merge: T71 prompt.py polish (witness role + ACTIVITIES + NICE trim docs)

T18 review (Phase 1) noted the NICE-tier trim drops previous-scene
FIRST while §6.3 spec lists previous-scene LAST in the NICE tier
group. Decision: keep the existing greedy order (previous-scene
first), and document why.

Rationale (now in code at the trim ladder):
  1. Cheapest-impact-first — a per-POV previous-scene summary loses
     less narrative continuity than the older dialogue turns or
     memory hits it competes with.
  2. Greedy lookahead is more expensive than the marginal narrative
     loss. Dropping previous-scene typically clears the soft-budget
     slack in one step.

Test added: test_nice_trim_order_documented pins the observed order
(previous-scene -> memories -> dialogue) so a future refactor can't
silently invert it. Sized so that all-NICE config overflows soft but
dropping just previous-scene fits — proves memories and older
dialogue turns survive while previous-scene is the FIRST drop.

chat/db/connection.py

@@ -5,9 +5,9 @@ from pathlib import Path
 
 
 @contextmanager
-def open_db(path: Path):
+def open_db(path: Path, *, check_same_thread: bool = True):
     path.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(path)
+    conn = sqlite3.connect(path, check_same_thread=check_same_thread)
     conn.execute("PRAGMA journal_mode=WAL")
     conn.execute("PRAGMA foreign_keys=ON")
     try:

chat/services/prompt.py

@@ -273,6 +273,18 @@ def _resolve_previous_scene_summary(
     return mem[0]
 
 
+def _witness_role_for(speaker_bot_id: str, host_bot_id: str | None) -> str:
+    """Return the witness POV role for the speaker's memory query.
+
+    The host bot of a chat queries memories with ``witness_role="host"``;
+    the guest bot queries with ``witness_role="guest"``. Phase 2 T46
+    pinned the contract on ``search_memories``; this helper applies it
+    at the call site so a guest-as-speaker doesn't silently retrieve
+    memories under the wrong POV mask.
+    """
+    return "host" if speaker_bot_id == host_bot_id else "guest"
+
+
 def _resolve_addressee(
     conn: Connection, addressee: str, you: dict | None
 ) -> tuple[str, str]:
@@ -356,34 +368,57 @@ def assemble_narrative_prompt(
         addressee_name,
     )
 
-    # Activity for present entities. Core (MUST): you + speaker bot.
-    # Phase 2 (SHOULD-tier): when a third party (guest) is present in
-    # the chat, append their activity in a separate block so it can be
-    # trimmed independently under tight budget.
-    activities: list[dict] = []
+    # Activity for present entities — single ACTIVITIES: block with up
+    # to three bullets (you, speaker, guest). The block itself is
+    # MUST-tier and survives all trims, but bullet-level trim drops
+    # bullets in the order guest -> you, keeping the speaker bullet
+    # (the speaker's own current activity is the load-bearing slice).
+    #
+    # T71.2 chose Option B from the polish plan: pre-truncate the
+    # bullets list at trim time before _build_activity_block runs,
+    # rather than introducing a granular tier mode in the trim
+    # machinery. The single-block render avoids the dual-ACTIVITIES:
+    # header that Phase 2 T43 introduced (read by some LLMs as a
+    # duplicate-section bug).
+    you_activity: dict | None = None
     you_act = get_activity(conn, "you")
     if you_act is not None:
-        you_act = dict(you_act)
-        you_act["_display_name"] = (you or {}).get("name") or "you"
-        activities.append(you_act)
+        you_activity = dict(you_act)
+        you_activity["_display_name"] = (you or {}).get("name") or "you"
+
+    speaker_activity: dict | None = None
     bot_act = get_activity(conn, speaker_bot_id)
     if bot_act is not None:
-        bot_act = dict(bot_act)
-        bot_act["_display_name"] = bot["name"]
-        activities.append(bot_act)
-    activity_block = _build_activity_block(activities)
+        speaker_activity = dict(bot_act)
+        speaker_activity["_display_name"] = bot["name"]
 
-    # SHOULD-tier guest activity extension (Phase 2 / Task 43).
-    guest_activity_block: str | None = None
+    guest_activity: dict | None = None
     if guest_id is not None:
         guest_act = get_activity(conn, guest_id)
         if guest_act is not None:
-            guest_act = dict(guest_act)
+            guest_activity = dict(guest_act)
             guest_bot = get_bot(conn, guest_id)
-            guest_act["_display_name"] = (
+            guest_activity["_display_name"] = (
                 guest_bot["name"] if guest_bot else guest_id
             )
-            guest_activity_block = _build_activity_block([guest_act])
+
+    def _activity_block_for(
+        *, include_you: bool, include_guest: bool
+    ) -> str | None:
+        """Render the single ACTIVITIES: block with the requested bullets.
+
+        Speaker bullet is always included (it's the MUST-tier baseline);
+        ``you`` and ``guest`` bullets are toggled by the caller during
+        trim. Returns None when no bullets remain.
+        """
+        bullets: list[dict] = []
+        if include_you and you_activity is not None:
+            bullets.append(you_activity)
+        if speaker_activity is not None:
+            bullets.append(speaker_activity)
+        if include_guest and guest_activity is not None:
+            bullets.append(guest_activity)
+        return _build_activity_block(bullets)
 
     # SHOULD-tier group-node block (Phase 2 / Task 43): rendered only
     # when the group_node row is present AND it covers all three of
@@ -433,7 +468,12 @@ def assemble_narrative_prompt(
         memory_summaries = []
         if query:
             try:
-                hits = search_memories(conn, speaker_bot_id, "host", query, k=4)
+                witness_role = _witness_role_for(
+                    speaker_bot_id, chat.get("host_bot_id")
+                )
+                hits = search_memories(
+                    conn, speaker_bot_id, witness_role, query, k=4
+                )
                 memory_summaries = [h["pov_summary"] for h in hits]
             except Exception:
                 memory_summaries = []
@@ -452,11 +492,18 @@ def assemble_narrative_prompt(
     last4 = dialogue_full[-4:] if dialogue_full else []
     must_dialogue_block = _build_dialogue_block(last4, earlier_summary=None)
 
