Fix: remove artificial fallback scores in surface similarity functions

[Apprentice2907]

Fixes #28
Three similarity functions in surface_matcher.py were returning
artificially high fallback scores (0.85–0.95) when feature data
was missing, silently inflating overall match scores and causing
unrelated fragment pairs to pass the 0.6 similarity threshold.

Changes:

• compute_shape_similarity: return 0.0 instead of 0.95 when eigenvalues are missing
• compute_curvature_similarity: return 0.0 instead of max(intersection, 0.90) when histograms are empty
• compute_boundary_similarity: return 0.0 instead of 0.85 when boundary length or compactness is missing

These three functions account for 50% of the weighted overall score,
so the bad fallbacks were directly causing false matches in the
reconstruction pipeline.