Make runANN work, and run it in CI

[HugoFara]

Closes#163. runANN has failed on line 2 since it was committed, and it is the only coverage for ANNPropagationModel and BMapLoggerForANNTraining — both read a .ffann network in their constructor and abort without one, so the C++ unit suite cannot construct them.

diff: result.txt.ref: No such file or directory
ANN results differs from reference.

Why a rename would not have fixed it

result.txt.ref is not in the repository, and it looks like result.txt was committed in its place. But regenerating that output changes a handful of lines in the last digit:

- result: 12412.1 observed 12412.1
+ result: 12412.2 observed 12412.1

Five lines out of 1424 here. An exact diff is red on any machine but the one that produced the file.

What it checks now

ANN_test already computes the root mean squared error between what the network predicts and what the propagation model produced — and run.bash was discarding that line with grep result. That is now the check. No reference file, no dependence on the last digit.

The tolerance is calibrated against this fixture rather than guessed:

	RMSE
trained network	0.0235
predicting the mean of every output	0.0966
normalisation weights scaled by 1%	17817
tolerance	0.05

The middle row is the one that set the number. My first attempt used 0.1, which a network ignoring its inputs entirely would have passed — a test that cannot tell those apart is not worth running. 0.05 leaves the real network a factor of two of headroom and still fails a constant predictor.

The fixture is weak, and that is recorded

The expected outputs take four distinct values spanning 1.1 in 12412. That is why the margin between a working network and a constant one is only a factor of four, and why perturbing a single normalisation weight by 5% moves the error by 2%.

The check is still meaningful — 1% damage to the normalisation layer takes the error to 17817 — but it is closer to a smoke test for the ANN path than to a numerical regression test. Replacing modelrun.csv with inputs that produce a real spread of rates of spread would fix that, and needs someone who knows what the fixture ought to represent.

Also in here

• result.txt is removed. It was the golden file for a comparison that no longer exists; ANN_test ... print regenerates that output on demand.
• clean.bash deleted results.txt, which is not the name of anything the suite writes. There is nothing to clean now.
• TESTING.md said runff was the only test validated in CI. That stopped being true when the unit suite and the moisture invariants landed. It now documents runANN, the calibration table above, and what tests/run.bash actually runs.

Not done here

tests/python/idealizedwind.py, farsite_flat.py and percolation.py are listed in #163 as unwired. They are examples rather than tests — they produce plots and assert nothing — so wiring them into CI means deciding what they should assert first. They are documented in TESTING.md as the worked examples they are, and left for their own change. The mnh_* cases need a Meso-NH build, which is its own question.

Verification

• runANN passes on the host and in ubuntu:24.04, with the same RMSE to seven digits — which is the portability this replaces the exact diff for.
• Negative control: scaling the normalisation weights by 1% takes the RMSE to 17817 and fails the check.

---

This pull request, including its code changes and this description, was generated by Claude Opus 5, and reviewed manually before submitting.

[HugoFara]

Test only, merging.