Bug: _parse_code in ProgramOfThought corrupts valid Python code

Description

I've am using CodeAct and encountered issues where the _parse_code method corrupts valid Python code generated by the LLM. There are two separate problems in the parsing logic that break otherwise correct code.

Issue 1: .replace("\\n", "\n") corrupts escape sequences in string literals

Location: dspy/predict/program_of_thought.py, line 139

The .replace("\\n", "\n") call replaces all occurrences of \n with actual newline characters, including those inside Python string literals where they should remain as escape sequences.

Example:
When the LLM generates:

print(f"\nTotal Users: {total_users}")

After _parse_code, it becomes:

print(f"
Total Users: {total_users}")

This causes: Invalid Python syntax

Issue 2: Regex substitutions corrupt code with = in strings or complex expressions

Location: dspy/predict/program_of_thought.py, lines 144-158

The regex patterns attempt to fix single-line multi-assignment code but don't respect string boundaries or Python syntax:

code_block = re.sub(
    r"([a-zA-Z_]\w* *=.*?)(?=[a-zA-Z_]\w* *=)",
    r"\1\n",
    code_block,
)

Example:
When the LLM generates:

data = "users: Alice=25, Bob=30, Carol=28"
youngest = [name for name, age in users.items() if age == min_age]

After the regex substitutions:

data = "users: 
Alice=25, 
Bob=30, 
Carol=28"
youngest = [name for name, age in users.items() if 
age == min_age]

The regex inserts newlines inside string literals and breaks list comprehensions, causing syntax errors.

Steps to reproduce

import dspy
from dspy.predict import CodeAct

lm = dspy.configure(lm=dspy.LM("anthropic/claude-sonnet-4-5"))

class DataSummarySignature(dspy.Signature):
    """Analyze data and print a formatted summary report."""
    data: str = dspy.InputField(desc="Data to analyze")
    summary: str = dspy.OutputField(desc="Summary report")

act = CodeAct(signature=DataSummarySignature, tools=[], max_iters=3)
result = act(data="users: Alice=25, Bob=30, Carol=28")

Suggested Fix

Simplify _parse_code to only extract code from markdown blocks without attempting to reformat it:

def _parse_code(self, code_data):
    code = code_data.get("generated_code", "").split("---", 1)[0].split("\n\n\n", 1)[0]
    code_match = re.search(r"```python[ \n](.*?)[ \n]```?", code, re.DOTALL)
    code_block = code_match.group(1) if code_match else code
    if not code_block:
        return code, "Error: Empty code after parsing."
    return code_block, None

I'm happy to submit a PR with these changes if this approach looks reasonable. Thank you for the excellent work on DSPy!

DSPy version

3.1.0b1