claude suggested speed/memory improvements:

  #1 — read_hdf5_datasets: eliminate np.append in loop
  Replaced the per-file np.append pattern (which reallocates the full array on every call) with collecting chunks in a list and calling np.concatenate once at the end. For N files of M
   points each, this reduces allocations from O(N) full-array copies to one.

  #2 — fill_nan_with_neighbor_mean: vectorise triple-nested loop
  Replaced the Python for i / for j / for neighbour loop with scipy.ndimage.generic_filter, which operates in compiled C over the entire array in one pass.

  #3 — plot_2D_posterior: deduplicate sorted-histogram computation
  The same np.sort + np.cumsum + normalisation was computed twice in the same function call (once for CR masking, once for contour drawing). Now initialised to None and computed at
  most once, with the second use reusing the cached result.

  #4 — Replace deepcopy with shallow copies throughout
  - Numeric bounds (xy_bounds, x_bounds): replaced with [list(b[0]), list(b[1])] / list(b) — sufficient since only scalars are mutated.
  - requested_datasets (list of immutable tuples): replaced with list().
  - NumPy array slices (y_data[mask], z_data[mask], etc.): removed entirely — fancy indexing already returns a copy.
  - deepcopy(y_data) in posterior shading: replaced with y_data.copy().
  - Removed the now-unused deepcopy import.

  #5 — plot_1D_profile confidence band loop: single mask computation
  The boolean mask (x_values > x_start) & (x_values < x_end) was evaluated twice per segment (once for x, once for y) and converted through a Python list. Now computed once with
  np.concatenate instead of list wrapping.

  #6 — bin_and_profile_2D: nested loop → np.meshgrid
  Replaced the double Python loop over bin centres with a single np.meshgrid call + .ravel(), keeping the same flat-array layout (index = y_bin_index * n_xbins + x_bin_index).

  Bonus — Module-level np.finfo constants
  np.finfo(float).max and .eps (called ~20 times across function signatures and bodies) are now computed once at import time as _FLOAT_MAX and _FLOAT_EPS.

Author: ChrisJChang