metabolomics: run targeted panels in raw mode (platforms, internal standards, LOD, tank units)

[sr320]

The metabolomics workflow's raw_reanalysis path had never been run. It fitted only two groups, using TIC normalization and log2(x + 1). That's reasonable for an untargeted feature table but wrong for a targeted MRM panel. DUPERRET2025_POMS_STARVE's Table S12 is the first real case:

• four platforms on different scales;
• spiked internal standards;
• 0.0001 below-detection placeholders, which log2(x + 1) would read as real intensities.

This PR adds what a targeted panel needs. Every option defaults to the old behaviour, and the processed-mode demo is unchanged.

New params.metabolomics options

Option	What it does
normalization	tic (default), median or internal_standard. Runs separately within each value of an optional platform column. Rows with is_internal_standard true are used for normalization only and are never tested.
detection_limit	Values at or below it become missing, not zero.
replicate_unit	Averages log2 values within a unit (e.g. a tank) before limma. Refuses a unit that sits in both arms. Matches the RNA-seq and methylation option.
min_detected_per_arm	Tests a feature only if it's detected in enough replicates per arm (default 2). Every feature, tested or not, goes in a new *_missingness.tsv with its status.
acquisition	untargeted or targeted. Written to each row's analysis_method.

Other changes

• limma now writes lfcSE (its posterior standard error). aree harmonize carries it into standard_error and reads analysis_method. Older processed tables without these columns harmonize as before; a new test covers both cases.
• The manifest records every params.metabolomics value. Before, it recorded only mode.
• The report template defines %||% when it's missing (R < 4.4) and summarizes normalization and the detection filter.
• CI runs the raw DAG for real, not stubbed, on a new simulated targeted-panel fixture (tests/fixtures/metabolomics_raw/). scripts/check_metabolomics_raw_smoke.py checks three things: the planted effects come back (+2.0 → +1.98, −1.5 → −1.61, +1.0 → +0.84); internal standards and undetected features stay out of the results; and the manifest records the parameters. The CI R setup gains bioc::limma.

Verified locally: the processed demo, the raw fixture (check_metabolomics_raw_smoke.py and check_manifest_versions.py), the real DUPERRET2025_POMS_STARVE panel, the full test suite (609 passed) and ruff.

The DUPERRET study files will come in a separate study PR, once the curation decision on its batch confound is made.

🤖 Generated with Claude Code