[Montana]

Closes #1217 (with #1348, which fixes its #1212 group).

What was wrong

Running the whole suite with every model forced through the MIR JIT (BNGSIM_CODEGEN_JIT=mir BNGSIM_CODEGEN_THRESHOLD=1) is the check that found #1208 and #1212. It is only a signal if it passes. On main 9a26f5b with #1348 merged in, 53 tests fail under it. That matches #1217's accounting once #1212's 38 are gone:

group	tests	why it cannot pass under a forced JIT
cc's compiled artifact	32	the test is about the .so, its cache key, the redirected cache dir, or the output-sensitivity and sens-RHS notes written beside it. The JIT compiles in process and writes none of them.
cc by name	7	the test asserts codegen_backend == "cc", and the JIT reports "mir"
a forced threshold	13	the test needs a 2-species model to run interpreted (exprtk, analytical, no .so, the interpreter's RHS counters), and THRESHOLD=1 compiles it. They fail under forced cc too.
engine-dependent answers	2	below

The 32 artifact tests are in test_output_sens_support_note.py (22), test_sens_rhs_decline_note.py (5), test_hpc_contract.py (2), and one each in test_codegen_cache_key.py, test_ar_species_output_sensitivities.py and test_sens_numerically_zero_base.py. The 7 cc-by-name tests are in test_engine_choice_accessors.py (4), test_tfun_analytical_jacobian.py (2) and test_lu_fill_routing.py (1). The 13 threshold tests are in test_engine_choice_accessors.py (3), test_rhs_obs_func_gate.py (3), test_steady_state_polish_jacobian.py (3), test_jacobian_discontinuous_fallback.py, test_lazy_jacobian.py, test_model_codegen_sensitivity.py, and test_steady_state_scaled_criterion.py's MODEL1712050008 retry test. That last one pins the arm64 interpreter's route (solver_jacobian_retried); compiled, the march reaches the same root, within 1.6e-6, without a retry.

The change

python/tests/_codegen_backend.py. A new helper, imported the way _fd_sens.py and _ci_workflow.py are. It provides three skipifs and one xfail:

• under_jit_no_cc_artifact and under_jit_engine_name read the JIT selection through bngsim._codegen._codegen_jit_backend(), the library's own reading of BNGSIM_CODEGEN_JIT.
• threshold_forced applies when BNGSIM_CODEGEN_THRESHOLD is set below the default of 256.
• xfail_when_threshold_forced(issue, why) is a strict xfail under the same condition.

Each skip reason is declared in conftest.py's _DECLARED_SKIPS (tier ANYWHERE), so BNGSIM_SKIP_AUDIT=strict accepts them.

+    @threshold_forced
     def test_small_model_uses_exprtk(self):
 ...
+    @under_jit_engine_name
     def test_explicit_codegen_uses_cc(self):

The two engine-dependent tests are not skipped. A forced run shows each gives a different answer on the compiled path. Both findings are filed, as #1357 and #1358. Each test carries a strict xfail under a lowered threshold, naming its issue, so the mark has to go when the fix lands:

1. Steady-state refusal (steady_state(method='integration') on the compiled RHS and Jacobian refuses a state at residual 2e-39 and does not retry on difference quotients #1357). steady_state(method="integration") on the compiled RHS refuses glioblastoma_egfrviii_seed.net's steady state at residual 2.1e-39, with continued_run_distance = inf, because its run test is never retried on difference quotients. That affects test_a_march_whose_integrator_fails_returns_the_state_it_passed[integration]. Fixed in steady state: take a run test the closed form cannot take once more on differences (#1357) #1370.
2. Stale get_param (get_param on an assignment-rule parameter after a compiled run returns its load-time value, not the rule's #1358). After a compiled run, get_param on a parameter an assignment rule sets returns its load-time value: q reads 1.0, where an interpreted run reads the rule's end-of-run 0.0183. The trajectory is unaffected. This is test_a_write_after_a_run_does_not_move_it_to_the_rules_current_value, whose precondition reads the slot.

Verified

On this branch with #1348 merged in, macOS arm64 MIR build, --max-worker-restart=500, BNGSIM_SKIP_AUDIT=strict:

run	passed	skipped	xfailed	failed
BNGSIM_CODEGEN_JIT=mir BNGSIM_CODEGEN_THRESHOLD=1	16,476	185	4	0 (53 before)
BNGSIM_CODEGEN_THRESHOLD=1 (forced cc)	16,512	149	4	0 (15 before)
default	16,527	137	1	0, the same skips as main

The suite's existing xfail is the 1 in each count; the forced runs added three. That was measured when MODEL1712050008's retry test still carried an xfail. It is now a forced-threshold skip, so a forced run reports 3 xfailed and one more skip. Rerun after that change, test_steady_state_scaled_criterion.py under BNGSIM_CODEGEN_THRESHOLD=1 with the strict audit gives 222 passed, 3 skipped and 1 xfailed (glioblastoma, #1357), and 224 passed by default. The strict audit accepts every new skip.

Negative control: with the artifact reason removed from _DECLARED_SKIPS, the strict audit reports 1 undeclared skip reason(s).

Not in this PR

• The scheduled forced-MIR job Make the full suite a clean signal under BNGSIM_CODEGEN_JIT=mir: 93 forced-MIR failures hide new ones (cc-artifact tests, engine-name asserts, #1212, #1213) #1217 proposes. It is a workflow edit, for a maintainer to add once this and codegen: write C for what ExprTk means, or decline before compiling (#1212) #1348 are in. A sketch, on the MIR build mir.yml already makes:

  - name: Whole suite, every model through the JIT (issue #1217)
    env: {BNGSIM_CODEGEN_JIT: mir, BNGSIM_CODEGEN_THRESHOLD: "1", BNGSIM_SKIP_AUDIT: strict}
    run: python -m pytest python/tests -n 3 --max-worker-restart=500

• Fixes for the two engine-dependent findings: steady_state(method='integration') on the compiled RHS and Jacobian refuses a state at residual 2e-39 and does not retry on difference quotients #1357 and get_param on an assignment-rule parameter after a compiled run returns its load-time value, not the rule's #1358.