async function selfTest() {
       + 'line that QUOTES the spelling held apart from one that declares a value in BOTH halves '
       + 'of that property, the input record whose field roster is the same on exit 0, on exit 4 '
       + 'and on a refusal — with the selected carrier, its body fingerprint and the rejected '
-      + 'candidates each stated — and the exit register).',
+      + 'candidates each stated, the DECORATED claim that enters the pool and governs through '
+      + 'the sibling\'s one reading — the constant unwidened, and the file\'s two undecoration '
+      + 'paths pinned as the two jobs they are — and the exit register).',
   );
 
   selfTestReachedVerdict = true;