+    # MUST-tier ACTIVITIES floor: the speaker bullet alone (you and
+    # guest bullets are dropped first under bullet-level trim before
+    # the block bottoms out at speaker-only).
+    must_activity_block = _activity_block_for(
+        include_you=False, include_guest=False
+    )
+
     must_blocks: list[str | None] = [
         speaker_identity,
         edge_to_addressee,
         scene_block,
-        activity_block,
+        must_activity_block,
         must_dialogue_block,
         closing,
     ]
@@ -481,6 +528,7 @@ def assemble_narrative_prompt(
         include_previous_scene: bool,
         include_memories_top_k: int,
         dialogue_keep: int,
+        include_you_activity: bool = True,
         include_guest_activity: bool = True,
         include_group_node: bool = True,
     ) -> tuple[str, int, list[dict]]:
@@ -503,13 +551,20 @@ def assemble_narrative_prompt(
             if include_previous_scene else None
         )
 
+        # Single ACTIVITIES: block, bullet-level trim (T71.2). Guest
+        # bullet drops first, then the you bullet; speaker bullet is the
+        # MUST-tier floor and always present when an activity row exists.
+        activity_block = _activity_block_for(
+            include_you=include_you_activity,
+            include_guest=include_guest_activity,
+        )
+
         body = _join_blocks([
             speaker_identity,
             edge_to_addressee,
             other_edges_block if include_other_edges else None,
             scene_block,
             activity_block,
-            guest_activity_block if include_guest_activity else None,
             group_node_block if include_group_node else None,
             prev_block,
             memories_block,
@@ -527,16 +582,18 @@ def assemble_narrative_prompt(
     nice_memories_k = min(4, len(memory_summaries))
     include_prev = previous_scene_summary is not None
     include_other = other_edges_block is not None
-    include_guest_activity = guest_activity_block is not None
+    include_you_activity = you_activity is not None
+    include_guest_activity = guest_activity is not None
     include_group_node = group_node_block is not None
 
     def _build(*, prev: bool, mem_k: int, dlg: int, other: bool,
-               guest_act: bool, group: bool) -> tuple[str, int]:
+               you_act: bool, guest_act: bool, group: bool) -> tuple[str, int]:
         body, total, _ = assemble(
             include_other_edges=other,
             include_previous_scene=prev,
             include_memories_top_k=mem_k,
             dialogue_keep=dlg,
+            include_you_activity=you_act,
             include_guest_activity=guest_act,
             include_group_node=group,
         )
@@ -544,8 +601,8 @@ def assemble_narrative_prompt(
 
     body, total = _build(
         prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
-        other=include_other, guest_act=include_guest_activity,
-        group=include_group_node,
+        other=include_other, you_act=include_you_activity,
+        guest_act=include_guest_activity, group=include_group_node,
     )
 
     # If under soft, we're done.
@@ -554,12 +611,32 @@ def assemble_narrative_prompt(
 
     # Drop NICE in order: previous scene → memories beyond top-2 →
     # older dialogue turns (collapse to 4).
+    #
+    # T71.3 — order rationale: the §6.3 spec lists NICE-tier members
+    # with previous-scene LAST, which read as a literal trim order
+    # during T18 review. We deliberately keep the greedy order shown
+    # here (previous-scene FIRST) for two reasons:
+    #
+    #   1. Cheapest-impact-first: a per-POV previous-scene summary is
+    #      a single short paragraph that loses very little narrative
+    #      continuity when dropped, while the older dialogue turns it
+    #      is competing with carry the speaker's last few beats — those
+    #      ground the next response far more concretely.
+    #   2. Greedy lookahead is more expensive than the marginal
+    #      narrative loss. Dropping previous-scene typically clears
+    #      the soft-budget slack in one step; trying memories or
+    #      dialogue first would routinely require multiple recompute
+    #      passes through the assembler.
+    #
+    # The pin test test_nice_trim_order_documented locks this order so
+    # a future refactor can't quietly invert it without surfacing the
+    # decision.
     if include_prev:
         include_prev = False
         body, total = _build(
             prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
-            other=include_other, guest_act=include_guest_activity,
-            group=include_group_node,
+            other=include_other, you_act=include_you_activity,
+            guest_act=include_guest_activity, group=include_group_node,
         )
         if total <= budget_soft:
             return _emit(body, user_turn_prose)
@@ -568,8 +645,8 @@ def assemble_narrative_prompt(
         nice_memories_k = 2
         body, total = _build(
             prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
-            other=include_other, guest_act=include_guest_activity,
-            group=include_group_node,
+            other=include_other, you_act=include_you_activity,
+            guest_act=include_guest_activity, group=include_group_node,
         )
         if total <= budget_soft:
             return _emit(body, user_turn_prose)
@@ -578,8 +655,8 @@ def assemble_narrative_prompt(
         nice_dialogue_keep = baseline_keep
         body, total = _build(
             prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
-            other=include_other, guest_act=include_guest_activity,
-            group=include_group_node,
+            other=include_other, you_act=include_you_activity,
+            guest_act=include_guest_activity, group=include_group_node,
         )
         if total <= budget_soft:
             return _emit(body, user_turn_prose)
@@ -589,35 +666,47 @@ def assemble_narrative_prompt(
         nice_memories_k = max(0, nice_memories_k - 1)
         body, total = _build(
             prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
-            other=include_other, guest_act=include_guest_activity,
-            group=include_group_node,
+            other=include_other, you_act=include_you_activity,
+            guest_act=include_guest_activity, group=include_group_node,
         )
 
