Share the strength GP fit across tests: 399s -> 135s

[SebastianAment]

Speeds up the Python test suite from ~399 s to ~135 s (2.9×) with no change to what is tested.

The problem

The runtime wasn't spread across the 270 tests. In a profiling run totalling 476 s, the five slowest entries accounted for 456 s, and four of those were fits of the same production strength GP:

time	test
108.6 s	test_models.py::TestPredictiveQualityRegression (setup)
108.3 s	test_models.py::TestGetModelListWithCost (setup)
106.9 s	test_lengthscale_identifiability
102.9 s	test_strength_curve_monotonicity

(Those durations come from that profiling run, which was on a busier machine than the before/after figures below — they are not directly comparable.)

They really are one fit: fit_strength_model calls fit_strength_gp(X, Y, Yvar, X_bounds) on data.strength_data — exactly what the other two call directly — fit_strength_gp defaults to seed=0 and applies it via torch.manual_seed, and DATA_PATH is load_concrete_strength's default. Same inputs, same seed, same function. Verified bit-exactly: old and new paths produce state_dict diffs of 0.0.

The fix

test/shared_fits.py fits once per process and hands out deepcopys. The fitted objects live in a closure with no module-level name bound to them, so no caller reaches one by accident.

Measured rather than assumed:

• mutating a copy's parameters leaves the original untouched
• zero shared tensor identities and zero shared storage between copy and original
• a deepcopy costs ~0.00 s against a ~100 s fit

The provider drives the public fit_strength_model rather than reaching for fit_strength_gp, so the production path under test is unchanged.

Results

before   405 s, 393 s
after    141 s, 130 s      (same machine, coverage on)

270 tests pass, coverage 100%, CI green on 3.11 and 3.12. Verified from a clean worktree of the commit itself, not just my working tree.

Why there is no parallelism here

pytest-xdist was prototyped and dropped. It helped before the fits were shared (435 → 200 s, on a busier machine), but afterwards stopped paying for itself: 135 s serial against ~184 s at -n 4, since workers are separate processes that each re-import torch and, without grouping, each refit. With the default --dist load the shared fit scatters and the suite measured 265 s — slower than not parallelising at all. The numbers are in the Makefile so this isn't rediscovered.

Risk

Test-only; no boxcrete/ changes. The thing worth a reviewer's eye is test isolation — a shared mutable fixture can create order-dependent failures. Ordering was varied deliberately (reverse module order, and under xdist with several worker counts and distribution modes) and all 270 passed in every arrangement.