[claude]

Adds tests/test_norta.py, which tests rupture_generator/sampling/norta.py. Only test files are changed.

What the new tests check

• The normal marginal is the identity. Uses Hypothesis. NORMAL.apply returns its input unchanged as float64, and integer input comes back as floats. latent_correlation(NORMAL, NORMAL, t) returns t unchanged.
• Pairing a normal with a non-normal is a linear correction. The normal transform is He_1 alone, so g(rho) = rho * corr(Z, h(Z)). The test computes corr(Z, h(Z)) for a gamma(0.75) marginal with its own scipy.integrate.quad and checks latent_correlation against target / corr at several points. It also checks that swapping the two marginals gives the same answer.
• Unattainable correlations are refused. A target just past g(1) or g(-1) raises RuptureGeneratorError; both the ceiling and the floor side are tested. A target past the edge by less than NORTA_CORRELATION_SLACK is clipped to latent correlation ±1.
• An unimplemented family raises ValueError the first time it is transformed. This is the case unknown branch.
• The doubly truncated normal fit is checked against independent numbers. For several parent scales and maxima:
  • the parent location found by _doubly_truncated_normal_location gives a cut normal with mean 1, computed by direct quadrature of the normal pdf on [0, maximum];
  • _doubly_truncated_normal_spread matches the quadrature coefficient of variation;
  • a narrow parent is left essentially uncut (location 1, spread equal to the scale).
• The fitted doubly truncated normal behaves correctly through apply. Its output is monotone and stays within [0, maximum], and the transformed standard normal has unit mean and the requested spread (Gauss-Hermite).

Coverage of norta.py

	Coverage	Lines missed
Before	95%	187, 291, 327, 328, 451, 463, 472
After	99%	291, 463

Both remaining lines are left uncovered on purpose:

• 291 (_distribution for normal) can't be reached through the public API, because Marginal.apply returns early for the normal marginal.
• 463 (correlation map not increasing) can't happen: by Price's theorem g never decreases for monotone transforms.

Surviving mutants

_doubly_truncated_normal mutants 1–12 (dropping or garbling a, b, loc or scale) are killed by the new tests. I checked each variant by patching it in through a scratch pytest plugin; every one fails 6–8 of the new doubly truncated normal tests. The existing tests in test_numerics.py (mean 1, support (0, maximum)) look like they should also have caught these. Mutmut's run reported 107 of 221 mutants as timeouts, so its test mapping for this module looks unreliable.

_doubly_truncated_normal_location mutants:

• Killed:
  • 4–10, 12 pass None or drop an argument. They crash inside SciPy or the root finder.
  • 11 (+ 1.0) never brackets the root, so the search never ends.
  • 13 (- 2.0) moves the cut mean to 2, which the quadrature mean check rejects.
• Killed by any test that reaches the code: 2 and 3 (invalid="XXignoreXX" / "IGNORE") raise ValueError in np.errstate.
• Equivalent: 1 (invalid=None) only stops a warning being silenced. With both errstate guards removed, the fits for maxima 1.5, 2, 4.3 and 10 raised no floating-point warnings, so it changes nothing that can be observed.

Issues filed

None. An unknown family string gets past Marginal.__post_init__ and only fails later, with a ValueError (whose message has a typo, "Unimplementhed"). It can only arrive by going against the MarginalFamily type hint, so I didn't file it as a bug.

🤖 Generated with Claude Code

[github-actions]

Benchmarks

7fa9b6e against base 1a72c3e, on INTEL(R) XEON(R) PLATINUM 8573C. Times are medians over the rounds. The median change is 1.3%; changes past 5% are in bold.

Kernels

kernel	base	head	change	head IQR	rounds	padded cells
embed[Mw6.5]	42.4 ms	42.9 ms	+1.3%	210 µs	23	448,000 (+0.0%)
circulant_draw[Mw6.5]	4.31 ms	4.38 ms	+1.6%	233 µs	174
embed[Mw7.5]	134 ms	136 ms	+1.6%	968 µs	8	1,204,224 (+0.0%)
circulant_draw[Mw7.5]	9.88 ms	10.1 ms	+1.9%	391 µs	90
eikonal_solve	1.27 s	1.29 s	+1.4%	944 µs	5
synthesise_pulses[impulse]	825 µs	831 µs	+0.8%	26.5 µs	607
synthesise_pulses[liu]	473 ms	474 ms	+0.2%	2.03 ms	5

End to end

Each round is the command line in a fresh interpreter, start-up included. Peak RSS is the largest the child process reached over the rounds.

example	base time	head time	change	base peak RSS	head peak RSS	change
alpine_hope	9.18 s	9.06 s	-1.3%	294 MiB	294 MiB	-0.1%
beavan	3.46 s	3.48 s	+0.4%	232 MiB	232 MiB	+0.2%
colombia	8.75 s	8.56 s	-2.1%	3,168 MiB	3,168 MiB	-0.0%
crustal	1.44 s	1.45 s	+1.3%	244 MiB	245 MiB	+0.5%
landers	2.75 s	2.81 s	+1.9%	385 MiB	385 MiB	+0.0%
northridge	1.41 s	1.42 s	+0.9%	294 MiB	293 MiB	-0.2%
two_faults	2.39 s	2.43 s	+1.6%	354 MiB	353 MiB	-0.2%
two_faults_segments	1.46 s	1.47 s	+1.1%	190 MiB	190 MiB	+0.2%
venezuela	16.5 s	16.8 s	+1.9%	5,768 MiB	5,766 MiB	-0.0%