V3 post step 2 refinement: add temperature & top-p settings to config (and .env), add prompt fields back to job table so that the prompt settings get frozen at runtime for all sources being processed.

tests/integration/test_pipeline_flow.py

@@ -73,10 +73,13 @@ class TestPipelineSuccessFlow:
             image_path,
             *,
             prompt_name="transcribe_document.md",
+            prompt_text=None,
+            temperature=None,
+            top_p=None,
             settings=None,
             provider=None,
         ) -> TranscriptionResult:
-            _ = (image_path, prompt_name, settings, provider)
+            _ = (image_path, prompt_name, prompt_text, temperature, top_p, settings, provider)
             return TranscriptionResult(
                 text="Pipeline transcript",
                 provider="openrouter",
@@ -102,10 +105,10 @@ class TestPipelineSuccessFlow:
         assert job is not None
         assert job.status == JobStatus.TRANSCRIBED
         assert any(job_source.raw_transcription == "Pipeline transcript" for job_source in job.job_sources)
-        assert any(job_source.prompt_name == "transcribe_document.md" for job_source in job.job_sources)
-        assert any(job_source.user_prompt is not None for job_source in job.job_sources)
-        assert any(job_source.temperature == 0.2 for job_source in job.job_sources)
-        assert any(job_source.top_p == 0.85 for job_source in job.job_sources)
+        assert job.prompt_name == "transcribe_document.md"
+        assert job.user_prompt is not None
+        assert job.temperature == 0.2
+        assert job.top_p == 0.85
         assert any(job_source.ai_metadata == {"finish_reason": "stop", "usage": {"total_tokens": 42}} for job_source in job.job_sources)
         assert any(
             job_source.raw_api_response == {"id": "resp_123", "choices": [{"message": {"content": "Pipeline transcript"}}]}
@@ -142,11 +145,14 @@ class TestPipelineSuccessFlow:
             image_path,
             *,
             prompt_name="transcribe_document.md",
+            prompt_text=None,
+            temperature=None,
+            top_p=None,
             settings=None,
             provider=None,
         ) -> TranscriptionResult:
             page_name = Path(image_path).name
-            _ = (prompt_name, settings, provider)
+            _ = (prompt_name, prompt_text, temperature, top_p, settings, provider)
             return TranscriptionResult(
                 text=f"Transcript for {page_name}",
                 provider="openrouter",
@@ -171,6 +177,7 @@ class TestPipelineSuccessFlow:
         assert all(job_source.status == JobSourceStatus.TRANSCRIBED for job_source in job.job_sources)
         assert all(job_source.raw_transcription for job_source in job.job_sources)
         assert all(job_source.source is not None and job_source.source.raw_transcription for job_source in job.job_sources)
+        assert job.prompt_name == "transcribe_document.md"
 
     @pytest.mark.asyncio
     async def test_worker_marks_partial_success_when_some_sources_fail(
@@ -202,12 +209,15 @@ class TestPipelineSuccessFlow:
             image_path,
             *,
             prompt_name="transcribe_document.md",
+            prompt_text=None,
+            temperature=None,
+            top_p=None,
             settings=None,
             provider=None,
         ) -> TranscriptionResult:
             nonlocal call_count
             call_count += 1
-            _ = (prompt_name, settings, provider)
+            _ = (image_path, prompt_name, prompt_text, temperature, top_p, settings, provider)
             if call_count == 2:
                 raise RuntimeError("simulated page failure")
             return TranscriptionResult(
@@ -279,11 +289,14 @@ class TestPipelineSuccessFlow:
             image_path,
             *,
             prompt_name="transcribe_document.md",
+            prompt_text=None,
+            temperature=None,
+            top_p=None,
             settings=None,
             provider=None,
         ) -> TranscriptionResult:
             nonlocal call_count
-            _ = (image_path, prompt_name, settings, provider)
+            _ = (image_path, prompt_name, prompt_text, temperature, top_p, settings, provider)
             call_count += 1
             return TranscriptionResult(
                 text="new transcript",
@@ -332,10 +345,13 @@ class TestPipelineFailureFlow:
             image_path,
             *,
             prompt_name="transcribe_document.md",
+            prompt_text=None,
+            temperature=None,
+            top_p=None,
             settings=None,
             provider=None,
         ) -> TranscriptionResult:
-            _ = (image_path, prompt_name, settings, provider)
+            _ = (image_path, prompt_name, prompt_text, temperature, top_p, settings, provider)
             raise RuntimeError("pipeline provider failure")
 
         monkeypatch.setattr(