-    # Drop SHOULD-tier blocks in order: guest activity → group node →
-    # other edges. (Guest activity goes first per Task 43 spec — it's
-    # the most expendable additive context.)
+    # Drop SHOULD-tier extras in order:
+    #   1. guest activity bullet (T71.2: bullet-level trim within the
+    #      single ACTIVITIES: block — guest goes first per Task 43 spec)
+    #   2. group node block
+    #   3. you activity bullet (still SHOULD-tier; speaker bullet is the
+    #      MUST-tier floor and never dropped)
+    #   4. other edges
     if include_guest_activity and total > budget_hard:
         include_guest_activity = False
         body, total = _build(
             prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
-            other=include_other, guest_act=include_guest_activity,
-            group=include_group_node,
+            other=include_other, you_act=include_you_activity,
+            guest_act=include_guest_activity, group=include_group_node,
         )
 
     if include_group_node and total > budget_hard:
         include_group_node = False
         body, total = _build(
             prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
-            other=include_other, guest_act=include_guest_activity,
-            group=include_group_node,
+            other=include_other, you_act=include_you_activity,
+            guest_act=include_guest_activity, group=include_group_node,
+        )
+
+    if include_you_activity and total > budget_hard:
+        include_you_activity = False
+        body, total = _build(
+            prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
+            other=include_other, you_act=include_you_activity,
+            guest_act=include_guest_activity, group=include_group_node,
         )
 
     if include_other and total > budget_hard:
         include_other = False
         body, total = _build(
             prev=include_prev, mem_k=nice_memories_k, dlg=nice_dialogue_keep,
-            other=include_other, guest_act=include_guest_activity,
-            group=include_group_node,
+            other=include_other, you_act=include_you_activity,
+            guest_act=include_guest_activity, group=include_group_node,
         )
 
     if total > budget_hard:

chat/services/scene_summarize.py

@@ -334,26 +334,92 @@ async def apply_scene_close_summary(
             timeout_s=timeout_s,
         )
 
-    # Group node update: naive per-POV concat for v2. Only fires when
-    # both POVs ran (i.e. the guest is present) and a group_node row
-    # exists for this chat.
+    # Group node update: T70 runs a third classifier call to merge the
+    # two per-POV summaries into a coherent group-level view + a brief
+    # group-dynamic note. Falls back to the Phase 2 naive concat on
+    # classifier failure (see :func:`merge_group_summary`). Only fires
+    # when both POVs ran (i.e. the guest is present) and a group_node
+    # row exists for this chat.
     if guest_pov is not None and get_group_node(conn, chat_id) is not None:
         host_bot = get_bot(conn, host_bot_id) or {"name": host_bot_id}
         guest_bot = get_bot(conn, guest_bot_id) or {"name": guest_bot_id}
         host_name = host_bot.get("name", host_bot_id) or host_bot_id
         guest_name = guest_bot.get("name", guest_bot_id) or guest_bot_id
-        group_summary = (
-            f"{host_name}: {host_pov.summary}\n\n"
-            f"{guest_name}: {guest_pov.summary}"
+        merged = await merge_group_summary(
+            client,
+            classifier_model=classifier_model,
+            host_name=host_name,
+            host_pov_summary=host_pov.summary,
+            guest_name=guest_name,
+            guest_pov_summary=guest_pov.summary,
+            timeout_s=timeout_s,
         )
         append_and_apply(
             conn,
             kind="group_node_updated",
             payload={
                 "chat_id": chat_id,
-                "summary": group_summary,
-                "dynamic": "",
+                "summary": merged.summary,
+                "dynamic": merged.dynamic,
             },
         )
 
     return host_pov
+
+
+class GroupMetaSummary(BaseModel):
+    """Classifier output: a merged group-level view of a closed scene.
+
+    Defaults are an empty no-op so callers can use the schema's default
+    as a sentinel; in practice :func:`merge_group_summary` builds an
+    explicit naive-concat fallback rather than returning these defaults
+    directly so existing Phase 2 behavior is preserved on classifier
+    failure.
+    """
+
+    summary: str = ""
+    dynamic: str = ""
+
+
+_GROUP_MERGE_SYSTEM = (
+    "Given two per-POV scene summaries from a 3-entity scene (you + "
+    "host + guest), produce a coherent group-level summary capturing "
+    "the shared events as both witnesses experienced them, plus a "
+    "brief 'dynamic' note describing the trio's group dynamic during "
+    "the scene. Output strict JSON matching schema."
+)
+
+
+async def merge_group_summary(
+    client: LLMClient,
+    *,
+    classifier_model: str,
+    host_name: str,
+    host_pov_summary: str,
+    guest_name: str,
+    guest_pov_summary: str,
+    timeout_s: float = 30.0,
+) -> GroupMetaSummary:
+    """Merge two per-POV scene summaries into a coherent group-level
+    summary + group-dynamic note. Falls back to the naive concat (the
+    existing behavior) on classifier failure."""
+    user = (
+        f"{host_name} (host) POV summary:\n{host_pov_summary}\n\n"
+        f"{guest_name} (guest) POV summary:\n{guest_pov_summary}"
+    )
+    fallback = GroupMetaSummary(
+        summary=(
+            f"{host_name}: {host_pov_summary}\n\n"
+            f"{guest_name}: {guest_pov_summary}"
+        ),
+        dynamic="",
+    )
+    return await classify(
+        client,
+        model=classifier_model,
+        system=_GROUP_MERGE_SYSTEM,
+        user=user,
+        schema=GroupMetaSummary,
+        default=fallback,
+        timeout_s=timeout_s,
+    )

