session: gate retrieval keywords on anaphora, scope to parent only

Two related fixes for fox-flagged "the convo gets narrowed":

1. Anaphora gate. The retrieval_keywords flow from ancestor cited
   titles into FTS5 was firing on every turn — even unrelated topic
   switches. Asking "when was batman comic created?" in a Sonic-heavy
   session pulled 8 Sonic titles into FTS5 and drowned the Batman
   article → UNGROUNDED. Now gated on the question containing a
   pronoun ("it", "him", "her", "they", "this", "that", "one"). For
   subject-bearing questions, retrieval stands on the question text
   alone — the LLM-side conversation_history (added 832297f) gives
   the model pronoun-resolution context, but retrieval doesn't need
   keywords when the new question already has a subject.

2. Parent-only scoping. When the question IS anaphoric, use ONLY the
   immediate parent's cited titles — not the union across all
   ancestors. Pronouns almost always refer to the most recent
   subject; accumulating across a long Sonic-then-Batman thread let
   Sonic dominate keyword weight and the model resolved "it" to
   Sonic instead of Batman. Parent-scoped keywords align with how
   English antecedent resolution actually works.

Live verified (3-turn smoke):
  Sonic created?            → HYBRID, Sonic article cited
  Batman comic created?     → HYBRID, Batman article cited
                              (was UNGROUNDED with bare-conv polluter)
  who created it?           → HYBRID, Bob Kane + Bill Finger via
                              List of Batman creators
                              (was Sonic Naoto Ōshima — wrong topic)

24 session+providence+run-query tests pass.

arborist/cli.py

@@ -7524,6 +7524,35 @@ _POINTER_LINE_RE = re.compile(
 )
 _INLINE_POINTER_RE = re.compile(r'\[E\d+(?:\s*\|[^\]]*?)?\]')
 
+# Pronouns / anaphoric markers the session-keyword gate looks for.
+# Conservative set — only words whose referent is almost certainly
+# in a prior turn. "the" / "a" / "what" are NOT here (those start
+# fresh-subject questions like "what is the capital of Paris?").
+_ANAPHORIC_WORDS = frozenset({
+    "it", "its", "itself",
+    "he", "him", "his", "himself",
+    "she", "her", "hers", "herself",
+    "they", "them", "their", "theirs", "themselves",
+    "this", "that", "these", "those",
+    "one", "ones",  # "which one was first?"
+})
+_WORD_RE = re.compile(r"[A-Za-z']+")
+
+
+def _question_needs_anaphor_keywords(question: str) -> bool:
+    """True iff the question contains a pronoun / anaphoric word whose
+    antecedent is presumed to live in the conversation history.
+
+    When True: ancestor cited-titles flow to FTS5 retrieval as keywords.
+    When False: retrieval stands on the question text alone — prevents
+    a Sonic-heavy prior conversation from polluting an unrelated
+    follow-up like "when was batman comic created?".
+    """
+    for m in _WORD_RE.finditer(question or ""):
+        if m.group(0).lower() in _ANAPHORIC_WORDS:
+            return True
+    return False
+
 
 def _session_strip_pointers(answer_text: str) -> str:
     """Strip claim-lattice pointer markers from a prior turn's answer
@@ -7870,7 +7899,6 @@ def _session_repl(args: argparse.Namespace, sess) -> int:
             #    questions ("who created it?" after "what is
             #    permaculture?"). Folds into conversation_hash so each
             #    distinct chat path gets its own cache_key.
-            seen_titles: list[str] = []
             history: list[dict] = []
             # current_bates is the PARENT of the soon-to-be-minted
             # node, so its path-to-root IS the conversation history.
@@ -7880,9 +7908,6 @@ def _session_repl(args: argparse.Namespace, sess) -> int:
             for n in reversed(ancestors):
                 if n.seq == 0:
                     continue  # skip synthetic root
-                for t in n.cited_titles:
-                    if t and t not in seen_titles:
-                        seen_titles.append(t)
                 history.append({"role": "user", "content": n.question})
                 # Strip claim-lattice pointer-line excerpts before
                 # passing prior answers to the model — the [E#] bindings
@@ -7894,8 +7919,30 @@ def _session_repl(args: argparse.Namespace, sess) -> int:
                     "content": _session_strip_pointers(n.answer_text),
                 })
             policy_local = dict(policy)
-            if seen_titles:
-                policy_local["retrieval_keywords"] = " ".join(seen_titles[:8])
+            # Retrieval keyword flow is GATED on the new question being
+            # anaphoric — only inject titles when the question itself
+            # has no subject (pronouns "it", "him", "her", "they",
+            # "this", "that"). Without the gate, asking "when was batman
+            # comic created?" in a session whose prior turns are about
+            # Sonic pulls Sonic-titled keywords into FTS5 and drowns
+            # the Batman article. The LLM-side conversation_history
+            # carries antecedent context for pronoun questions anyway;
+            # keywords are only needed when the question text alone is
+            # subject-less.
+            #
+            # When anaphoric, use ONLY the immediate parent turn's
+            # cited titles — pronouns almost always refer to the most
+            # recent subject. Accumulating titles across all ancestors
+            # lets a long Sonic thread dominate keyword weight when the
+            # parent is actually about Batman. (2026-06-01 fox-flagged
+            # conversation narrowing + antecedent drift.)
+            if _question_needs_anaphor_keywords(line):
+                parent = sess.get_node(sess.current_bates)
+                parent_titles = list(parent.cited_titles) if parent else []
+                if parent_titles:
+                    policy_local["retrieval_keywords"] = " ".join(
+                        parent_titles[:8]
+                    )
 
             try:
                 result = providence_query(