ALOHA: sort the tag in combine_name so Majorana MadLoop output links

[oliviermattelaer]

fa710f7ad ("fixing some MSSM flavor merging issue") added tag.sort() to aloha/aloha_writers.py :: get_routine_name, so ALOHA now writes a loop wavefunction that also carries a conjugation flag as FFS4C1L3_2.f. combine_name, which builds the corresponding CALL, was not given the same normalisation and still emits CALL FFS4L3C1_2, in the order the caller assembled the flags — helas_call_writers inserts the loop tag in front of the conjugation tags.

Every Majorana loop line therefore calls a routine no file defines, and MadLoop output fails at link time:

ld: symbol(s) not found for architecture arm64
  "_mp_ffs4l3c1_2_", referenced from: _ml5_0_mp_coef_construction_1_
  "_mp_ffv1l1c1_2_", ...

Nothing in the SM or EW test surface puts a conjugation flag on a loop line, so the default suites stayed green while all gluino and neutralino NLO output was broken.

Why sort rather than revert

06e163cb0 (2013, "a small fix for calling the conjugated loop helas function, which will appear in process with majorana fermions") deliberately changed flag.append("L") to flag.insert(0,"L"): the L-before-C order at the call site was intentional and matched what ALOHA wrote when nothing sorted. fa710f7ad reordered the routine name from the other side and silently broke that.

Reverting tag.sort() also makes the tests pass, but presumably re-breaks the flavour merging it was added for. Sorting both sides retires the order-sensitivity instead of re-pinning it.

Not the same as 3a3e2c4

3a3e2c4a4 ("ALOHA: route the FLV flag through 'tags' for amplitudes in the cross-base name"), already on main, fixes the same class of bug in the same function — call site names a routine ALOHA did not write — but a different aspect: which placeholder carries the FLV M flag (%(propa)s vs %(tags)s) in the cross-base scheme. That flag never enters the tag list; it arrives through the format placeholders, and both naming schemes join tag before appending them. Disjoint statements, disjoint data. The link failure above reproduces on a pristine main that already contains 3a3e2c4a4.

MadGraph7 only

fa710f7ad is not on mg5/3.8.0 — get_routine_name there has no sort — so this fix must not be ported upstream: it would make the call side sort while the routine side does not, which is how you would introduce the mismatch. Confirmed empirically: the MSSM MadLoop comparisons link and pass on 3.8.0 with no change to combine_name.

Tests

suite	result
./tests/test_manager.py -p U loop	2 failures, both pre-existing standalone_cpp naming, identical on an untouched tree
LoopSquaredOrder_IOTest	1/1
short_ML_SMQCD_default / _optimized	2/2 each
long_ML_SMQCD_default / _optimized	2/2 each
short_ML_SMQCD_LoopInduced	1/1
IOExportFKSTest	6/6
test_short_mssm_vs_stored_HCR_gg_gogo_QCD	OK (link error before this change)
test_long_mssm_vs_stored_HCR_gg_gogog_QCD	OK
test_long_mssm_vs_stored_HCR_uux_gogog_QCD, gg_t1t1xg, gg_n1n1, uux_gogo	OK

Note combine_name is also called from the ALOHA writer side with the routine's own tag (aloha_writers.py:1329, 2400, 3114), and that same list is read a few lines later for data['addon'] = ''.join(self.tag). Sorting in place therefore reorders it — the goldens above cover that path and are unchanged, because those tags already arrive sorted. The function already mutated tag in place via the propagator shuffle, so this is not a new class of side effect.

Related

The loop-wavefunction mother corruption originally in this PR (LP#2159043) is an upstream bug present in 3.8.0 verbatim, and has moved to mg5amcnlo/mg5amcnlo#414.

🤖 Generated with Claude Code

[oliviermattelaer]

Scope changed: the LP#2159043 loop-wavefunction fix has been moved to mg5amcnlo/mg5amcnlo#414, targeting 3.8.0, since that bug is present upstream verbatim (reproducer crashes on a pristine 1ce45e2ae).

This PR now carries only the ALOHA combine_name fix, which is MadGraph7-specific — fa710f7ad is not on 3.8.0, so the fix must not be ported there.

Rebased onto current main (d9fc84d22) and re-validated after the rebase.