Fix MLM clustering through flavour-forbidden merged graphs (apply_flavor_grouping)

[oliviermattelaer]

Fixes the pre-existing failure of test_MLM_reweight.py::TestMLMReweightAutoMatch::test_mlm_rewgt_auto_match_qq_to_qq (39.1 sigma, reproduced on main and at ce5ebc6). The failure had two causes: a wrong test, and a real MLM bug that the wrong test was hiding.

1. The test compared different flavour content (factor ~3.8)

With group_subprocesses=False, apply_flavor_grouping=False splits q q~ > q q~ (q = u d s c) over three P-dirs:

P-dir	rows	channels
P0_uux_uux	8 (flavour-diagonal)	s-channel g, t-channel g
P0_uux_ddx	12 (q q~ > q' q~')	s-channel g only
P0_udx_udx	12 (q q~' > q q~')	t-channel g only

The flavour-grouped P0_QQx_QQx integrates all 32 rows. The test compared it with only the P-dir that has the most channels (P0_uux_uux).

2. Real bug: cluster() used flavour-forbidden graphs

After summing over P-dirs (50k points per channel):

channel	fg=True (main)	fg=False, summed
s-channel g	134610 ± 782	34680 + 84725 = 119405 ± 438 (17 sigma)
t-channel g	5.028e6 ± 0.070e6	5.095e6 ± 0.049e6 (0.8 sigma)

With ickkw=0 the s-channel numbers agree (134610 vs 134065, 0.6 sigma), so the difference is in REWGT. mlevel=8 traces show the cause. For the q q~ > q' q~' rows (flavour class 1,1,2,2), fg=True clusters beam q with outgoing q' through the t-channel gluon of the merged graph. That vertex does not exist for those flavours, and the merge gives REWGT = 1. The ungrouped P0_uux_ddx only has the s-channel graph: it clusters q' q~' -> g and applies alpha_s(ptij)/alpha_s(mtlast), about 0.85.

cluster() already filtered candidate graphs with flavor_compatible, but never checked that any graph survived. The filtered icgs was a dead local, so the filter had no effect at all.

Fix (Template/LO/SubProcesses/cluster.f, reweight.f)

• flavor_filter replaces flavor_compatible. It works on the event's merged-flavour indices, i.e. the class pattern from GET_FLAVOR that reaches REWGT, where equal index means equal flavour, as in check_flavor. It does not use the remapped PDG codes.
  • A g/γ/Z/h may join two merged legs only if they have the same index.
  • Everything else stays permissive, as it effectively was: unmerged or internal legs, W, unmapped merged leptons.
  • It is applied to every candidate pair, in the first pass and in the recomputation, and to igraphs after a winner is chosen. Pairs whose graph list is empty are no longer clustered.
  • The old function would have been wrong once switched on: it checked W up/down parity and sign on class indices, rejected unmapped 82 leptons, and rejected boson daughters such as W+ W- -> Z.
  • fg=False passes flavor = 1,1,..., so the filter is a no-op there.
• setclscales: the first event of a channel, and the "jets fail, recluster" path, force chcluster (this channel's graph only). For a flavour row that forbids that graph (s-channel config, u d~ > u d~ row), clustering now returns .false., which the callers already turn into zero weight. Before, this case would stop. The event has zero multichannel weight there anyway (AMP2(config)=0 for that flavour). Any other clustering failure still stops as before.

Test (tests/acceptance_tests/test_MLM_reweight.py)

• Compiles every P-dir. Each flavour-grouped channel is compared with the sum of its ungrouped counterparts, matched by s- and t-channel propagators with explicit flavours mapped onto 81/82/83.
• 50k points per channel: at 5k points the +13% only shows up as 4.4 sigma.
• run_single_channel deletes results.dat before each run. A Fortran STOP exits with 0 and otherwise leaves the previous channel's result to be read; this hid the chcluster stop above during development.

Validation

• New test fails on the old Fortran (17.0 sigma: 134610 vs 119405) and passes with the fix: s-channel 119810 ± 677 vs 119405 ± 438 (0.5 sigma), t-channel 0.8 sigma.
• -p tests/acceptance_tests '.*mlm_rewgt.*': Ran 5 tests, OK. This covers the manual qq, gg>tt~gg and DY+4j cases.
• test_flavor_grouping_consistency_mlm, test_madevent_dy3j_mlm, test_DY_onejet: Ran 3 tests, OK. No clustering errors in the logs.
• IOTests not run: the change touches only Template Fortran that the goldens do not cover, plus the acceptance test.