Separate noise estimation from sample weight computation

[jeipollack]

Description of Refactoring/Improvement
The SHE validation workflow requires access to the noise estimate associated with each observation to compute goodness-of-fit metrics for the WaveDiff model.

Currently, train_utils.calculate_sample_weights performs two distinct operations:

• estimate the noise level using the utils.NoiseEstimator class;
• convert this estimate into a scalar sample weight using a median-based reduction.

The current implementation does not expose the intermediate noise estimate independently, limiting its reuse for validation metrics such as reduced χ².

Goals and Objectives

• Extract the noise estimation step into a dedicated estimate_noise_sigma function.
• Update calculate_sample_weights to call estimate_noise_sigma internally.
• Preserve the existing weighting behaviour (median-based reduction).
• Maintain backwards compatibility with the current training pipeline.
• Expose the per-observation noise estimate for use in future uncertainty models and goodness-of-fit calculations.

Current Code Behaviour
calculate_sample_weights currently combines two responsibilities:

• Estimate the noise level for each observation using utils.NoiseEstimator, which uses a robust MAD-based estimator on unmasked background pixels.
• Convert this estimate into a scalar sample weight used during training.

The coupling of these operations prevents the noise estimate from being accessed independently by downstream validation workflows.

Proposed Changes

Introduce a dedicated helper:

estimate_noise_sigma(...)

which returns the per-observation noise standard deviation estimated by utils.NoiseEstimator.

Refactor:

calculate_sample_weights(...)

to use:

sigma = estimate_noise_sigma(...)
weights = ...

while preserving the current training behaviour.

Expected Benefits

• Separates data-derived noise estimation from training-specific weighting.
• Makes the uncertainty information available for future goodness-of-fit calculations.
• Improves clarity around the statistical assumptions used by the training pipeline.
• Avoids duplicating noise estimation logic in validation workflows.

Dependencies

• Compatibility with the SHE validation pipeline.
• Conversion from noise estimates to an inverse covariance representation remains dependent on the uncertainty model adopted by the validation workflow and is outside the scope of this refactor.