[DESIGN][MEDIUM] GW likelihood uses hardcoded 10% distance error instead of per-source Fisher bound

[JasperSeehofer]

File: master_thesis_code/bayesian_inference/bayesian_inference.py

The GW likelihood uses a uniform fractional error for every event:

sigma_luminosity_distance = FRACTIONAL_LUMINOSITY_ERROR * mu_luminosity_distance  # 10%

The simulation already computes the actual per-source Cramér–Rao lower bound on d_L and stores it as delta_luminosity_distance_delta_luminosity_distance in the CSV output (accessible via the Detection dataclass).

Using a hardcoded 10% ignores source-by-source variation in SNR and orbital geometry: bright nearby sources become artificially uncertain while faint distant sources appear artificially precise. This flattens the contribution of well-localised events to the H₀ posterior.

Fix: Pass d_L_uncertainty = sqrt(detection.delta_luminosity_distance_delta_luminosity_distance) from the Detection dataclass into BayesianInference and use it in place of FRACTIONAL_LUMINOSITY_ERROR * d_L in both gw_likelihood() and gw_detection_probability().