madspace: evaluate the matrix element on the momenta the event is written with

[oliviermattelaer]

The initial-state mirror covers the beam-swapped copy of an asymmetric initial
state by rotating the event by pi about x (py, pz -> -py, -pz,
kernel_mirror_momenta), which moves each leg onto the other beam. Only the
written event was rotated: Integrand::build_common_part kept feeding the
cross section momenta_acc, the momenta as generated.

That is sound only if |M|^2 is invariant under the rotation.

Why it is wrong

It is invariant for any Lorentz invariant matrix element, and for a
longitudinally polarised one. It is not for a transversely polarised
particle evaluated in a frame that holds it at rest: HELAS quantises such a
particle along the frame z axis (the pp == 0 branch of vxxxxx) rather
than along its own momentum, the mirror flips that axis, and the + and - states
swap.

Measured in FP64, |M(mirror p)|^2 / |M(p)|^2 - 1 on g q > z{+} q:

frame the ME is evaluated in	unpolarised	z{0}	z{+}
as generated	6e-16	2e-14	1e-15
partonic CM	4e-16	8e-15	3e-16
rest frame of beam 1 + Z	8e-16	1e-14	6e-15
Z rest frame	6e-15	6e-15	4.8e-01

So it is not "any boosted frame" — it is specifically a transversely polarised
particle held at rest, which is what asking for the Z rest frame does.

What it cost

On a symmetric pp collider the two beams cannot be told apart, and the mirror is
precisely what averages over them, so sigma(p p > z{+} j) and
sigma(p p > z{-} j) have to be equal. They were not:

	p p > z{+} j	p p > z{-} j	split
before	4485 +- 39 pb	6060 +- 54 pb	23 sigma
after	5299 +- 32 pb	5265 +- 33 pb	0.7 sigma

The sum over the three polarisations was always right (10545 +- 67 vs
10564 +- 46), so the bug only misallocated between + and -. That is why no
unpolarised result moves — p p > z j is 12420 +- 69 pb before and
12430 +- 110 pb after — and why nothing ever caught it.

The fix

Evaluate the matrix element on the momenta the event is written with. No
symmetry assumption at all. Two things already assumed this:

• madevent does the same thing: it flips the momentum array itself and
  passes that same array to both the matrix element and the event record
  (super_auto_dsig_group_v4.inc, "Flip momenta (rotate around x axis)";
  write_leshouche's "flip back particle 1 and 2" only relabels the two beam
  legs, it does not undo the rotation). So this aligns madspace with madevent
  rather than inventing a convention.
• SystematicsCalculator::matrix_elements reads the momenta back out of the
  event buffer, i.e. the mirrored ones, so the nominal weight and its
  scale/PDF variations were computed on different momenta for the same event.

Impact and testing

Half the events of a mirrored subprocess now have their matrix element
evaluated on sign-flipped momenta, so existing results shift in the last bits
and LHE hashes move. The reproducibility tests compare runs to each other and
the xsec references have a 1% tolerance, so nothing should break — the
unpolarised control above is the direct check.

The regression test that guards this needs the me_frame run-card setting to
reach the broken branch at all, so it travels with the follow-up PR that adds
me_frame (test_polarised_mirror_symmetry_mg7, CI job
acceptancetest_mg7_polarised_mirror). It fails on this PR's parent commit and
passes with it.

madspace test suite: 1545 pass. Full unit suite: unchanged against main (the
two reds on main — a missing scipy and a MadLoop IOTest golden drift that only
appears in a full-suite run — reproduce on a clean checkout).

🤖 Generated with Claude Code