Create synthetic data degradation tools

[KedoKudo]

Parent Epic

Part of #172 (2D Resonance Imaging)

Description

Create tools to artificially degrade clean synthetic data to test recovery methods.

Motivation

We have:

• Clean synthetic data (ground truth known)
• Real sparse data (ground truth unknown)

To validate recovery methods, we need to:

1. Degrade synthetic data in controlled ways
2. Apply recovery methods
3. Compare to known ground truth

Degradation Types

1. Poisson Noise

• Add realistic counting statistics
• Scale by total counts parameter

2. Sparsification

• Remove fraction of counts
• Simulate flux-limited scenarios

3. Detector Effects

• Dead pixels
• Non-uniform efficiency
• Dark current

4. Background Noise

• Flat background
• Time-dependent background

Proposed Interface

class DataDegrader:
    """Degrade clean synthetic data for testing."""
    
    def __init__(self, random_seed: int | None = None):
        self.rng = np.random.default_rng(random_seed)
    
    def add_poisson_noise(
        self,
        data: np.ndarray,
        total_counts: float,
    ) -> np.ndarray:
        """Add Poisson counting statistics."""
        ...
    
    def sparsify(
        self,
        data: np.ndarray,
        keep_fraction: float,
    ) -> np.ndarray:
        """Randomly remove counts to simulate sparse data."""
        ...
    
    def add_dead_pixels(
        self,
        data: np.ndarray,
        dead_fraction: float,
    ) -> tuple[np.ndarray, np.ndarray]:  # data, mask
        """Add dead pixel regions."""
        ...
    
    def degrade_to_level(
        self,
        data: np.ndarray,
        target_level: int,  # L1-L4
    ) -> np.ndarray:
        """Degrade to specified severity level."""
        ...

Tasks

• Implement DataDegrader class
• Implement each degradation type
• Calibrate degradation to match real data
• Create degradation presets (L1-L4)
• Add unit tests
• Create example degraded datasets

Acceptance Criteria

• Degraded data statistically similar to real data
• Reproducible with random seed
• Works with existing synthetic data