Fix static_inclusion_mode=INCLUDE against a cohort with no statics (fixes #16)

[florian6973]

Closes #16.

Two commits, and the split matters — the first alone would have been a fixture change wearing a fix's clothes.

1. The guard (c1e5613) — the actual fix

A model whose datamodule sets static_inclusion_mode: INCLUDE, run against a MEDS dataset with no baseline variables, dies inside the dataloader with:

ValueError: Cannot infer dtype from empty values; provide an explicit `schema=`

meds-torch-data builds its static tensors from whatever the cohort holds; when that is nothing at all, JointNestedRaggedTensorDict has no values to infer a dtype from. The error names neither the config key that asked for static data nor the cohort that lacks it, and it arrives once training has started.

The mismatch is knowable before any work happens. The tokenized cohort already records each subject's statics in tokenization/schemas/*/*.parquet as a static_code list, so "no subject has any static measurement" is one scan of one column. require_statics_if_requested does that scan and raises with the cause and both ways out; it runs inside build_datamodule, the single place every command constructs a datamodule, so no command can skip it.

Same shape as _require_split_sharded: cheap to check, expensive to discover later.

datamodule.config.static_inclusion_mode is INCLUDE, but no subject in /…/data/patients has any
static measurement, so meds-torch-data would fail to collate a batch at all ('Cannot infer dtype
from empty values').

Static measurements are MEDS rows with a null `time` — baseline variables such as age, sex or
ethnicity. Either:
  * the source dataset genuinely has none, in which case set
    `datamodule.config.static_inclusion_mode=OMIT` and drop them from the model; or
  * they were lost in preprocessing — check that `external_meds_dir` carries null-time rows and
    that any `pipeline=` you passed preserves them.

A cohort with no statics is legitimate MEDS, so this is a mismatch between cohort and request, not a broken dataset. Whether meds-torch-data should additionally collate an empty static set is a separate question, upstream of here — this makes the failure actionable, it does not make the case work.

2. The fixture (009f9ba) — why this was invisible

build_signal_dataset emitted nothing but timestamped events, so the cohort it produced had no static measurements at all. That is why the learnability tier could not exercise any model that reads baseline data: it failed before reaching it. The fast tier could not see the problem either, because the meds_testing_helpers fixture does have statics.

Both builders now give every subject one value-less code (BASELINE//GROUP_{A,B}) and one numeric one (BASELINE//AGE). with_statics=False reproduces the old cohort — which is what the guard's test uses.

The independent RNG is load-bearing, not tidiness. The statics come from random.Random(seed + 9973), separate from the generator that decides the labels. A static correlated with the outcome would be a second signal: a model could score well on the designed-signal test without ever reading SIGNAL_CODE, and the negative control would have something real to learn — the vacuous pass the control exists to rule out.

Tests

template/tests/test_synthetic_statics.py, in the payload rather than in the template's own suite — the template's tests install no torch, so they cannot collate anything, which is exactly why this was invisible here. Unmarked, so it runs in the fast tier. Five checks:

1. every subject has both kinds of static measurement, with the value/no-value split right;
2. the codes reach metadata/codes.parquet, so they land in the vocabulary;
3. neither group is perfectly separable by label (the leakage check);
4. a statics-free cohort is refused with the actionable error, through build_datamodule, and OMIT still works — the guard;
5. a real tensorized batch collates under INCLUDE with populated static tensors — the crash itself.

Verification

• Template suite: 70 passed (was 64); ruff check / ruff format --check clean, payload included.
• Dropped into the TECO repo, which is what copier update will do: fast tier 76 passed, 0 skipped, lint clean. Its local add_baseline_variables workaround — ~40 lines of exactly commit 2, written because a generated repo cannot fix the fixture from the inside — was deleted, and the learnability tier still passes: designed signal AUROC 1.0000, negative control 0.3158 against a 0.70 ceiling, identical across two runs. (The control moves from the 0.3885 that repo measured with its own workaround: same fixture shape, different draw.)

🤖 Generated with Claude Code