feat(review): 取证层多源化 + 视觉多图（整体+2x2 象限）

- 新增 evidence_model.py：统一证据模型（source_type + images 列表），访问器兼容旧 image_b64/image_mime 单图字段
- phases/evidence.py：新增 collect() 多源分发 seam（仅 CAD）、capture_images() 整体+等分象限多图（结束复位视口）、sub_windows()
- phases/review.py：多图逐张下采样喂 LLM，payload 增加 source_type
- config.py：EVIDENCE_DETAIL_TILES=4 / EVIDENCE_MAX_IMAGES=5（默认开启多图）
- 新增 tests/test_evidence_model.py，更新 evidence/review 断言与 fakes 归一

tests/test_review.py

@@ -28,8 +28,8 @@ from tests.fakes import (
 def test_evidence_payload_strips_image():
     ev = make_evidence(image_b64="QUJD", image_mime="image/png")
     payload = evidence_payload(ev)
+    assert "images" not in payload
     assert "image_b64" not in payload
-    assert "image_mime" not in payload
     assert payload["frame_index"] == 1
 
 
@@ -243,3 +243,20 @@ def test_review_frame_intermediate_tool_turn_json_not_parsed():
     out = review_frame(ctx, make_evidence(), "sys", log=lambda *a: None)
     assert out["summary"] == "最终结论"
     assert out["issues"] == []
+
+
+def test_review_frame_includes_multiple_image_blocks():
+    llm = FakeLlmClient(script=[llm_text("ok")])
+    ctx = make_ctx(client=FakeMcpClient(), llm=llm)
+    ev = make_evidence(
+        images=[
+            {"b64": "A", "mime": "image/png", "label": "overall"},
+            {"b64": "B", "mime": "image/png", "label": "tile-1"},
+        ]
+    )
+    review_frame(ctx, ev, "sys", log=lambda *a: None)
+    user_content = llm.calls[0]["messages"][1]["content"]
+    image_blocks = [b for b in user_content if b.get("type") == "image_url"]
+    assert len(image_blocks) == 2
+    assert any(b.get("text") == "（截图：overall）" for b in user_content)
+    assert any(b.get("text") == "（截图：tile-1）" for b in user_content)