chat/state/entities.py

@@ -48,6 +48,17 @@ def _apply_bot_reset(conn: Connection, e: Event) -> None:
             "SELECT id FROM chats WHERE host_bot_id = ?", (bot_id,)
         ).fetchall()
     ]
+    # T69: purge orphaned "you" activity rows pointing at containers in this
+    # bot's chats BEFORE the containers/chats themselves are deleted, otherwise
+    # the subqueries find nothing and the FK constraint on activity.container_id
+    # blocks the container delete.
+    conn.execute(
+        "DELETE FROM activity WHERE entity_id = 'you' "
+        "AND container_id IN (SELECT id FROM containers WHERE chat_id IN ("
+        "    SELECT id FROM chats WHERE host_bot_id = ?"
+        "))",
+        (bot_id,),
+    )
     for chat_id in chat_ids:
         conn.execute("DELETE FROM scenes WHERE chat_id = ?", (chat_id,))
         conn.execute("DELETE FROM containers WHERE chat_id = ?", (chat_id,))
@@ -74,8 +85,6 @@ def _apply_bot_reset(conn: Connection, e: Event) -> None:
         (bot_id,),
     )
     # NOTE: bots row itself is preserved (identity, kickoff_prose intact).
-    # NOTE: "you" activity (entity_id="you") may linger from a deleted chat;
-    # acceptable for v1 — Phase 1.5 cleanup if needed.
 
 
 def get_bot(conn: Connection, bot_id: str) -> dict | None:

chat/web/bots.py

@@ -1,10 +1,10 @@
 from __future__ import annotations
-import sqlite3
 from pathlib import Path
 from fastapi import APIRouter, Depends, Form, HTTPException, Request
 from fastapi.responses import RedirectResponse, HTMLResponse
 from fastapi.templating import Jinja2Templates
 
+from chat.db.connection import open_db
 from chat.eventlog.log import append_event
 from chat.eventlog.projector import project
 from chat.state.entities import list_bots
@@ -19,15 +19,8 @@ REQUIRED_FIELDS = ("id", "name", "persona", "initial_relationship_to_you", "kick
 def get_conn(request: Request):
     settings = request.app.state.settings
     db_path: Path = settings.db_path
-    db_path.parent.mkdir(parents=True, exist_ok=True)
-    conn = sqlite3.connect(db_path, check_same_thread=False)
-    conn.execute("PRAGMA journal_mode=WAL")
-    conn.execute("PRAGMA foreign_keys=ON")
-    try:
+    with open_db(db_path, check_same_thread=False) as conn:
         yield conn
-        conn.commit()
-    finally:
-        conn.close()
 
 
 def _split_voice_samples(text: str) -> list[str]:

tests/test_open_db_threading.py

@@ -0,0 +1,57 @@
+from __future__ import annotations
+import sqlite3
+import threading
+
+from chat.db.connection import open_db
+
+
+def test_open_db_default_uses_check_same_thread_true(tmp_path):
+    """Default open_db must reject cross-thread use (safe default)."""
+    db = tmp_path / "t.db"
+    captured: list[BaseException | None] = []
+
+    with open_db(db) as conn:
+        conn.execute("CREATE TABLE t (x INTEGER)")
+
+        def worker():
+            try:
+                conn.execute("SELECT 1").fetchall()
+                captured.append(None)
+            except BaseException as e:  # noqa: BLE001
+                captured.append(e)
+
+        t = threading.Thread(target=worker)
+        t.start()
+        t.join()
+
+    assert len(captured) == 1
+    err = captured[0]
+    assert isinstance(err, sqlite3.ProgrammingError), (
+        f"expected sqlite3.ProgrammingError on cross-thread use, got {err!r}"
+    )
+
+
+def test_open_db_can_disable_check_same_thread(tmp_path):
+    """open_db(check_same_thread=False) must allow cross-thread use."""
+    db = tmp_path / "t.db"
+    captured: list[BaseException | None] = []
+    rows: list[object] = []
+
+    with open_db(db, check_same_thread=False) as conn:
+        conn.execute("CREATE TABLE t (x INTEGER)")
+        conn.execute("INSERT INTO t (x) VALUES (42)")
+
+        def worker():
+            try:
+                result = conn.execute("SELECT x FROM t").fetchall()
+                rows.extend(result)
+                captured.append(None)
+            except BaseException as e:  # noqa: BLE001
+                captured.append(e)
+
+        t = threading.Thread(target=worker)
+        t.start()
+        t.join()
+
+    assert captured == [None], f"worker thread raised: {captured}"
+    assert rows == [(42,)]

tests/test_per_pov_summary.py

@@ -636,8 +636,10 @@ async def test_close_with_guest_updates_both_edges(tmp_path):
 @pytest.mark.asyncio
 async def test_close_with_group_node_updates_group_summary(tmp_path):
     """When a group_node row exists, scene close emits group_node_updated
-    with a non-empty summary that mentions both bots' names (v2 naive
-    concat of per-POV summaries)."""
+    with a non-empty summary that mentions both bots' names. T70 swapped
+    the Phase 2 naive concat for an LLM-merged summary; this regression
+    test feeds bad-JSON merge responses so the helper falls back to the
+    original naive-concat shape, preserving the original assertions."""
     db = tmp_path / "t.db"
     apply_migrations(db)
     import chat.state.group_node  # noqa: F401  -- register handlers
@@ -660,7 +662,11 @@ async def test_close_with_group_node_updates_group_summary(tmp_path):
         _seed_two_bot_scene(conn, with_group_node=True)
         project(conn)
 
