fix(test): seed the flaky interpolator covariance tests

[Jammy2211]

Closes #1450.

Makes test_autofit/interpolator/test_covariance.py deterministic. This is the
Class B release blocker — the sole reason the 2026.8.2.1 live release failed
(PyAutoHands run 30736527569, release_test_pypi (3.12, PyAutoFit, main) step 9:
1641 passed, 1 failed, assert 30.121646313498022 == 25.0 ± 5).

Test-only. No library source or public API change. One file, +53/−2.

Measured pre-fix failure rate

Measured before changing anything, so seeding would not be applied over a
genuinely wrong tolerance. 2000 independent runs under the same conditions CI
uses (the autouse limit_maxcall fixture caps the search at maxcall=1):

tolerance	failures	rate	95% CI (Wilson)
abs=5.0 (the assertion)	33 / 2000	1.65%	1.18% – 2.31%
abs=4.0	109 / 2000	5.45%	4.54% – 6.53%
abs=3.0	293 / 2000	14.65%	13.17% – 16.27%

Recovered value: mean 25.24, std 2.05, range 17.82 – 32.04. The observed CI
failure (30.12) sits comfortably inside that distribution — a marginal miss, not
a regression.

The tolerance is fine; the estimator is broken

With a full untruncated search the same test recovers 25.0496 ± 0.0039
(n=20) — 500× tighter and nowhere near the abs=5.0 boundary. So abs=5.0 is
correctly calibrated and should not be loosened.

limit_maxcall (added with #1386 for speed) caps every search in this module at
a single likelihood call, so the recovered value is an unconverged draw whose
spread is set entirely by the search's randomness. Stage-wise check with all
seeds fixed: the interpolator's inputs (x, y, inverse covariance) are
bit-identical; all the variance is inside search.fit(...).

Seeding np.random alone would not have fixed it

Three independent generators feed the result:

1. numpy.random — the two np.random.random() calls in the test body.
   Replacing them with a fixed-seed local Generator leaves std at 1.95 (of
   2.05) and the test still fails. Not the dominant term.
2. the stdlib random module — autofit/non_linear/initializer.py:301 draws the
   search's initial unit values with random.uniform.
3. dynesty's rstate — NestedSampler defaults it to
   np.random.Generator(PCG64(None)), seeded from OS entropy and reachable from
   neither of the above (dynesty/dynesty.py:607-608, 814).
   This is the dominant term.

It also fixes a worse sibling

test_single_variable contains no np.random call at all, yet over 500 runs
produced 500 distinct values and missed its abs=2.0 tolerance
18/500 = 3.6% of the time (std 0.88, range 22.26 – 29.66). It depends on
source (3) alone, so seeding np.random would never have touched it.

Ruled out: the scipy.linalg.LinAlgError guards on test_interpolate /
test_relationships / test_interpolate_other_field /
test_linear_analysis_for_value (e29c69ef2) are not this nondeterminism in
disguise. The interpolator fixture in conftest.py uses no RNG; its covariance
matrix is bit-identical across builds and rank-deficient (rank 6 of 9, each
3×3 block rank 2 of 3, condition number 1.3e17). Those guards cover inverting a
singular matrix — deterministic input, platform-LAPACK dependent. Separate issue.

The change

1. import random and a module-level SEED = 20260802.
2. A new autouse seed_search_randomness(monkeypatch) fixture beside
   limit_maxcall, which patches dynesty.dynesty.get_random_generator (the
   binding both NestedSampler and DynamicNestedSampler resolve at call time)
   and seeds the two global generators, restoring their state in a finally
   after yield so the seeding cannot leak into whatever runs next in the same
   process.
3. test_variable_and_constant builds its samples from a local
   np.random.default_rng(SEED) instead of the global np.random.

limit_maxcall is unchanged; the module still runs in ~1s.

Testing

• Recovered values bit-identical across 8 pytest invocations and in both
  selection contexts (whole module vs -k filtered):
  test_single_variable → 24.577316739353 (dev 0.42 vs tol 2.0);
  test_variable_and_constant → 22.718814166116 (dev 2.28 vs tol 5.0).
• Full pytest test_autofit/: 1642 passed, 6 skipped, confirming the state
  save/restore leaks nothing.

Could not verify locally

test_autofit/non_linear/search/nest/test_nautilus.py::test__single_core_builds_no_pool
fails with ModuleNotFoundError: No module named 'nautilus' — the optional
sampler is absent from the sandbox. Pre-existing and unrelated: it reproduces
identically with this change stashed. CI has the optional deps, so it should pass
there.

Reviewer note — two judgement calls left to you

These assertions are now weaker than they look. Under maxcall=1 both are
frozen-RNG regression checks, not accuracy checks: a seeded draw 2.28 away from
the true 25.05 still passes at abs=5.0. Restoring real meaning costs suite time
— measured maxcall=200 → std 1.06 at 0.76 s/draw, maxcall=1000 → std 0.46 at
2.67 s/draw, full → std 0.004 at ~25 s/draw. I left limit_maxcall alone
deliberately; that trade is yours.

Patching a third-party module from a test is not where this belongs. It is
only necessary because AbstractDynesty.search_kwargs is a closed dict, so
rstate cannot be threaded through even though dynesty accepts it. The same
gap means users cannot reproduce a fit. Filed as a follow-up in PyAutoMind
(draft/feature/autofit/search_seed_reproducibility.md) rather than folded in
here, to keep the release-unblocking change test-only.

🤖 Generated with Claude Code

https://claude.ai/code/session_016vdGab4C5rWumTCXkWHPyh

---

Generated by Claude Code