test: leak checks match whole words, not letters inside random key hashes

services/api/tests/memory_stack.py

@@ -244,6 +244,13 @@ def stored_text(self) -> str:
         return "\n".join(found)
 
 
+def holds_word(text_dump: str, word: str) -> bool:
+    """Whether `word` appears in `text_dump` as a whole word, ignoring case. A dump of every
+    stored value also holds random key hashes and ids, whose letters can spell a short word
+    ("bus" in "...GdGbusxe..."); a stored sentence or token still has a boundary around it."""
+    return re.search(rf"(?<![^\W_]){re.escape(word)}(?![^\W_])", text_dump, re.IGNORECASE) is not None
+
+
 @pytest.fixture
 def stack(tmp_path, monkeypatch):
     engine = create_engine(

services/api/tests/test_boundary_forget.py

@@ -9,7 +9,7 @@
 import json
 
 import pytest
-from memory_stack import OWNER, request_id
+from memory_stack import OWNER, holds_word, request_id
 from memory_stack import stack as stack
 
 SECRET = "My spare house key hides under the blue flowerpot by the gate"
@@ -214,7 +214,7 @@ def test_a_memory_from_chat_stays_forgotten_and_its_message_can_be_forgotten_too
     # words too (Pi keeps the transcript). Pi can still forget the message afterwards,
     # and then nothing of it is left anywhere.
     assert "gym" not in json.dumps(stack.library(object_type="evidence"))
-    assert "gym" not in stack.stored_text()
+    assert not holds_word(stack.stored_text(), "gym")
     assert stack.unsay("msg_bike")["state"] == "deleted"
     assert "gym" not in json.dumps(stack.library())
-    assert "gym" not in stack.stored_text() and "gym" not in stack.backup()
+    assert not holds_word(stack.stored_text(), "gym") and not holds_word(stack.backup(), "gym")

services/api/tests/test_boundary_scope.py

@@ -9,7 +9,7 @@
 import json
 
 import pytest
-from memory_stack import READ_KEY
+from memory_stack import READ_KEY, holds_word
 from memory_stack import stack as stack
 
 TEA = "I prefer green tea in the afternoon"
@@ -147,7 +147,7 @@ def test_a_read_key_cannot_switch_namespace(seeded):
     assert (
         stack.http.put("/runtime/conversation/msg_x", headers=read, json=ingest).status_code == 401
     )
-    assert "chess" not in stack.stored_text()
+    assert not holds_word(stack.stored_text(), "chess")
 
 
 def test_all_scope_stays_inside_the_keys_namespace(seeded):

services/api/tests/test_boundary_supersede.py

@@ -11,7 +11,7 @@
 from concurrent.futures import ThreadPoolExecutor
 
 from app.routes import memory as memory_routes
-from memory_stack import request_id
+from memory_stack import holds_word, request_id
 from memory_stack import stack as stack
 
 OLD = "I always study physics in the library after school"
@@ -90,7 +90,7 @@ def test_a_correction_is_what_every_read_returns_and_history_keeps_the_old_text(
     assert stale.status_code == 409
     assert set(returned_texts(stack, memory_id)) == {NEWER}
     assert stack.index.holds(memory_id, NEWER)
-    assert "bus" not in stack.stored_text()
+    assert not holds_word(stack.stored_text(), "bus")
 
 
 def test_concurrent_corrections_one_wins_and_the_rest_leave_no_trace(stack):
@@ -222,3 +222,9 @@ def resolve(winner):
         json.loads(responses[[r.status_code for r in responses].index(409)].text)["detail"]["code"]
         == "revision_conflict"
     )
+
+
+def test_the_leak_check_ignores_letters_inside_a_random_hash():
+    # The check above once failed on a stored key hash that happened to spell "bus".
+    assert not holds_word("zI7eP5gQdGbusxeJu9C9KN5iuaFtieMX4=", "bus")
+    assert holds_word('["bus"]', "bus") and holds_word("study physics on the BUS", "bus")

services/api/tests/test_refusal_and_pending_search.py

@@ -9,6 +9,7 @@
 
 import json
 
+from memory_stack import holds_word
 from memory_stack import stack as stack
 
 GREY = "The weather was grey on the walk home"  # below the default value threshold
@@ -31,7 +32,7 @@ def test_a_refused_write_keeps_only_the_refusal(stack):
     payload = json.loads(refusal["payload_json"])
     assert payload["reason"] == "low_value" and payload["text_length"] == len(GREY)
     assert set(payload) == {"agent_id", "reason", "text_length", "value_score"}
-    assert "grey" not in stack.stored_text().lower()  # nor anywhere else
+    assert not holds_word(stack.stored_text(), "grey")  # nor anywhere else
 
 
 def test_search_finds_a_memory_still_queued_for_the_index_by_its_words(stack):