-        client = MockLLMClient(canned=[host_canned, guest_canned])
+        # 2 valid (host POV, guest POV) + 3 bad-JSON merge attempts ->
+        # merge_group_summary falls back to the naive concat default.
+        client = MockLLMClient(
+            canned=[host_canned, guest_canned, "bad1", "bad2", "bad3"]
+        )
         await apply_scene_close_summary(
             conn,
             client,
@@ -675,8 +681,167 @@ async def test_close_with_group_node_updates_group_summary(tmp_path):
         gn = get_group_node(conn, "chat_bot_a")
         assert gn is not None
         assert gn["summary"]  # non-empty
-        # Naive concat surfaces both bot names in the group summary.
+        # Naive-concat fallback surfaces both bot names in the group summary.
         assert "BotA" in gn["summary"]
         assert "BotB" in gn["summary"]
-        # Phase 2 v2 keeps dynamic empty (Phase 3 polishes).
+        # Naive-concat fallback keeps dynamic empty.
         assert gn["dynamic"] == ""
+
+
+# ---------------------------------------------------------------------------
+# T70: LLM-merged group meta-summary on scene close.
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_group_summary_merges_per_pov_via_classifier_when_guest_present(
+    tmp_path,
+):
+    """With a guest present and a group_node row, scene close runs the
+    merge classifier as a third call after the two per-POV summarize_scene
+    calls; its output drives the group_node summary + dynamic fields."""
+    db = tmp_path / "t.db"
+    apply_migrations(db)
+    import chat.state.group_node  # noqa: F401  -- register handlers
+
+    host_canned = json.dumps(
+        {
+            "summary": "BotA appreciated the calm.",
+            "knowledge_facts": [],
+            "relationship_summary": "BotA felt steady.",
+        }
+    )
+    guest_canned = json.dumps(
+        {
+            "summary": "BotB found the room friendly.",
+            "knowledge_facts": [],
+            "relationship_summary": "BotB warmed up.",
+        }
+    )
+    merge_canned = json.dumps(
+        {"summary": "merged group view", "dynamic": "warm rapport"}
+    )
+    with open_db(db) as conn:
+        _seed_two_bot_scene(conn, with_group_node=True)
+        project(conn)
+
+        # Canned-queue layout matches the production call order in
+        # apply_scene_close_summary: host POV summarize_scene runs first,
+        # then guest POV summarize_scene, then merge_group_summary.
+        client = MockLLMClient(
+            canned=[host_canned, guest_canned, merge_canned]
+        )
+        await apply_scene_close_summary(
+            conn,
+            client,
+            classifier_model="x",
+            chat_id="chat_bot_a",
+            scene_id=1,
+            host_bot_id="bot_a",
+        )
+
+        # All three canned entries consumed -> classifier ran exactly 3x.
+        assert client._canned == []
+
+        from chat.state.group_node import get_group_node
+
+        gn = get_group_node(conn, "chat_bot_a")
+        assert gn is not None
+        assert gn["summary"] == "merged group view"
+        assert gn["dynamic"] == "warm rapport"
+
+
+@pytest.mark.asyncio
+async def test_group_summary_falls_back_to_naive_concat_on_classifier_failure(
+    tmp_path,
+):
+    """If the merge classifier flaps (bad JSON across all 3 retries), the
+    helper falls back to the original Phase 2 naive concat shape and
+    leaves dynamic empty."""
+    db = tmp_path / "t.db"
+    apply_migrations(db)
+    import chat.state.group_node  # noqa: F401  -- register handlers
+
+    host_canned = json.dumps(
+        {
+            "summary": "BotA appreciated the calm.",
+            "knowledge_facts": [],
+            "relationship_summary": "BotA felt steady.",
+        }
+    )
+    guest_canned = json.dumps(
+        {
+            "summary": "BotB found the room friendly.",
+            "knowledge_facts": [],
+            "relationship_summary": "BotB warmed up.",
+        }
+    )
+    with open_db(db) as conn:
+        _seed_two_bot_scene(conn, with_group_node=True)
+        project(conn)
+
+        # 2 valid POV summaries + 3 bad-JSON merge attempts trip the
+        # classifier's retry-then-default path; the default is the naive
+        # concat fallback.
+        client = MockLLMClient(
+            canned=[host_canned, guest_canned, "bad1", "bad2", "bad3"]
+        )
+        await apply_scene_close_summary(
+            conn,
+            client,
+            classifier_model="x",
+            chat_id="chat_bot_a",
+            scene_id=1,
+            host_bot_id="bot_a",
+        )
+
+        from chat.state.group_node import get_group_node
+
+        gn = get_group_node(conn, "chat_bot_a")
+        assert gn is not None
+        expected = (
+            "BotA: BotA appreciated the calm.\n\n"
+            "BotB: BotB found the room friendly."
+        )
+        assert gn["summary"] == expected
+        assert gn["dynamic"] == ""
+
+
+@pytest.mark.asyncio
+async def test_group_summary_skipped_when_no_guest(tmp_path):
+    """No-guest path: scene close does NOT invoke merge_group_summary
+    and emits no group_node_updated event. Confirms the existing
+    `if guest_bot_id is not None` gating at the call site."""
+    db = tmp_path / "t.db"
+    apply_migrations(db)
+    canned = json.dumps(
+        {
+            "summary": "BotA helped you talk through the deadline anxiety.",
+            "knowledge_facts": ["Deadline next Friday."],
+            "relationship_summary": "BotA leaned in supportively.",
+        }
+    )
+    with open_db(db) as conn:
+        _seed_single_bot_scene(conn)
+        project(conn)
+
+        # Only 1 canned entry; if merge_group_summary were called the
+        # MockLLMClient would IndexError on the empty queue.
+        client = MockLLMClient(canned=[canned])
+        await apply_scene_close_summary(
+            conn,
+            client,
+            classifier_model="x",
+            chat_id="chat_bot_a",
+            scene_id=1,
+            host_bot_id="bot_a",
+        )
+
+        # Exactly the host POV call consumed, nothing else.
+        assert client._canned == []
+
+        # No group_node_updated event was emitted.
+        rows = conn.execute(
+            "SELECT 1 FROM event_log WHERE kind = 'group_node_updated'"
+        ).fetchall()
+        assert rows == []

tests/test_prompt.py

@@ -452,6 +452,273 @@ def test_assemble_when_speaker_is_guest_orients_edges_correctly(tmp_path):
     assert "68/100" in body
 
