Unit tests for every propagation and flux model, and -Wall -Wextra

[HugoFara]

Unit tests for every propagation and flux model, and -Wall -Wextra

Targets dev.

tests/runff runs one case, through one propagation model — Rothermel, set in params.ff. Nothing was watching the other sixteen propagation models or any of the sixteen flux models. This adds tests that exercise one model at a time, without running a simulation, and turns on warning flags.

21 test cases, 2082 assertions, ~0.07s. No new dependency: the framework is vendored.

How a model gets tested

getSpeedForNode splits in two — the DataBroker gathers properties out of the simulation into a double*, then the model does arithmetic on that array. Only the second half is physics, and it needs nothing but the array.

So ModelSandbox builds an empty FireDomain purely to instantiate models, and Inputs fills their property array by property name rather than by index:

ModelSandbox sandbox;
PropagationModel* model = sandbox.propagation("Balbi2020");
Inputs in(model);
fillStandardConditions(in, model);
in.set("normalWind", 5.0);
constdouble ros = model->getSpeed(in.data());

Addressing by name matters. A model's properties are numbered in the order its constructor calls registerProperty, so a test written against raw indices would keep passing after someone reorders that constructor — while silently measuring a different quantity.

Two kinds of assertion, and they are not equally trustworthy

Invariants — no spread without fuel, more wind never means less spread, a downslope is not an upslope, total released energy does not depend on how the time window is cut. These hold whatever the implementation is, and they are the ones worth trusting.

Pinned values — recorded rates of spread for the standard fuel records what ForeFire produces today, at 1e-5 relative tolerance so that -march=native and floating-point contraction differences between runners do not trip them. A pin moving means a model changed; that may well be intended, but it should be a decision rather than a surprise. The pins carry no claim of matching published values — someone who knows the literature should confirm them, and that is a good reason to look at test_propagation_models.cpp closely.

Compiler warnings

ForeFire's own sources now compile with -Wall -Wextra (/W4 on MSVC). The warnings are not clean yet — about 430 on Linux/GCC — so they are informational rather than fatal. Two options:

• -DFOREFIRE_WARNINGS_AS_ERRORS=ON — fatal, for whoever is clearing a file.
• -DFOREFIRE_ENABLE_WARNINGS=OFF — quiet.

NetCDF's headers are now included as SYSTEM, which stops netCDF::file_id defined but not used appearing in every translation unit.

One fix in src/, in its own commit

ForeFireModel's constructor left properties uninitialised. The six models that register no properties — Iso, heatFluxBasic, factorChemFlux, CraterVaporFluxModel, LavaSO2Flux, vaporFluxBasic — never overwrite it, so destroying one passed whatever the member happened to be built over to delete[]. The same destructor freed fuelPropertiesTable with delete[] while DataBroker::extractFuelProperties allocates it with a scalar new.

Neither has ever been observed, for the reason in the next section.

Three things found while writing these, not fixed here

Models are never destroyed.FireDomain keeps them in propModelsTable and fluxModelsTable and frees neither, so every model a simulation instantiates is leaked — and the code that would trip the bug above never runs.

properties is deleted twice. Seventeen flux models and two propagation models delete it in their own destructor, and ~ForeFireModel deletes it again; most also use scalar delete on a new[] allocation. Destroying any model that registers at least one property is a double free. The fix touches nineteen files and deserves its own PR — which is why test_model_registry.cpp only destroys the models that register none.

BalbiNov2011 responds non-physically to live fuel moisture at the values in the shipped fuel table. xsi exceeds 1 for fuel 1 of tests/runff/fuels.csv, the flame-temperature term goes negative, and R00 raises it to the fourth power. Rate of spread therefore falls with rising live moisture up to about Ml = 0.8 and then climbs again: 1.3e-3 m/s at Ml = 0.5, 1.9e-8 at Ml = 0.8, 8.2e-5 at Ml = 1.0 — which is the value the table ships. The monotonicity test stops short of that inversion rather than asserting it is correct. This one is worth a maintainer's eye; I did not want to change physics in a test PR.

Test isolation

Parameters are process-global — every SimulationParameters method reads and writes GetInstance() whatever instance it is called on — so a test setting Iso.speed, or burningDuration to zero, would change the result of whichever test ran next. ModelSandbox snapshots the parameter map on construction and restores it on destruction.

That restore is load-bearing, not precautionary: with it removed, three of five --order-by=rand seeds fail. tests/unit/README.md records the check.

CI

Both native workflows gained a ctest step. They also now trigger on pull requests to dev, which they did not before — main.yml already carried a commented-out dev push trigger.

The macOS run is the more useful of the two here: it is what checks the pinned rates of spread survive a different compiler and a different CPU.

Vendoring

doctest 2.5.3, single header, unmodified, under third_party/ with a README recording version, source URL and licence. Nothing there is compiled into libforefireL — CMakeLists.txt globs src/*.cpp only, and the directory is on the include path of the test target alone. A build and a CI run need no network beyond cloning.

FOREFIRE_BUILD_TESTS defaults on, except for wheel builds, which stay unaffected.

Verification

• 21/21 cases, 2082/2082 assertions; ctest 3/3.
• Isolation holds under --order-by=name|suite|file and five random seeds.
• tests/runff still passes — KML and NetCDF both within tolerance.
• -DFOREFIRE_BUILD_TESTS=OFF and -DFOREFIRE_ENABLE_WARNINGS=OFF both build clean.

Adding a test

tests/unit/README.md covers it. Models needing an external resource cannot be covered here — ANNPropagationModel and BMapLoggerForANNTraining read a .ffann network in their constructor and abort when it is missing; tests/runANN covers those.