tests: define aloha TMP temporaries in an explicit eval namespace (fix 9 errors on python 3.13+)

[oliviermattelaer]

Symptom

Nine tests in tests/parallel_tests/test_aloha.py error out locally with NameError: name 'TMP0' is not defined (or TMP2, ...):

./tests/test_manager.py -p P -l WARNING 'test_aloha.*'
Ran 124 tests ... FAILED (errors=9)

• testLorentzObject: test_short_expand_veto, test_short_part_spin32propagator, test_short_spin2propagator4
• test_aloha_creation: test_short_aloha_FFT2, test_short_aloha_FFV, test_short_aloha_FFVP1N, test_short_aloha_FFV_MG4, test_short_aloha_ZPZZ, test_short_use_of_library_spin2

Root cause

Not an aloha bug. Aloha's stringification is correct, and the writers already materialise the temporaries properly (AbstractRoutine.contracted / .fct, built in create_aloha.define_simple_output from KERNEL.reduced_expr2).

The tests stringify an aloha expression and eval() the result. Those strings refer to the TMP.../FCT... temporaries aloha introduces for common subexpressions (aloha_lib.Computation.add_expression_contraction), so each test first defines them with a bare

for name, cexpr in aloha_lib.KERNEL.reduced_expr2.items():
    exec('%s = %s' % (name, cexpr))

That only ever worked by accident. Inside a function, exec() with no explicit namespace writes into the dict returned by locals(). Before PEP 667, CPython cached that dict on the frame and refreshed it from the fast locals without deleting extra keys — so the TMP0 written by exec() survived, and the later eval(), which reads the same cached dict, still saw it.

PEP 667 (python 3.13) makes locals() return an independent snapshot of the function scope. The exec() write goes into a throwaway dict, the next eval() gets a fresh snapshot without TMP0, and it raises NameError.

Minimal reproducer, independent of this repo:

def f():
    exec('TMP0 = 3 + 4')
    return eval('TMP0')

python3.9 gives 7. python3.14 gives NameError: name 'TMP0' is not defined.

So these tests were always fragile — they leaned on a CPython implementation detail — and only now trip. It is worth a real fix rather than writing it off as an environment artefact, because every interpreter from 3.13 onwards hits it.

Why CI is green

.github/workflows/aloha.yml has no setup-python step, so ./tests/test_manager.py (shebang #!/usr/bin/env python3) runs on the ubuntu-24.04 runner's default python3, which is 3.12 — one release before PEP 667. The test selection is identical to the local one; only the interpreter differs.

The fix

Build the namespace explicitly and pass it to both exec() and eval(), which behaves the same on every version. A small module-level helper seeds it from the module globals plus the test's own locals:

def eval_scope(local_vars):
    scope = dict(globals())
    scope.update(local_vars)
    return scope

and each affected site becomes scope = eval_scope(locals()), then exec(..., scope) and eval(..., scope). This is the idiom the generated write_param_card.py files in this repo already use (exec(..., {"cmath": cmath}, local_vars)).

Only the nine failing tests are touched, and the surrounding try/except structure of each site is preserved exactly, so the pre-3.13 behaviour is reproduced rather than changed.

No assertion was loosened. The tests still compare against the same hardcoded MG4 reference values and the same cross-checks between routines.

Verification

./tests/test_manager.py -p P -l WARNING 'test_aloha.*' (124 tests), run on both interpreters available locally:

	python 3.9.6 (pre-PEP-667, proxy for CI's 3.12)	python 3.14.6 (PEP 667)
before	OK	FAILED (errors=9)
after	OK	OK

That 2x2 is the evidence for the diagnosis: the interpreter is the only variable, and the fix is version independent.

Liveness checks, so the green is not vacuous:

• The temporaries really are generated and really are consumed — test_short_aloha_ZPZZ builds TMP0..TMP7, test_short_aloha_FFV_MG4 builds TMP0/TMP1.
• Negative control: perturbing the test_short_aloha_ZPZZ reference by one unit makes it fail with AssertionError: 178727040j != 178727041j — i.e. the eval() is genuinely producing the right value through the temporaries.

Full unit suite (-p U): 892 tests, only the two known pre-existing failures (testIO_UnitProcOutputIOTests, order-dependent; test_DensityMatrixObservables22, missing scipy). No regression.

Note on a nearby, separate issue

.github/workflows/aloha.yml:51 is missing a space before -e, so the test_short_mssm_subset_creation exclusion is inert. That is not touched here — a fix already exists on claude/fix-dead-ci-test-names (63466ec45), which has not yet landed on main.

🤖 Generated with Claude Code