 
+def test_speaker_is_guest_uses_guest_witness_role(tmp_path, monkeypatch):
+    """T71.1: when the guest is the speaker, ``search_memories`` is
+    called with ``witness_role="guest"``, not the previously-hardcoded
+    ``"host"``. Pins the parametric witness role at the prompt call site
+    so guest-as-speaker honours the witness mask via Phase 2 T46.
+    """
+    db = tmp_path / "t.db"
+    apply_migrations(db)
+    captured: dict = {}
+
+    def _fake_search(conn, owner_id, witness_role, query, k=4):
+        captured["owner_id"] = owner_id
+        captured["witness_role"] = witness_role
+        captured["query"] = query
+        return []
+
+    # Patch the imported reference inside the prompt module so the
+    # production call site uses the fake.
+    import chat.services.prompt as prompt_mod
+    monkeypatch.setattr(prompt_mod, "search_memories", _fake_search)
+
+    with open_db(db) as conn:
+        _seed_with_guest(conn)
+        # Guest as speaker — must request memories with witness_role="guest".
+        assemble_narrative_prompt(
+            conn,
+            chat_id="chat_bot_a",
+            speaker_bot_id="bot_b",
+            recent_dialogue=[],
+            # retrieved_memory_summaries=None forces the search path.
+            retrieved_memory_summaries=None,
+        )
+    assert captured["owner_id"] == "bot_b"
+    assert captured["witness_role"] == "guest"
+
+
+def test_speaker_is_host_uses_host_witness_role(tmp_path, monkeypatch):
+    """T71.1 (regression): host-as-speaker still queries with
+    ``witness_role="host"``."""
+    db = tmp_path / "t.db"
+    apply_migrations(db)
+    captured: dict = {}
+
+    def _fake_search(conn, owner_id, witness_role, query, k=4):
+        captured["witness_role"] = witness_role
+        return []
+
+    import chat.services.prompt as prompt_mod
+    monkeypatch.setattr(prompt_mod, "search_memories", _fake_search)
+
+    with open_db(db) as conn:
+        _seed_with_guest(conn)
+        assemble_narrative_prompt(
+            conn,
+            chat_id="chat_bot_a",
+            speaker_bot_id="bot_a",  # host as speaker
+            recent_dialogue=[],
+            retrieved_memory_summaries=None,
+        )
+    assert captured["witness_role"] == "host"
+
+
+def test_single_activities_block_with_three_bullets_when_3_entities(tmp_path):
+    """T71.2: with you + host + guest present, the assembled prompt
+    contains exactly ONE ``ACTIVITIES:`` header and bullets for all
+    three entities (no duplicate header from the prior dual-block
+    rendering).
+    """
+    db = tmp_path / "t.db"
+    apply_migrations(db)
+    with open_db(db) as conn:
+        _seed_with_guest(conn)
+        msgs = assemble_narrative_prompt(
+            conn,
+            chat_id="chat_bot_a",
+            speaker_bot_id="bot_a",
+            recent_dialogue=[],
+            retrieved_memory_summaries=[],
+        )
+    body = msgs[0].content
+    # Exactly one ACTIVITIES: header.
+    assert body.count("ACTIVITIES:") == 1
+    # Bullets for all three entities (you=Sam, host=Aria, guest=Iris)
+    # — pin on the unique action verbs from the seed data.
+    assert "finishing emails" in body  # you bullet
+    assert "pretending to work" in body  # speaker (host) bullet
+    assert "smirking-distinctively" in body  # guest bullet
+
+
+def test_tight_budget_drops_guest_activity_bullet_first(tmp_path):
+    """T71.2: under tight budget the speaker bullet survives but the
+    guest activity bullet is the first ACTIVITIES: bullet to drop. The
+    block as a whole stays present (it's MUST-tier); only its body
+    contracts.
+    """
+    db = tmp_path / "t.db"
+    apply_migrations(db)
+    with open_db(db) as conn:
+        _seed_with_guest(conn)
+        dialogue = [
+            {"speaker": "you", "text": "line-16 hi there"},
+            {"speaker": "bot_a", "text": "line-17 hey"},
+            {"speaker": "you", "text": "line-18 quiet night"},
+            {"speaker": "bot_a", "text": "line-19 indeed"},
+        ]
+        msgs = assemble_narrative_prompt(
+            conn,
+            chat_id="chat_bot_a",
+            speaker_bot_id="bot_a",
+            recent_dialogue=dialogue,
+            retrieved_memory_summaries=[],
+            budget_soft=250,
+            budget_hard=340,
+        )
+    body = msgs[0].content
+    # Speaker bullet survives (MUST-tier floor).
+    assert "pretending to work" in body
+    assert "ACTIVITIES:" in body
+    # Guest bullet is dropped first under budget pressure.
+    assert "smirking-distinctively" not in body
+
+
+def test_nice_trim_order_documented(tmp_path):
+    """T71.3: pin the NICE-tier trim order so a future refactor can't
+    quietly invert it.
+
+    Order under NICE pressure is:
+      1. previous-scene summary (dropped FIRST)
+      2. memories beyond top-2
+      3. older dialogue turns (collapsed to last-4)
+
+    We size the budget so that all-NICE-included is over soft, but
+    dropping ONLY previous-scene gets us back under soft. The observed
+    behaviour we pin: previous-scene gone, memories/dialogue intact.
+    """
+    db = tmp_path / "t.db"
+    apply_migrations(db)
+    # Heavy previous-scene summary — large enough that dropping it
+    # alone clears the soft-budget overage. Defined out here so the
+    # marker is in scope for the assertions below.
+    prev_scene_blob = "PREVSCENE-MARKER " + ("filler " * 200)
