The SIR_v5 tests in test_power_edge_without_a_case.py fail on a loaded host: their ∂f/∂p derivation runs past the 20 s wall-clock budget #1416

The three SIR_v5 tests in python/tests/test_power_edge_without_a_case.py need SIR_v5's analytic sensitivity RHS. They derive it under the default BNGSIM_SENS_DERIV_BUDGET_S, a 20 s wall-clock deadline. On a loaded host the derivation can run past it, and then the tests fail on the budget rather than on what they check.

This happened once, in a full BNGSIM_SKIP_AUDIT=strict run under -n 6 on main 49c2931 + #1348, which doesn't touch these files:

FAILED python/tests/test_power_edge_without_a_case.py::test_sir_v5_with_a_singular_exponent_is_refused[dt_2023]
TypeError: argument of type 'NoneType' is not iterable

generate_sens_from_model returned None, as it does when the budget runs out. The test's _source helper is typed -> str and handed that None to not in. On reruns the test passed, alone and under -n 6, on that merge and on main.

Measured on macOS arm64, on a build whose _codegen.py, _jacobian.py and test file are main e5fcfa3's:

• The margin. Alone, generate_sens_from_model on SIR_v5 with a_2023 = 1.5 took 3.7 s in one run, so the budget leaves about 5x for load.
• The failure, on demand. With BNGSIM_SENS_DERIV_BUDGET_S=0.05:
  • both cases of test_sir_v5_with_a_singular_exponent_is_refused fail at line 1377 with that TypeError;
  • test_sir_v5_at_its_own_values_keeps_its_columns fails at line 1406 with SensitivityUnsupportedError ("it has no analytic sensitivity right-hand side (the build-time ∂f/∂p derivation exceeded its 0.05s budget …").
• Only this budget matters. With BNGSIM_JAC_DERIV_BUDGET_S=0.05 alone, all three pass.
• No silent wrong result. With the budget spent, the sensitivity run is refused, and the decline warning says why: CVODES' difference quotient would be wrong across this model's moving crossings.

This is the failure e8b6e05 fixed for SIR_v4. There, test_sir_v4_keeps_both_derivatives_now_its_model_is_fixed sets both budgets to inf, because what it checks is "whether the derivatives are kept, not how fast they are found".

Proposed fix: the same pin in the three tests, monkeypatch.setenv("BNGSIM_SENS_DERIV_BUDGET_S", "inf") before the first derivation. With it, all three passed locally even under an outer BNGSIM_SENS_DERIV_BUDGET_S=0.05. Optionally, _source could assert that the source isn't None, so a spent budget reads as one instead of as a TypeError.

test_zero_base_power_derivative.py::test_sir_v5_keeps_both_derivatives also derives SIR_v5 under the default budget and asserts has_analytic_sens_rhs. It runs only where the .net corpus is present, so it's exposed in the same way there. I haven't measured it, since it skipped here.

Not checked: other tests in the suite that derive a large model under the default budget.