fix: Phase 4 - update model_evaluation_utils.py for JSON parsing

- Fixed DYNAMIC_PRIMARY_TEXT: now unpacks (prompt, parser) tuple and uses parser
- Fixed REIMBURSEMENT_PRIMARY: corrected function signature (only takes context)
- Replaced extract_text_from_delimiters with JSON parser
- Replaced universal_json_load with parser from tuple
- Added proper instruction caching for both prompts

Other testbed/QC files verified: they process internal data formats (out of scope per PRD)

src/testbed/model_evaluation_utils.py

@@ -21,15 +21,17 @@ def prompt_dynamic_primary(
         for exhibit_page, exhibit_text in exhibit_dict.items():
             answer_dict = {}
             for field in dynamic_primary_fields.fields:
-                prompt = prompt_templates.DYNAMIC_PRIMARY_TEXT(
+                prompt, parser = prompt_templates.DYNAMIC_PRIMARY_TEXT(
                     exhibit_text, field.field_name, field.get_prompt(constants)
                 )
                 llm_answer_raw = llm_utils.invoke_claude(
-                    prompt=prompt, model_id=model_id, filename=filename
+                    prompt=prompt,
-                )
+                    model_id=model_id,
-                llm_answer_final = string_utils.extract_text_from_delimiters(
+                    filename=filename,
-                    llm_answer_raw, string_utils.Delimiter.PIPE
+                    cache=True,
+                    instruction=prompt_templates.DYNAMIC_PRIMARY_TEXT_INSTRUCTION(),
                 )
+                llm_answer_final = parser(llm_answer_raw)
                 answer_dict[field.field_name] = llm_answer_final
             answer_dicts.append(answer_dict)
 
@@ -171,17 +173,14 @@ def evaluate_reimbursement_primary(
     for model_name, model_id in model_id_dict.items():
         seen_pairs = set()
         for exhibit_page, exhibit_text in exhibit_dict.items():
-            field_prompts = reimbursement_primary_fields.print_prompt_dict(constants)
+            prompt, parser = prompt_templates.REIMBURSEMENT_PRIMARY(exhibit_text)
-            instruction, prompt = prompt_templates.REIMBURSEMENT_PRIMARY(
-                exhibit_text, field_prompts
-            )
             llm_answer_raw = llm_utils.invoke_claude(
                 prompt,
                 model_id,
                 filename,
                 max_tokens=25000,
                 cache=True,
-                instruction=instruction,
+                instruction=prompt_templates.REIMBURSEMENT_PRIMARY_INSTRUCTION(),
                 usage_label="REIMBURSEMENT_PRIMARY",
             )
             # Check for the special "no results" case
@@ -189,7 +188,7 @@ def evaluate_reimbursement_primary(
                 continue
 
             try:
-                llm_answer_final = string_utils.universal_json_load(llm_answer_raw)
+                llm_answer_final = parser(llm_answer_raw)
 
                 reimbursement_primary_answers = (
                     one_to_n_funcs.clean_reimbursement_primary(