`testImportModule` flakes in-suite with an empty-call-graph failure

[khatchad]

TestTensorflow2Model.testImportModule intermittently fails Function must exist in call graph in ~0.1s when the full suite (or full class) runs, while passing in isolation, in pairs, and in most full-class runs—including a run at the identical method order that had just failed. Occurrences so far: once on a full-suite run of the #691 branch (2026-07-03), once on a full-suite run of the #690 fix branch; an immediately-following identical full-class run passed 945/0.

The sub-second timing means the engine produced a call graph with the target function missing, without doing real analysis work—consistent with cross-test interference on shared JVM state rather than anything about the import fixture itself. Two leads:

• The root POM sets the parallel property to both (surefire reads it via its user-property alias) with threadCount=1 and reuseForks=true, so suite runs execute JUnit 4 tests with intra-JVM concurrency; any mutable static in the front end or engine is then a race candidate.
• A related deterministic poisoning was just documented in Multi-file mode binds script-toplevel class declarations and reads to different globals #692: a test that flips PythonCAstToIRTranslator.setSingleFileAnalysis(false) breaks 13 later tests across five classes even after restoring the flag, proving that translation mutates static state that survives engine construction. The flake may be the concurrent flavor of the same substrate.

Next step when picked up: rerun the suite with -Dparallel=none a few times; if the flake disappears, bisect the shared static.

[khatchad]

Third occurrence, same signature: full-suite run on the #661 shape-3 branch, testImportModule fails Function must exist in call graph in 0.112s; passes in isolation immediately afterward. The other 996 tests were green, so the flake is not correlated with the branch's changes (list-subscript read modeling).

[khatchad]

Occurrences 4 and 5 (2026-07-03)

testImportModule failed in-suite twice today, on consecutive full-suite runs of ponder-lab#536 branch (0.097s and 0.099s, Function must exist in call graph.), passing isolated immediately after each run. Same signature as the first three occurrences.