diff --git a/scripts/check_pr_perimeter.py b/scripts/check_pr_perimeter.py
index 34e697f1fc..74f8ed085d 100644
--- a/scripts/check_pr_perimeter.py
+++ b/scripts/check_pr_perimeter.py
@@ -717,10 +717,25 @@ def check_assertion(
     r"\b(?:editer|modifier|toucher|ouvrir|créer|creer|ajouter|changer|mettre\s+à\s+jour|mettre\s+a\s+jour|update|edit|modify|touch|open|create|add|change)\b",
     re.IGNORECASE,
 )
-# A named file as it would appear in a body: backticked path, bare basename
+# A named referent as it would appear in a body: backticked path, backticked
+# DOTTED SYMBOL (`pick_idle_grain.upsert_orphans_comment`), bare basename
 # (.py/.cs/.yml/.json/.md/.ipynb/.sh), or a known scripts/<x>.py shape.
+#
+# #13610 residual (measured 2026-09-02, po-2024): the tail class was
+# `[A-Za-z0-9]+`, which excludes the underscore. A dotted symbol -- the most
+# common way a French technical body points at code -- was therefore NOT a
+# named referent, and the FN-safety branch kept the rouge on the FOUNDING
+# sentence of #13539, whose referent was named but named as a SYMBOL:
+#   « L'upsert vit dans `pick_idle_grain.upsert_orphans_comment` ; le
+#     generaliser demanderait d'editer un fichier deja porteur de deux PRs »
+# The boundary this drew was arbitrary AND invisible: `pick_idle_grain.upsert`
+# passed while `pick_idle_grain.upsert_orphans` rouged, on the sole strength
+# of one underscore -- two spellings of the same code reference, opposite
+# verdicts. Widening the tail to `\w+` removes the inversion. It does NOT
+# touch the deliberate FN-safety choice of #13612: an ANONYMOUS referent
+# ("editer un fichier", nothing named on the line) still keeps the rouge.
 _NAMED_FILE_BODY = re.compile(
-    r"(?:`([^`]+\.[A-Za-z0-9]+)`|"  # backticked: `pick_idle_grain.py`
+    r"(?:`([^`]+\.\w+)`|"  # backticked: `pick_idle_grain.py`, `mod.fn_name`
     r"\b([\w./-]+\.(?:py|cs|yml|yaml|json|md|ipynb|ts|js|sh|toml|cfg|ini))\b)"  # bare basename
 )
 # A cited threshold ("< 15 fichiers", ">= 10 fichiers") quotes a rule, it does
diff --git a/scripts/tests/test_check_pr_perimeter.py b/scripts/tests/test_check_pr_perimeter.py
index 048372d4d9..41369e14b0 100644
--- a/scripts/tests/test_check_pr_perimeter.py
+++ b/scripts/tests/test_check_pr_perimeter.py
@@ -2865,6 +2865,110 @@ def test_13610_predicate_unit_unanimous():
     ) is False  # named file IS in scope
 
 
+# ---------------------------------------------------------------------------
+# #13610 residual (2026-09-02, measured by po-2024) -- the referent is named,
+# but named as a SYMBOL, not as a file. The tail class of _NAMED_FILE_BODY was
+# `[A-Za-z0-9]+`, which excludes the underscore; a dotted symbol was therefore
+# not recognized as a named referent, and the FN-safety branch kept the rouge
+# on the FOUNDING sentence of #13539 itself. The boundary was arbitrary AND
+# invisible: the same code reference passed or rouged on the sole strength of
+# one underscore. The pair below is the proof -- the first test alone would be
+# satisfied by simply disabling the guard, which is why the positive control
+# and the anonymous-referent control are pinned alongside it.
+# ---------------------------------------------------------------------------
+
+# The two rows of the #13610 table, verbatim in shape: same sentence, same
+# referent, one underscore apart.
