[lsimon125]

Summary

NativeRuntime::set_model_parameters did not refresh particle masses that are derived model parameters.

In a G_μ-scheme UFO (the serialized sm), MW is derived from aEWM1, Gf and MZ, so it has no runtime slot. The loop
over particle_mass_parameter_names only read masses through model_parameter_runtime_slots. A derived mass
therefore kept the value derived from the artifact defaults:

• model_parameter_value("MW") reported the refreshed value, e.g. 80.3692 GeV for aEWM1 = 132.1001485837534 and
  Gf = 1.1663788e-5;
• the W wavefunctions and propagators still used MW = 80.419 GeV.

The polarisation sum was then taken on a different mass shell from the external momenta. As a result, |M|² of
u d~ > W+ H was not Lorentz invariant: it changed by 5.0e-5 under a z-boost at √s = 500 GeV.

The fix applies to masses without a runtime slot. It takes the refreshed real component of the corresponding
derived_parameter_component, which is the value the derived-parameter evaluator has just written. Settable masses
are unchanged.

Test

New test: rust/crates/rusticol-core/tests/derived_particle_masses.rs. It runs only when
RUSTICOL_DERIVED_MASS_ARTIFACT points to an artifact generated with
pyamplicol generate 'u d~ > W+ H' DIR --model models/json/sm/sm.json --restriction no_b_mass --execution-mode compiled
(QED = 2).

The test:

1. sets aEWM1 and Gf;
2. checks that MW moved;
3. evaluates the process at an on-shell point (built from the refreshed MW and MH) and at the same point boosted
   along z with rapidities 0.5 and −1;
4. requires the relative change to be below 1e-12.

	before	after
boost with y = 0.5	5.03e-5 (fails)	< 1e-12 (passes)

Both runs used the same generator and artifact. The "before" run is loop at bae3b9e plus #26; the "after" run adds
this commit.

On loop at 36f4325 plus this commit (the PR branch):

• the new test passes;
• cargo test --release --features symbolica-runtime --lib gives 1132 passed, 0 failed, 1 ignored (with
  PYAMPLICOL_TEST_PYTHON set).

Downstream, ROULETTE's pp → W±H LO cross section against MG5_aMC with the same G_μ inputs moved from −0.06 % to
+0.001 % (0.09σ) with this fix.

🤖 Generated with Claude Code