SH-aLRT: draw RELL replicates as bootstrap samples whatever -j/-J selected (#198)

[cindykrafft]

Fixes #198.

PhyloTree::resampleLh() builds the RELL replicates of the SH-aLRT test (and of -lbp) with
Alignment::createBootstrapAlignment(), which follows the global --jack-prop setting. With -j
or -J every replicate was therefore a delete-half jackknife sample: a sum over half the sites,
whose expectation is half the full-alignment log-likelihood. The statistics lh_new - lh in
testOneBranch() were then no longer centred, and the support of every branch moved. On
test_scripts/test_data/turtle.fa with the test seed:

run	SH-aLRT values on the same ML tree
-B 1000 -alrt 1000	98.4 89.8 98.2 61.7 47.6 47.9 98 100 81.1
-J 1000 -alrt 1000 (before)	100 100 100 91.7 86 81.5 100 100 99.7
-J 1000 -alrt 1000 (after)	98.4 89.8 98.2 61.7 47.6 47.9 98 100 81.1

An independent numpy implementation of the testOneBranch() rule on IQ-TREE's own site
log-likelihoods gives the same picture on a simulated 24-taxon alignment: bootstrap RELL 60.3 /
54.6 / 74.6 / 85.8 on four branches, delete-half jackknife RELL 74.0 / 82.9 / 97.6 / 100.0, and
the jackknife pseudo-value (weights × 2, expectation restored) 57.5 / 53.8 / 74.0 / 85.1 — the
inflation is the missing centring, not sampling noise
(https://github.com/cindykrafft/mytochondria/tree/main/audits/iqtree/issue-fixes/198-sh-alrt-jackknife).

Change: random_resampling() and createBootstrapAlignment(int*, ...) (Alignment and
SuperAlignment) take an optional jackknife proportion, negative by default meaning "the global
setting", and resampleLh() passes 0 so the SH-aLRT replicates are bootstrap samples as in
Guindon et al. (2010) and Adachi & Hasegawa (1996). The tree-search resampling of -j/-J is
untouched (UFJack and UFBoot values are bit-identical before and after), as are the
GENE/GENESITE/SCALE= paths. If you would rather keep jackknife replicates and rescale them
(lh_new / (1 - jack_prop)), that is a two-line variant of the same change; I went with the
definition of the test.

Test: test_iqtree.sh/.ps1 run turtle.fa with -B 1000 -alrt 1000 and with -J 1000 -alrt 1000 and fail if the SH-aLRT labels differ (they do on master); two expect_ans.txt rows for
the log-likelihoods. The full test_iqtree.sh + verify_results.sh pass locally with the change.

Not changed here, same pattern: the RELL replicates of the tree topology tests (--test, KH/SH
and the weighted variants in treetesting.cpp) also follow --jack-prop, and the AU test stops
with "Unsupported jackknife with SCALE=" when -j/-J is given. Happy to send that as a
follow-up if you agree it should behave the same way. #196 (the --jack-prop wording) already
has its fix on master (b8ef423) and could be closed.

---

Generated by Claude Code

[bqminh]

Can you remove the CI stuffs? This is not really the intension of the CI.

The CI is not really meant to be extensive test suite. It's only to check a few most popular use cases by users, that should mimic functionality of IQ-TREE v2. We also want to build binaries, so that some users can download and early beta test it, before releasing. The CI has to run quick enough, right now keeping ~20 min. We don't want to make it run longer.

If you want more testing, then this needs to be done with other means. You can post some input files and commands to reproduce it with the PR, if you want.

[cindykrafft]

Understood! I have removed the test_scripts changes
(test_iqtree.sh, test_iqtree.ps1, expect_ans.txt) and force-pushed; the branch is now the
six-file code change only, nothing that lengthens the CI run.

Here is the reproduction instead. It needs no extra input files: it uses turtle.fa, which is
already in the repository, and two runs of about a minute each.

iqtree3 -s test_scripts/test_data/turtle.fa -B 1000 -alrt 1000 --prefix turtle.alrt.boot -T 1 -seed 73073
iqtree3 -s test_scripts/test_data/turtle.fa -J 1000 -alrt 1000 --prefix turtle.alrt.jack -T 1 -seed 73073

grep -o ')[0-9.]*/' turtle.alrt.boot.treefile   # SH-aLRT values, UFBoot run
grep -o ')[0-9.]*/' turtle.alrt.jack.treefile   # SH-aLRT values, UFJack run

Both runs find the same ML tree with the same log-likelihood (-5370.3730), so the SH-aLRT values
should be the same up to the replicate noise of the test. On master they are not:

run	SH-aLRT on the nine internal branches
-B 1000 -alrt 1000	98.4 89.8 98.2 61.7 47.6 47.9 98 100 81.1
-J 1000 -alrt 1000, master	100 100 100 91.7 86 81.5 100 100 99.7
-J 1000 -alrt 1000, this branch	98.4 89.8 98.2 61.7 47.6 47.9 98 100 81.1

The UFBoot and UFJack support values themselves are unchanged by the patch, on this data
bit-identical before and after; only the SH-aLRT column moves.

A second check on simulated data (AliSim, 24 taxa, JC, 800 sites, -alrt 1000 alone / with
-bb 1000 / with --ufjack 1000) and an independent numpy re-implementation of the
testOneBranch() rule, which separates the missing centring from Monte-Carlo noise, are in the
kit if they are useful:
https://github.com/cindykrafft/mytochondria/tree/main/audits/iqtree/issue-fixes/198-sh-alrt-jackknife

Happy to run anything else you would like to see before merging.

---

Generated by Claude Code