Repair what the 'standalone' rename left behind

[oliviermattelaer]

Why

PR #64 renamed the export formats — the Fortran standalone became standalone_fortran, and standalone now names the MadMatrix (C++) output. Those commits are dated 2026-08-11 but sat on feat-big-renaming and only reached main on 2026-09-02 (d7f4f8b21). Every branch written against the old vocabulary in the meantime kept saying standalone and silently changed meaning at the merge. No textual conflict, so git flagged nothing.

What was broken

A user-facing regression. LOOP_INDUCED_FORMATS still named standalone, so for g g > h [noborn=QCD]:

command	before	after
output standalone_fortran	refused ("does not support loop-induced processes")	writes the MadLoop directory
output standalone (MadMatrix)	KeyError: 5 in export_mg7.set_channels_colors_map	refused, pointing at [sqrvirt=]

The Fortran standalone is the format PR #81 taught to accept loop-induced processes, and it is the one with the MadLoop backend; the crash is exactly what the guard in ExportCPPFactory exists to prevent. [sqrvirt=] / [virt=] through the MadLoop interface are untouched (checked by hand and by the tests).

Three tests that named the format.

• test_loop_induced_output.py asked ExportV4Factory for 'standalone' and asserted a sqrvirt refusal from standalone_cpp, a format removed in ea8593da3. The prefix trap it guards has flipped round — 'standalone' is now the prefix without a loop backend — so that is what it pins now.
• testIO_FDgauge_standalone_fortran generated a madmatrix directory, which has no Source/DHELAS, so its two reference files stopped being produced. 7840ba898 ("update IOTest", the evening of the Feat big renaming #64 merge) then ran with -U and deleted them. The test has reported OK ever since while comparing nothing. Both files are restored here byte-for-byte (git hash-object matches the blobs deleted in 7840ba898, and they match today's output); tampering one now fails the test again, so it has its detection power back.
• testIO_loop_induced_standalone_output likewise — see the caveat below.

An order dependency of the same family. testIO_UnitProcOutputIOTests was the only IOTest in its file with no @set_global, so it exported in whatever gauge an earlier test left behind. tests/unit_tests/iolibs/test_ufo_parsers imports a loop model through a MasterCmd, which switches aloha.unitary_gauge to Feynman and never switches it back; loop_sm then keeps the charged Goldstone and get_color.f starts at CASE(-251) instead of the ghost CASE(-82). Pinned on one side, restored on the other.

Two things deliberately left open

1. testIO_loop_induced_standalone_output fails on this branch. Pointing it at standalone_fortran makes it run for the first time since Feat big renaming #64 (it crashes on main), and it then finds its goldens stale in two independent ways: check_sa.f has a debug CALL ..._COMPUTE_RES_FROM_JAMP that only ever lived on the dropped PR loop-induced: stop the run when the poles do not cancel #80 branch, and loop_matrix.f is missing the KEEP_OFFSHELL_MASS block that reached main on 2026-08-25 — four days after PR loop-induced output: fix standalone, refuse the formats with no MadLoop backend #81 recorded these goldens, and PR loop-induced output: fix standalone, refuse the formats with no MadLoop backend #81 merged on 09-05 without re-recording. Regenerating them is a maintainer call, so this PR does not.
2. testIO_UnitProcOutputIOTests still fails in a full-suite run, from a different leak the gauge fix does not cover: tests/unit_tests/fks/test_ewsudakov imports loop_qcd_qed_sm_Gmu_forSudakov, and a later import_ufo.import_model('loop_sm') then groups mdl_MW under ('aEWM1', 'Gf') instead of ('aEWM1',), which reorders mp_intparam_definition.inc. Same expression, same externals — only the dependency key changes, and the leaked key is unprefixed (Gf, not mdl_Gf), which points at the shared UFO module registry in models/__init__.py:load_model. Pre-existing and unrelated to the rename.

Why CI did not catch any of it

• Merges on main cancel each other. cancel-in-progress is keyed on the ref and main is a ref like any other: of the last 25 IOtest runs on main, 9 were cancelled — loop-induced output: fix standalone, refuse the formats with no MadLoop backend #81 and fix: keep loop_Nc_power out of the process-wide colour memo #84 landed 9 s apart, madspace: bind LHEEvent.alpha_qcd to the right member #87/madspace: spell the LHEParticle momentum kwargs like its attributes (px/py/pz) #88 21 s apart, Cheaper colour sum: DDM colour basis, Kleiss-Kuijf flows, and the same for madmatrix #60 was killed by CI: stop the xsec and madspace workflows running twice per push #129 36 s later. The post-merge run is the only one that sees the merged tree, which is where this class of conflict lives.
• unittest.yml runs a hand-maintained list of test names, not the suite. PR loop-induced output: fix standalone, refuse the formats with no MadLoop backend #81's two loop-induced tests are in no workflow at all — they have never run in CI.
• A missing golden does not fail. The missing-file branch in IOTests.py only prompts interactively; otherwise the test degrades to comparing nothing and reports OK. Only ./tests/test_manager.py -R -F 10 lists them as MISSING.

Testing

• ./tests/test_manager.py -p U -r 1 test_loop_induced_output — 16/16 green (was 1 failure + 1 error).
• ./tests/test_manager.py testIO_FDgauge_standalone_fortran testIO_FDgauge_madmatrix -pA -t0 — green, and verified non-vacuous by tampering a reference file.
• ./tests/test_manager.py -R -F 10 — no MISSING left (was 2).
• Full ./tests/test_manager.py: 1 failure + 93 errors, down from 3 + 94. The remaining failure is item 2 above; the 93 errors are the pre-existing madspin-density/numpy ones, identical on main.
• By hand: output standalone_fortran on g g > h [noborn=QCD] writes Cards/MadLoopParams.dat; output standalone refuses; [sqrvirt=] through bare output standalone still writes a MadLoop directory.

🤖 Generated with Claude Code