+    with open_db(db) as conn:
+        # Append all events first, project once at the end (project is
+        # not idempotent — it replays every event in the log).
+        from chat.eventlog.log import append_event as _append
+        _append(conn, kind="bot_authored", payload={
+            "id": "bot_a",
+            "name": "Aria",
+            "persona": "reserved coworker who notices things",
+            "voice_samples": ["I — sorry, I didn't mean to."],
+            "traits": ["introverted"],
+            "backstory": "An archivist who joined the firm last spring.",
+            "initial_relationship_to_you": "coworker",
+            "kickoff_prose": "you stay late at the office",
+        })
+        _append(conn, kind="you_authored", payload={
+            "name": "Sam",
+            "pronouns": "they/them",
+            "persona": "tired analyst",
+        })
+        _append(conn, kind="chat_created", payload={
+            "id": "chat_bot_a",
+            "host_bot_id": "bot_a",
+            "guest_bot_id": None,
+            "initial_time": "2026-04-26T20:00:00+00:00",
+            "narrative_anchor": "Day 1 evening",
+            "weather": "clear",
+        })
+        _append(conn, kind="container_created", payload={
+            "chat_id": "chat_bot_a",
+            "name": "office bullpen",
+            "type": "workplace",
+            "properties": {"public": False, "moving": False, "audible_range": "room"},
+        })
+        _append(conn, kind="edge_update", payload={
+            "source_id": "bot_a",
+            "target_id": "you",
+            "affinity_delta": 12,
+            "trust_delta": 5,
+            "knowledge_facts": ["they work on the same floor"],
+        })
+        _append(conn, kind="activity_change", payload={
+            "entity_id": "you",
+            "container_id": 1,
+            "posture": "sitting at your desk",
+            "action": {"verb": "finishing emails"},
+            "attention": "the screen",
+        })
+        _append(conn, kind="activity_change", payload={
+            "entity_id": "bot_a",
+            "container_id": 1,
+            "posture": "sitting at her desk",
+            "action": {"verb": "pretending to work"},
+            "attention": "you, in glances",
+        })
+        _append(conn, kind="scene_opened", payload={
+            "chat_id": "chat_bot_a",
+            "container_id": 1,
+            "started_at": "2026-04-26T20:00:00+00:00",
+            "participants": ["you", "bot_a"],
+        })
+        # Close the seeded scene and write a per-POV summary memory so
+        # _resolve_previous_scene_summary returns a non-empty string.
+        _append(conn, kind="scene_closed", payload={
+            "scene_id": 1,
+            "ended_at": "2026-04-26T20:30:00+00:00",
+            "significance": 2,
+        })
+        _append(conn, kind="memory_written", payload={
+            "owner_id": "bot_a",
+            "chat_id": "chat_bot_a",
+            "scene_id": 1,
+            "pov_summary": prev_scene_blob,
+            "witness_you": 1,
+            "witness_host": 1,
+            "witness_guest": 0,
+            "source": "direct",
+            "reliability": 1.0,
+            "significance": 2,
+        })
+        project(conn)
+
+        # Six dialogue turns — last 4 plus 2 older. If older turns are
+        # dropped under NICE pressure, the unique markers for turns 0/1
+        # disappear; we'll assert they REMAIN to prove dialogue trim
+        # didn't fire.
+        dialogue = [
+            {"speaker": "you", "text": "DLG-OLD-00 hello"},
+            {"speaker": "bot_a", "text": "DLG-OLD-01 hi"},
+            {"speaker": "you", "text": "DLG-LAST-16 ok"},
+            {"speaker": "bot_a", "text": "DLG-LAST-17 sure"},
+            {"speaker": "you", "text": "DLG-LAST-18 night"},
+            {"speaker": "bot_a", "text": "DLG-LAST-19 indeed"},
+        ]
+        # Four small memories — if "memories beyond top-2" trim fires,
+        # MEM-C/MEM-D disappear; we'll assert they REMAIN to prove
+        # memories trim didn't fire either.
+        memories = ["MEM-A short", "MEM-B short", "MEM-C short", "MEM-D short"]
+
+        # Soft tuned so the all-NICE config (with the heavy previous
+        # scene summary) overflows, but dropping just previous-scene
+        # fits comfortably. Hard set high so SHOULD-tier never trims.
+        msgs = assemble_narrative_prompt(
+            conn,
+            chat_id="chat_bot_a",
+            speaker_bot_id="bot_a",
+            recent_dialogue=dialogue,
+            retrieved_memory_summaries=memories,
+            budget_soft=400,
+            budget_hard=8000,
+        )
+    body = msgs[0].content
+    # Previous-scene summary was the FIRST NICE drop — its unique
+    # marker must be absent.
+    assert "PREVSCENE-MARKER" not in body
+    # Memories beyond top-2 stayed (proves memories trim did NOT fire).
+    assert "MEM-A" in body
+    assert "MEM-B" in body
+    assert "MEM-C" in body
+    assert "MEM-D" in body
+    # Older dialogue turns stayed (proves dialogue trim did NOT fire).
+    assert "DLG-OLD-00" in body
+    assert "DLG-OLD-01" in body
+    # Last-4 dialogue turns of course present.
+    assert "DLG-LAST-19" in body
+
+
 def test_assemble_with_tight_budget_drops_guest_activity_first(tmp_path):
     """Under tight budget MUST blocks survive but SHOULD-tier guest
     activity is dropped first."""

