Wire lab z-scoring, run provenance, and a named eval split

[Rian354]

Five small changes on top of cdeb3e0, plus the paper launchers. Nothing here changes what data is collected.

Not included on purpose: the time_origin fix. cdeb3e0 already does it, and by inspection it's identical to the version on the trunk — same _hours_since helper, same anchor on admissions_to_process[0].timestamp, same call sites. No need to revert anything; this PR doesn't touch that logic.

1. Lab z-scoring — pyhealth/processors/lab_standardizer.py (130 lines) + ~13 lines of wiring

Per-feature z-score fit on the training split only; missing values stay missing. UnifiedMultimodalEmbeddingModel already accepted numeric_standardizers, so this is just the fit and the hand-off.

--no-lab-standardization runs raw labs as an ablation, so this is a default you can switch off rather than a commitment.

Measured on EHRMamba / labs+notes / seed 1 — one cell, one seed:

	best val PR-AUC
z-scores on	0.8011
z-scores off	0.7473

Why it's worth the lines: the ten labs sit on very different scales (Na 138.3 ± 5.2, K 4.14 ± 0.79, Ca 1.12 ± 0.32). Un-standardised, sodium's numbers are ~33× potassium's, so the model weights a feature by how large its units happen to be rather than how abnormal the value is. Models with input normalisation (LayerNorm in the transformers) absorb some of that; a plain GRU doesn't.

The transform itself is one line — z = (values - mean) / std, unobserved → 0. Most of the remaining lines are _indices, which exists because SampleDataset subclasses litdata.StreamingDataset, whose __iter__/__len__ shard by WORLD_SIZE: naive iteration would silently fit on 1/N of the train split under torchrun.

Not clipped, deliberately — there's no universally valid physiological range for these MIMIC category aggregates, so out-of-support values stay as large finite z-scores and remain auditable.

2. write_run_config — pyhealth/utils.py

metrics_history.json records what a run scored but not the conditions that produced it. This writes a run_config.json next to it with the resolved settings, not the raw flags.

That distinction matters: on an earlier branch a per-model optimizer override meant run_config stored adam_eps: null while the optimizer actually used 1e-6. A config logging flags would have shown nothing wrong.

Also records source_sha256 (a hash of the runner source), which is how you demonstrate a whole table came from one build, and torch. Written atomically, so a killed run can't leave a half-written config.

3. Named eval split

The inference loader fell back test or val or train. A run without a test split would evaluate on the training set and report the numbers as test, with nothing indicating it. Now the split is named, warned about, and recorded in run_config.json.

4. exp_name includes the task

It was {model}_seed{seed}, so --task labs and --task notes_labs at the same seed wrote to the same directory and the second run silently overwrote the first. With 36 paired cells planned, that would quietly destroy half the table. One line.

5. Restore emitted_data_version

cdeb3e0 removed self.emitted_data_version = 1. That field is part of vars(task), which is what the task-cache uuid5 key is built from — it's the only mechanism that invalidates a stale cache when emitted data changes. Without it, a cache built before a change is silently reused. cdeb3e0 changes every event timestamp, which is exactly when the bump is needed, so it's restored and set to 2.

scripts/paper/

common.sh holds the Tranche 1 protocol (identical for every cell); will.sh and rian.sh add only the data roots and the CPU tuning for their machine.

TASK=labs_notes MODEL=ehrmamba SEED=3 bash scripts/paper/rian.sh

will.sh keeps --num-workers 4 and no thread pinning, so it behaves exactly as it does today. rian.sh pins OMP threads and adds loader workers, because those nodes run several cells at once: unpinned, four concurrent cells put ~800 threads on 128 cores and epoch time went from 191 s to 8600 s with the GPUs sitting at 0–1%.

Split of work: rian seeds 1/3/5 on labs and labs+notes plus all of labs+notes+CXR; will seeds 2/4 on labs and labs+notes.

Verification

Ran a dev split before and after on this branch:

	before (cdeb3e0)	after, z-score on	after, z-score off
output dir	mlp_seed1	notes_labs_mlp_seed1	notes_labs_mlp_seed1
run_config.json	absent	present	present
lab_standardization	—	True	False
eval_split	—	test	test
standardizer buffers in checkpoint	—	mean/std/count present	none (correct)

Fitted values on the dev train split look right (Na 138.39 ± 5.27, K 4.13 ± 0.57). tests/test_p2_lab_standardizer.py passes (4 tests).

[will-pang]

Looks good to me