fix(inference): #87 redact image bytes from inference_log

Log full text + image_ref (photo_id/seq/bytes) only; wire path still
sends data-URLs. Call 1/2 SELECT id+seq for refs. Cap is safety net
after redaction (METADATA-66).

backend/tests/test_inference.py

@@ -274,6 +274,112 @@ def test_fit_long_edge_png_leaves_small_unchanged_dims():
     assert im.size == (400, 300)
 
 
+def test_redact_request_for_log_strips_data_urls_keeps_text():
+    """#87: log sanitizer replaces data-URL images with photo refs; keeps text."""
+    from backend.inference import _redact_request_for_log
+
+    b64 = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mP8z8BQDwAEhQGAhKmMIQAAAABJRU5ErkJggg=="
+    messages = [{
+        "role": "user",
+        "content": [
+            {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}},
+            {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64 * 20}"}},
+            {"type": "text", "text": "LISTS:\n{\"species\":[]}\n\nSKU=X length_in=1 thickness_in=2"},
+        ],
+    }]
+    meta = [
+        {"photo_id": "p1", "seq": 1, "bytes": 100, "infer_long_edge_px": 64},
+        {"photo_id": "p2", "seq": 2, "bytes": 200, "infer_long_edge_px": 128},
+    ]
+    out = _redact_request_for_log(messages, meta)
+    blob = json.dumps(out)
+    assert "data:image" not in blob
+    assert "base64" not in blob.lower()
+    assert "LISTS:" in blob
+    assert "SKU=X" in blob
+    parts = out[0]["content"]
+    assert parts[0]["type"] == "image_ref"
+    assert parts[0]["photo_id"] == "p1"
+    assert parts[0]["seq"] == 1
+    assert parts[1]["photo_id"] == "p2"
+    assert parts[2]["type"] == "text"
+    assert "LISTS:" in parts[2]["text"]
+
+
+def test_call1_log_has_text_and_image_refs_no_base64(env, monkeypatch):
+    """#87 METADATA-66: inference_log stores text+refs, not image bytes."""
+    client, conn = env
+    conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('inference_enabled','true',NULL)")
+    conn.execute(
+        "INSERT INTO slabs(id,sku,length_in,thickness_in,status,server_rev,client_rev,created_at,updated_at) "
+        "VALUES ('s1','SLAB-01',100,2,'calibrated',1,1,'t','t')"
+    )
+    png = _rgba_png(400, 300)
+    conn.execute(
+        "INSERT INTO slab_photos(id,slab_id,seq,processed_blob,created_at) VALUES ('photo-abc','s1',1,?, 't')",
+        (png,),
+    )
+    conn.commit()
+    from backend.inference import InferenceClient
+
+    def fake_chat(self, messages):
+        # Wire path still has data-URL for the model
+        content = messages[0]["content"]
+        assert any(
+            isinstance(p, dict) and p.get("type") == "image_url"
+            and str(p.get("image_url", {}).get("url", "")).startswith("data:image")
+            for p in content
+        )
+        return json.dumps({"species_id": 1, "species_confidence": 0.9})
+
+    monkeypatch.setattr(InferenceClient, "chat", fake_chat)
+    r = client.post("/api/v1/slabs/s1/inference/call1")
+    assert r.status_code == 200, r.text
+    row = conn.execute(
+        "SELECT request_payload, length(request_payload) AS n FROM inference_log ORDER BY id DESC LIMIT 1"
+    ).fetchone()
+    payload = row["request_payload"]
+    assert row["n"] < 50000
+    assert "data:image" not in payload
+    assert "LISTS:" in payload
+    assert "image_ref" in payload
+    assert "photo-abc" in payload
+    assert '"seq": 1' in payload or '"seq":1' in payload
+    logged = json.loads(payload)
+    refs = [p for p in logged[0]["content"] if p.get("type") == "image_ref"]
+    assert len(refs) == 1
+    assert refs[0]["photo_id"] == "photo-abc"
+    assert refs[0]["bytes"] == len(png)
+
+
+def test_call2_log_redacts_images(env, monkeypatch):
+    """#87: Call 2 log also redacts image data-URLs."""
+    client, conn = env
+    conn.execute("INSERT OR REPLACE INTO settings(key,value,value_enc) VALUES ('content_llm_enabled','true',NULL)")
+    conn.execute(
+        "INSERT INTO slabs(id,sku,length_in,thickness_in,width_avg_in,bdft,title,status,server_rev,client_rev,created_at,updated_at) "
+        "VALUES ('s1','SLAB-01',100,2,20,18,'Walnut','ready',1,1,'t','t')"
+    )
+    png = _rgba_png(200, 200)
+    conn.execute(
+        "INSERT INTO slab_photos(id,slab_id,seq,processed_blob,created_at) VALUES ('p2','s1',1,?, 't')",
+        (png,),
+    )
+    conn.commit()
+    from backend.inference import InferenceClient
+
+    monkeypatch.setattr(InferenceClient, "chat", lambda self, messages: "<p>ok</p>")
+    r = client.post("/api/v1/slabs/s1/inference/call2")
+    assert r.status_code == 200, r.text
+    payload = conn.execute(
+        "SELECT request_payload FROM inference_log WHERE call=2 ORDER BY id DESC LIMIT 1"
+    ).fetchone()["request_payload"]
+    assert "data:image" not in payload
+    assert "image_ref" in payload
+    assert "p2" in payload
+    assert "DIMENSION FACT BLOCK" in payload
+
+
 def test_call1_sends_long_edge_at_most_1024(env, monkeypatch):
     """METADATA-35 end-to-end: messages to the model use <=1024 long edge."""
     client, conn = env