Add sativaepang/* modules: reference, lootasks, looplace, looscore

[erikrikarddaniel]

PR checklist

• This comment contains a description of changes (with reason).
• If you've fixed a bug or added code that should be tested, add tests!
• If you've added a new tool - have you followed the module conventions in the contribution docs
• If necessary, include test data in your PR.
• Remove all TODO statements.
• Broadcast software version numbers to topic: versions - See version_topics
• Follow the naming conventions.
• Follow the parameters requirements.
• Follow the input/output options guidelines.
• Add a resource label
• Use BioConda and BioContainers if possible to fulfil software requirements.
• Ensure that the test works with either Docker / Singularity. Conda CI tests can be quite flaky:
  • For modules:
    • nf-core modules test <MODULE> --profile docker
    • nf-core modules test <MODULE> --profile singularity
    • nf-core modules test <MODULE> --profile conda

Description

Adds four new modules exposing sativa-epang's staged
leave-one-out CLI, one nf-core module per natural stage, replacing what would otherwise be a single
monolithic-CLI wrapper:

• sativaepang/reference — builds a SATIVA reference (refjson + model) from an alignment and
  taxonomy, optionally handing off a tree built elsewhere (e.g. RAxML-NG, see
  raxmlng/search) via -reftree/-refmodel
  instead of running sativa-epang's own constrained RAxML search.
• sativaepang/lootasks — deals the refjson into leave-one-out folds via -stage loo-tasks, one
  self-contained task directory per fold plus a manifest.json.
• sativaepang/looplace — places every leave-one-out fold with EPA-ng via -stage loo-place,
  writing a jplace (plus logs) into each fold directory.
• sativaepang/looscore — reads the placed folds and reports mislabelled sequences via
  -stage loo-score, completing the chain.

These four modules will back a fasta_tax_detectmislabels_sativaepang subworkflow (opened
separately once this and raxmlng/search have landed) used by
nf-core/taxmarker.

Notable implementation details

• lootasks works around a real sativa-epang bug: -stage loo-tasks -r (a refjson built from
  an externally-supplied -reftree/-refmodel, as ours always is) leaves no RAxML_info file for
  the tool to recover its model from, and embeds the .model file's raw content — including
  RAxML-NG's own trailing partition clause (, name = range) — verbatim into manifest.json.
  loo-place then rejects every fold ("Wrong model specification"), since inline-mode EPA-ng
  doesn't accept that clause the way file-mode does at the reference stage. Worked around by
  exporting SATIVA_EPANG_MODEL ourselves, trimmed to the bare model expression. Reported
  upstream; not yet fixed as of this PR.
• looplace stages a private, writable copy of its input taskdir before running loo-place,
  rather than mutating the staged directory from lootasks in place — the same class of
  resume-breaking bug fixed in transdecoder/predict cannot be resumed: it writes checkpoints into its staged input, then fails on re-execution #12799 (TRANSDECODER_PREDICT). Verified by hand: staging a
  read-only unplaced taskdir through the same mv+mkdir+symlink script confirmed the original
  directory gained zero new files.
• reference's refjson output embeds its own build invocation (working-directory-relative
  paths and all), so it isn't byte-reproducible across engines — masked via sanitizeOutput's
  unstableKeys, with explicit JSON-structure assertions covering what the module actually
  guarantees instead.
• sativa-epang has no --version flag and its own internal version string is stale (inherited
  from upstream SATIVA); the installed conda package version is read back out of the share/
  path baked into the sativa-epang wrapper script at build time.

Test data

Both tiers now live in nf-core/test-datasets#2266 (merged): the tool's own bundled 38-sequence
tiny example, and a curated 121-sequence GTDB archaeal 16S dataset (which carries a deliberate
swapped-label pair used as a real detection assertion in looscore's tests, not just
process.success).

🤖 Generated with Claude Code

https://claude.ai/code/session_014eSqkuYZXfNiWy87y3srup

[erikrikarddaniel]

The conda CI job failed: EPA-ng's placement likelihoods aren't byte-reproducible between the conda environment and the container builds (identical inputs/logs, differing epa_result.jplace content). epa_info.log/epang.log also carry a wall-clock "Elapsed Time: Ns" line that can flip between runs under load, on any engine.

Pushed df09b4f masking all three via sanitizeOutput's unstablePatterns (same precedent as sigprofiler/snapaligner elsewhere in this repo). The structural asserts already in the tests (every fold directory contains a jplace) remain the real correctness guarantee. Verified stable across repeated runs on Docker, Singularity, and conda.

🤖 Generated with Claude Code

[SPPearce]

Can you have a pass to trim out a bunch of the LLM comments, both in the modules and the tests.
Any # comment you put into the script block will be present in the .command.sh I think.

Can you also use sanitizeOutput basically everywhere in the nf-test files.

[SPPearce]

This is a lot of code, is it really required?

[erikrikarddaniel]

Some form of private writable staging is required: sativa-epang writes results into every fold directory, and taskdir can be staged as a plain symlink straight into lootasks' own task directory (default local-executor staging) — writing into it in place would mutate that other task's output and break -resume the same way #12799 did.

You're right that the generic version was overkill though: it was walking an arbitrary file/directory tree, but lootasks' output schema is actually fixed (manifest.json + fold_*/{ref.nwk,ref.fasta,query.fasta}, confirmed against a real multi-fold run — nothing else ever appears). Pushed in adf8590: hardcoded to that shape instead, about a third of the size.

[SPPearce]

Surely just stage the original with this name instead.

[erikrikarddaniel]

Same reasoning as in my reply above — happy to take another look if you still see room to trim further once you've seen the shrunk version.

[erikrikarddaniel]

Thanks for the review! Pushed adf8590 addressing all four points:

• Version eval: simplified to one grep -m1 -oE '[0-9]+\.[0-9]+\.[0-9]+\.[0-9]+' call across all four modules, same output.
• looplace's restaging (the three inline comments): shrunk considerably. It no longer walks an arbitrary file/directory tree -- lootasks' output schema is fixed (manifest.json + fold_*/{ref.nwk,ref.fasta,query.fasta}, confirmed against a real multi-fold run, nothing else ever appears), so the private-writable-copy logic is now ~6 lines instead of ~20.
  Kept the symlink-based approach rather than switching to stageInMode 'copy': lootasks already writes each fold's ref.fasta/ref.nwk as a real per-fold copy (not shared), so at GTDB scale (hundreds of thousands of genomes, thousands of folds) a second full copy on top of that in looplace isn't affordable -- happy to discuss further if you still think a full-copy approach is preferable for simplicity.
• LLM-style comments: trimmed throughout, including the two that were inside script: blocks and leaking into .command.sh.
• sanitizeOutput everywhere: added to looscore's two non-stub tests (the other three modules already had it) -- as a bonus it also dropped the redundant positional (0/1) snapshot entries that duplicated the named outputs.

Re-verified all 12 tests (4 modules) pass under Docker, Singularity, and conda.

🤖 Generated with Claude Code

https://claude.ai/code/session_014eSqkuYZXfNiWy87y3srup

[Aaramis]

Hi Daniel,

Small correction to "not yet fixed": the loo-tasks model bug is fixed in 5c6bd1e on
epa-ng. It just is not in any tagged release. v0.9.3.6 is the commit right before it,
and bioconda is on 0.9.3.4, so the pending bump to 0.9.3.6 would not help either.

I can tag v0.9.3.7 and push it to bioconda if that is useful. Happy to do it now if
you want to drop the workaround in this PR, or after it lands as a follow-up. Your call