fix(projects): close Codex Wave 1.5 P2/P3 — stronger negatives + lowercase tokens

Codex's audit of e8ac8bb returned GO with two cheap improvements worth
folding in:

P2: the "short token does not match" test was trivially true because
apm and drill share no tokens at all. Replaced with a 4-label setup
that exercises both directions: apm + apm-fpga must NOT cluster (only
shared token is the 3-char 'apm'); foo-fpga + bar-fpga + apm-fpga MUST
cluster via the 4-char 'fpga'. Now a regression that lets <4-char
tokens through would fail.

P3: token comparison was case-sensitive. Lowercased before the length
check so 'HydroTech-Mining' clusters with 'hydrotech-split-tank' the
same way the all-lowercase variants do. Added a regression test.

Also added the registered-token-leak test Codex specifically called
out: p04-gigabit registered, gigabit-other unregistered — gigabit-other
must NOT surface p04-gigabit as a suggested alias (filter happens
before clustering).

Test count: 594 -> 596.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tests/test_emerging_project_proposals.py

@@ -103,20 +103,65 @@ def test_proposals_suggest_sibling_aliases_via_shared_tokens(env):
 
 
 def test_proposals_short_token_does_not_match(env):
-    """Two-or-three-letter tokens are too noisy to suggest aliases on.
-    'apm' (3 chars) and 'apm-fpga' (3 + 4) share 'apm' (3 chars) — the
-    'apm' token alone is too short, but 'fpga' (4) is long enough so
-    the match comes from the longer token. We test the negative: 'apm'
-    and 'drill' must NOT be siblings."""
-    for i in range(11):
-        create_memory("knowledge", f"apm fact {i}", project="apm")
-    for i in range(11):
-        create_memory("knowledge", f"drill fact {i}", project="drill")
+    """Per Codex Wave 1.5 P2: previously this test only asserted apm
+    and drill have empty siblings, which is trivially true because they
+    share no tokens at all. The real risk is an accidental relaxation
+    that lets <4-char tokens trigger clustering. Construct a setup where
+    that would matter:
+      - 'apm' and 'apm-fpga': only the 3-char 'apm' is shared. They must
+        NOT cluster, because 'apm' is too short.
+      - 'foo-fpga' and 'bar-fpga': the 4-char 'fpga' is shared. They
+        MUST cluster.
+    """
+    for label in ("apm", "apm-fpga", "foo-fpga", "bar-fpga"):
+        for i in range(11):
+            create_memory("knowledge", f"{label} fact {i}", project=label)
 
     client = TestClient(app)
     proposals = {p["project_id"]: p for p in client.get("/admin/projects/proposals").json()["proposals"]}
-    assert proposals["apm"]["suggested_aliases"] == []
-    assert proposals["drill"]["suggested_aliases"] == []
+
+    # Negative: short-token match must not happen
+    assert "apm-fpga" not in proposals["apm"]["suggested_aliases"], (
+        "'apm' (3 chars) is below the 4-char minimum; 'apm' and 'apm-fpga' "
+        "must not cluster via the 'apm' token."
+    )
+
+    # Positive: long-token match must happen — both directions
+    assert "bar-fpga" in proposals["foo-fpga"]["suggested_aliases"]
+    assert "foo-fpga" in proposals["bar-fpga"]["suggested_aliases"]
+    # And 'apm-fpga' clusters with the others via 'fpga'
+    assert "apm-fpga" in proposals["foo-fpga"]["suggested_aliases"]
+
+
+def test_proposals_clustering_is_case_insensitive(env):
+    """Token comparison must be case-insensitive so labels captured
+    with mixed casing still cluster. Codex Wave 1.5 P3."""
+    for label in ("HydroTech-Mining", "hydrotech-split-tank"):
+        for i in range(11):
+            create_memory("knowledge", f"{label} fact {i}", project=label)
+
+    client = TestClient(app)
+    proposals = {p["project_id"]: p for p in client.get("/admin/projects/proposals").json()["proposals"]}
+    assert "hydrotech-split-tank" in proposals["HydroTech-Mining"]["suggested_aliases"]
+    assert "HydroTech-Mining" in proposals["hydrotech-split-tank"]["suggested_aliases"]
+
+
+def test_proposals_registered_token_does_not_leak_into_sibling_set(env, monkeypatch):
+    """Registered project ids must be filtered BEFORE clustering so a
+    registered token doesn't get suggested as an alias for an
+    unregistered sibling. p04-gigabit is registered in env; an
+    unregistered 'gigabit-other' must not list 'p04-gigabit' as alias."""
+    for i in range(15):
+        create_memory("knowledge", f"p04 fact {i}", project="p04-gigabit")
+    for i in range(11):
+        create_memory("knowledge", f"gigabit-other fact {i}", project="gigabit-other")
+
+    client = TestClient(app)
+    proposals = {p["project_id"]: p for p in client.get("/admin/projects/proposals").json()["proposals"]}
+    assert "p04-gigabit" not in proposals
+    assert "gigabit-other" in proposals
+    # And the registered name must not surface as a sibling
+    assert "p04-gigabit" not in proposals["gigabit-other"]["suggested_aliases"]
 
 
 def test_proposals_include_sample_memories_and_guessed_root(env):