wasm-backend AC3.2 boundary: inline-fixture VM tests bypass the wasm parity hook

[bpowers]

Summary

The wasm-backend parity gate (wasm-backend.AC3.2 -- "every core-simulation corpus model runs through both the VM and the wasm backend") is enforced by an inline wasm_parity_hook that lives inside the shared corpus helpers simulate_path_with_excluding / simulate_mdl_path / simulate_mdl_path_with_data (src/simlin-engine/tests/simulate.rs) and the systems-format helper (src/simlin-engine/tests/simulate_systems.rs). Any VM-simulated model that flows through one of those helpers also runs through the wasm backend, and an Unsupported outcome is now a hard failure.

However, a number of tests in tests/simulate.rs build inline string-fixture models and run them through the VM with bespoke assertions without calling those helpers, so they never reach the wasm hook. A literal reading of AC3.2's success text ("every XMILE, MDL, and systems-format model in the corpus", phase_08.md line 22) would include these inline fixtures, so they sit outside the wasm gate by construction.

This is filed as a deliberate-coverage / AC-boundary note identified during the Phase 8 final review of the wasm-backend. It is not a defect -- the reviewer approved the phase for merge. It documents the boundary so a future session can decide whether to tighten it.

Concrete examples (verified)

Inline-fixture VM tests that bypass the hook (each calls open_vensim/open_vensim_with_data -> compile_vm -> Vm::new -> run_to_end with hand-written assertions, never simulate_*_path / wasm_parity_hook):

• simulates_except_basic_mdl (tests/simulate.rs:1306)
• na_existence_test_and_arithmetic_finite (tests/simulate.rs:1384)
• the GET DIRECT CSV trio simulates_get_direct_data_scalar_csv / simulates_get_direct_constants_scalar_csv / simulates_get_direct_lookups_scalar_csv (tests/simulate.rs:3286 / 3349 / 3389)
• several inline-fixture simulates_macro_* tests (e.g. simulates_macro_multi_expression_mdl, tests/simulate.rs:3566, which adds a namespace-isolation assertion the helper can't express)

Why the practical impact is small (verified)

Every wasm-supported feature these inline tests exercise is independently routed through the hook by a real corpus model, so no wasm-supported feature rests only on a non-hooked inline test:

• EXCEPT: simulates_except (:1288) and simulates_except2 (:1293) both call simulate_mdl_path(...) (-> wasm_parity_hook).
• GET DIRECT (data/constants/lookups/subs): directdata.mdl / directconst.mdl / directlookups.mdl / directsubs.mdl all go through simulate_mdl_path_with_data(...) (:3265-:3280), which ends in the hook.
• Macros: many simulates_macro_* tests route through simulate_mdl_path (e.g. simulates_macro_expression_mdl, simulates_macro_stock_mdl, simulates_macro_cross_reference_mdl).
• :NA: finite-sentinel arithmetic and basic vector ops / allocation likewise have corpus-routed simulate_*_path tests.

Why it matters

This is internal test-coverage texture, not user-facing. It matters only for the strength of the AC3.2 invariant: the gate's literal scope ("every model in the corpus") is broader than its implemented scope ("every model the VM simulates through the simulate_*_path / systems helpers in the default suite"). The implementer's notes already scope it this way deliberately (phase_08.md line 29: "The end-state gate applies to models the VM actually simulates in the default suite"), so the gap between literal AC text and implementation is intentional and reasonable -- but undocumented outside the plan.

Component(s) affected

• src/simlin-engine/tests/simulate.rs (inline-fixture tests + the wasm_parity_hook / simulate_*_path helpers)
• src/simlin-engine/tests/simulate_systems.rs (systems-format helper)
• src/simlin-engine/tests/test_helpers.rs (wasm_results_for / ensure_wasm_matches, the alternative hook entry points)

Possible approaches (optional future enhancement)

• Route the inline-fixture VM tests through the wasm hook too -- either by adding a wasm_parity_hook(&datamodel_project, &expected, &[]) call after their VM assertions, or by asserting parity directly via the wasm_results_for helper in test_helpers.rs. That would make AC3.2 hold in its strongest, most literal form.
• Alternatively, leave the boundary as-is and simply document (in the engine CLAUDE.md test notes and/or a comment near the inline fixtures) that the wasm parity gate covers only the simulate_*_path/systems-helper-routed corpus, so the boundary is discoverable without reading the implementation plan.

How it was discovered

Identified during the Phase 8 final review of the wasm-backend (branch wasm-backend-poc). The reviewer flagged it as a deliberate coverage-boundary worth tracking and approved the phase for merge. The "small impact" claim above was independently verified by spot-checking simulates_except_basic_mdl, the GET DIRECT CSV tests, and the macro fixtures, and confirming each feature is covered by a hooked corpus model elsewhere.