E2E gate degrades silently: clang #if-hollowing, silent local skip, and sibling-checkout drift (verified J2 unload regression vs vector-solver)

[petlenz]

From the 2026-08-02 full-repo review (finding F-13). Drift verified live.

Three ways the strongest test layer silently weakens:

1. clang hollowing — tests/generated/numsim_material_check_driver.cpp:34 compiles out 11 of 17 e2e tests under clang (#if defined(__GNUC__) && !defined(__clang__)) with no skip report. Per tests/CMakeLists.txt:299 the numsim-core pin now includes the numsim-core#16/Open-set extensibility: user-registered emitters for custom cas nodes #17 history_property fix, so the gate is likely stale. Replace with GTEST_SKIP() (visible in ctest output) or remove after re-testing under clang-19.
2. Silent local skip — with no local checkout and NUMSIM_CODEGEN_FETCH_MATERIALS=OFF (the local default), the whole Phase-B gate SKIPs with only a configure STATUS line (tests/CMakeLists.txt:386-390). Local green proves much less than CI green, invisibly.
3. Sibling-checkout drift — local checkouts silently override the CI pins. Verified concretely: against numsim-materials feature/vector-solver (0ea3801, of which the pin aaa6614 is an ancestor), NumSimMaterialEndToEnd.J2PathDependentRetainsPlasticStrainOnUnload FAILS (plastic strain grows 0.0077→0.0120 during elastic unloading), while the pinned SHAs pass 17/17. So (a) local red/green and CI red/green are different claims, and (b) the next pin bump imports a real behavioral regression unless the vector-solver unload semantics are resolved upstream first.

Proposals: GTEST_SKIP() instead of #if; a loud configure warning when a local sibling checkout's git rev-parse HEAD differs from the pin; raise the unload behavior with numsim-materials before any pin bump; optionally a CI step asserting the expected e2e test count so a fetch-flag regression can't silently drop the gate.

[petlenz]

Part (c) is now filed upstream where the fix/decision lives: NumSim-Stack/numsim-materials#22 (plastic strain accumulating during elastic unloading on feature/vector-solver, verified regression vs the aaa6614 pin). This issue's remaining scope: hold the materials-pin bump until numsim-materials#22 resolves, then bump and re-run the e2e gate.