[codex] Refactor benchmark execution orchestration

conftest.py

@@ -13,17 +13,17 @@ def response():
 @pytest.fixture(scope="session")
 def mock_openai():
     with patch("openai.ChatCompletion.create") as mock_create:
-        mock_create.return_value = {"choices": [{"message": {"content": response()}}]}
+        mock_create.return_value = {"choices": [{"message": {"content": mock_response}}]}
         yield mock_create
 
 
 @pytest.fixture(scope="session")
 def mock_together():
-    mock_response = MagicMock()
-    mock_response.choices = [MagicMock(message=MagicMock(content=mock_response))]
+    response_obj = MagicMock()
+    response_obj.choices = [MagicMock(message=MagicMock(content=mock_response))]
 
     mock_client = MagicMock()
-    mock_client.chat.completions.create.return_value = mock_response
+    mock_client.chat.completions.create.return_value = response_obj
 
     with patch("together.Together", return_value=mock_client):
         yield mock_client

scripts/README.md

@@ -2,6 +2,24 @@
 
 This directory contains utility scripts for running experiments and managing the project.
 
+## Benchmark runner
+
+`run_benchmark.py` plans and runs paper-style benchmark suites with explicit
+model overrides, repeat-aware output directories, manifest files, and safety
+checks for partial outputs.
+
+Preview the Qwen final-main plan without writing outputs:
+
+```bash
+uv run python scripts/run_benchmark.py --suite final-main --dry-run
+```
+
+Run with API preflight and resume support:
+
+```bash
+uv run python scripts/run_benchmark.py --suite final-main --preflight-api --resume
+```
+
 ## Directories
 
 ### slurm/
@@ -20,4 +38,4 @@ Data preprocessing and analysis scripts from various experiments.
 - `bentz2024_extract_text.py` - Extracts and combines paragraphs from CSV data
 - `bentz2024_get_sentence_suspense.py` - Calculates sentence-level suspense from word-level ratings
 
-Note: These scripts were originally part of the Bentz2024 experiment for processing German suspense data.
+Note: These scripts were originally part of the Bentz2024 experiment for processing German suspense data.