tests/test_reset.py

@@ -292,6 +292,189 @@ def test_reset_clears_guest_reference_in_other_chats(client, tmp_path):
         ).fetchone()[0] == 1
 
 
+def test_reset_purges_orphaned_you_activity_rows(client, tmp_path):
+    """T69: when a bot's chats are deleted, "you" activity rows tied to those
+    chats' containers should also be purged (otherwise they linger orphaned)."""
+    db = tmp_path / "test.db"
+    with open_db(db) as conn:
+        append_event(
+            conn,
+            kind="bot_authored",
+            payload={
+                "id": "bot_a",
+                "name": "BotA",
+                "persona": "thoughtful",
+                "voice_samples": [],
+                "traits": [],
+                "backstory": "",
+                "initial_relationship_to_you": "coworker",
+                "kickoff_prose": "",
+            },
+        )
+        append_event(
+            conn,
+            kind="chat_created",
+            payload={
+                "id": "chat_bot_a",
+                "host_bot_id": "bot_a",
+                "initial_time": "2026-04-26T20:00:00+00:00",
+                "narrative_anchor": "Day 1",
+                "weather": "",
+            },
+        )
+        append_event(
+            conn,
+            kind="container_created",
+            payload={
+                "chat_id": "chat_bot_a",
+                "name": "office",
+                "type": "workplace",
+                "properties": {},
+            },
+        )
+        append_event(
+            conn,
+            kind="activity_change",
+            payload={
+                "entity_id": "you",
+                "container_id": 1,
+                "posture": "standing",
+                "action": {"verb": "watching"},
+            },
+        )
+        project(conn)
+        # Sanity: the "you" activity row exists and points at the container.
+        assert conn.execute(
+            "SELECT COUNT(*) FROM activity WHERE entity_id = 'you'"
+        ).fetchone()[0] == 1
+
+    response = client.post(
+        "/bots/bot_a/reset",
+        data={"confirm_name": "BotA"},
+        follow_redirects=False,
+    )
+    assert response.status_code == 303
+
+    with open_db(db) as conn:
+        # The orphaned "you" activity row tied to bot_a's purged container is gone.
+        assert conn.execute(
+            "SELECT COUNT(*) FROM activity WHERE entity_id = 'you'"
+        ).fetchone()[0] == 0
+
+
+def test_reset_does_not_purge_you_activity_in_other_chats(client, tmp_path):
+    """T69: resetting bot_a must leave a "you" activity row pointing at
+    bot_b's container intact — only orphans from the reset bot's chats go."""
+    db = tmp_path / "test.db"
+    with open_db(db) as conn:
+        # bot_a + its chat + container.
+        append_event(
+            conn,
+            kind="bot_authored",
+            payload={
+                "id": "bot_a",
+                "name": "BotA",
+                "persona": "thoughtful",
+                "voice_samples": [],
+                "traits": [],
+                "backstory": "",
+                "initial_relationship_to_you": "coworker",
+                "kickoff_prose": "",
+            },
+        )
+        append_event(
+            conn,
+            kind="chat_created",
+            payload={
+                "id": "chat_bot_a",
+                "host_bot_id": "bot_a",
+                "initial_time": "2026-04-26T20:00:00+00:00",
+                "narrative_anchor": "Day 1",
+                "weather": "",
+            },
+        )
+        append_event(
+            conn,
+            kind="container_created",
+            payload={
+                "chat_id": "chat_bot_a",
+                "name": "office",
+                "type": "workplace",
+                "properties": {},
+            },
+        )
+        # bot_b + its chat + container.
+        append_event(
+            conn,
+            kind="bot_authored",
+            payload={
+                "id": "bot_b",
+                "name": "BotB",
+                "persona": "curious",
+                "voice_samples": [],
+                "traits": [],
+                "backstory": "",
+                "initial_relationship_to_you": "friend",
+                "kickoff_prose": "",
+            },
+        )
+        append_event(
+            conn,
+            kind="chat_created",
+            payload={
+                "id": "chat_bot_b",
+                "host_bot_id": "bot_b",
+                "initial_time": "2026-04-26T20:00:00+00:00",
+                "narrative_anchor": "Day 1",
+                "weather": "",
+            },
+        )
+        append_event(
+            conn,
+            kind="container_created",
+            payload={
+                "chat_id": "chat_bot_b",
+                "name": "kitchen",
+                "type": "home",
+                "properties": {},
+            },
+        )
+        # The activity table is keyed on entity_id (PRIMARY KEY), so only one
+        # "you" row exists at a time. Point it at bot_b's container so reset of
+        # bot_a should NOT touch it.
+        append_event(
+            conn,
+            kind="activity_change",
+            payload={
+                "entity_id": "you",
+                "container_id": 2,  # kitchen, in chat_bot_b
+                "posture": "sitting",
+                "action": {"verb": "reading"},
+            },
+        )
+        project(conn)
+        # Sanity: the "you" activity row is in bot_b's container.
+        row = conn.execute(
+            "SELECT container_id FROM activity WHERE entity_id = 'you'"
+        ).fetchone()
+        assert row is not None and row[0] == 2
+
+    response = client.post(
+        "/bots/bot_a/reset",
+        data={"confirm_name": "BotA"},
+        follow_redirects=False,
+    )
+    assert response.status_code == 303
+
+    with open_db(db) as conn:
+        # The "you" activity in bot_b's container is preserved.
+        row = conn.execute(
+            "SELECT container_id FROM activity WHERE entity_id = 'you'"
+        ).fetchone()
+        assert row is not None
+        assert row[0] == 2
+
+
 def test_reset_purges_guest_memories_from_other_chats(client, tmp_path):
     db = tmp_path / "test.db"
     _seed_two_bots_with_guest_link(