+_SYMBOL_SHAPE = (
+    "L'upsert vit dans `pick_idle_grain.{symbole}` ; le generaliser "
+    "demanderait d'editer un fichier deja porteur de deux PRs ouvertes "
+    "de la meme lane"
+)
+
+
+def test_13610_symbol_referent_without_underscore_passes():
+    """Table row 1: `pick_idle_grain.upsert` -- passed even before the fix.
+    Pinned so a future tightening of the tail class cannot silently take it
+    back, which would re-open the inversion from the other side."""
+    line = _SYMBOL_SHAPE.format(symbole="upsert")
+    problems = check_assertion(FILES_13610, line)
+    assert problems == [], (
+        "a dotted symbol without underscore must not rouge; got: "
+        + repr(problems)
+    )
+
+
+def test_13610_symbol_referent_with_underscore_passes():
+    """Table row 2: `pick_idle_grain.upsert_orphans` -- ROUGED before the fix,
+    on the sole strength of the underscore. This is the regression the
+    residual names."""
+    line = _SYMBOL_SHAPE.format(symbole="upsert_orphans")
+    problems = check_assertion(FILES_13610, line)
+    assert problems == [], (
+        "a dotted symbol with an underscore must not rouge -- the underscore "
+        "is not a semantic boundary; got: " + repr(problems)
+    )
+
+
+def test_13610_founding_sentence_of_13539_passes():
+    """The sentence of #13539 that founded the whole thread. Its referent was
+    named all along -- named as a symbol. Acceptance line 1 of #13610."""
+    line = _SYMBOL_SHAPE.format(symbole="upsert_orphans_comment")
+    problems = check_assertion(FILES_13610, line)
+    assert problems == [], (
+        "the founding sentence of #13539 must pass the guard; got: "
+        + repr(problems)
+    )
+
+
+def test_13610_symbol_widening_keeps_positive_control_rouge():
+    """Acceptance line 2 of #13610, stated as a PAIR with the three tests
+    above: widening the tail class must NOT extinguish the positive control.
+    A fix that made the symbol cases pass by weakening the guard would pass
+    those three and fail this one -- which is the whole point of pinning it
+    here rather than relying on the older copy elsewhere in the file."""
+    problems = check_assertion(FILES_13610, "Cette PR ne touche qu'un fichier.")
+    assert problems, (
+        "the positive control must still rouge after the widening -- "
+        "3 files in the PR, assertion claims 1"
+    )
+
+
+def test_13610_symbol_widening_keeps_anonymous_referent_rouge():
+    """FN-safety, restated post-widening: widening WHAT counts as a named
+    referent must not turn an ANONYMOUS referent into a named one. #13612's
+    deliberate default-fail-loud on ambiguous shapes is untouched."""
+    line = (
+        "le generaliser demanderait d'editer un fichier deja porteur de "
+        "deux PRs ouvertes"
+    )
+    problems = check_assertion(FILES_13610, line)
+    assert problems, (
+        "an anonymous referent must keep the rouge after the widening"
+    )
+
+
+def test_13610_symbol_predicate_unit_pair():
+    """Direct unit test of the predicate on the two table rows, decoupled from
+    check_assertion so a pipeline change cannot mask a predicate regression --
+    same rationale as test_13610_predicate_unit_unanimous."""
+    files = FILES_13610
+    assert _word_form_is_indef_non_pr_subject(
+        _SYMBOL_SHAPE.format(symbole="upsert"), files
+    ) is True
+    assert _word_form_is_indef_non_pr_subject(
+        _SYMBOL_SHAPE.format(symbole="upsert_orphans"), files
+    ) is True
+    assert _word_form_is_indef_non_pr_subject(
+        _SYMBOL_SHAPE.format(symbole="upsert_orphans_comment"), files
+    ) is True
+    # The widening is about the TAIL of a dotted referent, nothing else:
+    # an anonymous referent stays False.
+    assert _word_form_is_indef_non_pr_subject(
+        "editer un fichier quelque part", files
+    ) is False
+
+
 # ---------------------------------------------------------------------------
 # #13637 -- carried-from-main files. GitHub's /pulls/N/files diffs base-tip ->
 # head, so a branch that merged main gets main's own changes attributed to it