Assess PATH/environment semantics of command execution (venv exposure, bare-name resolution, -m runnability)

MEDS-DEV's whole run model is configured shell commands, which makes it the most PATH-semantics-sensitive repo in the ecosystem. This issue requests an assessment; below is a concrete audit (@ 60b678c) to start from. Context: this came out of the MEDS-Extract 0.7 migration work (Medical-Event-Data-Standard/MIMIC_IV_MEDS#58), where the same class of problem produced Medical-Event-Data-Standard/MIMIC_IV_MEDS#51; companion issues: mmcdermott/MEDS_transforms#398, mmcdermott/MEDS_extract#155, mmcdermott/meds-torch-data#115.

1. A latent bug found during the audit (probably wants its own fix regardless)

utils.py:385-388 — run_in_env appends source $VIRTUAL_ENV/bin/activate to script_lines after script has already been joined, so the activation line never lands in the written cmd.sh. Venv exposure therefore rests solely on the PATH prepend at utils.py:348. (The dead line also hardcodes POSIX /bin/activate, cf. #316.) Notably, closed issue #99 — meds-tab console scripts not found on Clalit despite existing in the venv bin/ — is field evidence that the PATH-prepend-only mechanism breaks on some server setups, and this dead code may be why the intended belt-and-suspenders never engaged.

2. Same-environment dependencies invoked via ambient PATH (activation assumed)

These are console scripts of MEDS-DEV's own pinned deps, run with no env= at all — they resolve via whatever PATH the calling shell has, i.e. they require the install env to be activated and can silently cross-resolve when multiple envs exist:

• tasks/__main__.py:34 — bare aces-cli (es-aces==0.6.1), shell=True
• tasks/__main__.py:40 — bare expand_shards inside $(...) command substitution (silent mis-resolution here corrupts shard lists rather than failing)
• evaluation/__main__.py:15 — bare meds-evaluation-cli (meds-evaluation==0.0.3)
• utils.py:284 — uv located by shutil.which on ambient PATH even though uv is a declared runtime dependency

Candidate directions: interpreter-pinned invocation (sys.executable -m, pending guards in those packages) or an activation-equivalent PATH (prepend sysconfig.get_path("scripts") — the approach just adopted in Medical-Event-Data-Standard/MIMIC_IV_MEDS#62, which also propagates transitively to grandchildren).

3. Cross-environment commands (the deliberate case — different trade-offs)

Per-dataset/model venvs are exposed to YAML commands: exclusively via the PATH prepend in temp_env (utils.py:348). That's intentionally cross-env, so sys.executable does not apply — but given #99, consider resolving commands against the already-computed venv_bin_path absolutely (or at minimum verifying the expected scripts exist there before running) instead of trusting prepended-PATH resolution through bash/shell=True. Also in this bucket: bare python in models/cehrbert/model.yaml:13 and models/random_predictor/model.yaml:15 (the latter has no venv at all — plain ambient env — and -ms into MEDS_DEV itself, so it only works if the PATH-first python happens to have MEDS-DEV installed).

4. -m runnability of MEDS-DEV's own six console scripts

All six entry points live in package __main__.py files with no guard and no top-level call — so python -m MEDS_DEV.datasets (etc.) currently succeeds silently having run nothing (exit 0, no output), the worst failure shape. All are @hydra.main apps, so guards are trivial and Hydra-safe. One wrinkle: results/__main__.py hosts two entry points (pack_result, validate_result), so a single guard is ambiguous — needs an argv dispatch or a module split.

Suggested outcome

A decision per bucket: (1) fix the dead activation line or delete it deliberately; (2) pick interpreter-pinned vs. activation-equivalent-PATH for same-env deps; (3) harden cross-env command resolution (absolute venv_bin_path?) with #99 as the regression case; (4) add guards + a dispatch story for results. Happy to follow up with PRs once directions are chosen.

Start with the audit at commit 60b678c, then inspect utils.py:284, 348, and 385-388, the task entry points in tasks/main.py, evaluation/main.py, and the package main.py files. Reproduce the silent python -m behavior and review the YAML commands in the two model files. Done means one documented decision and validated outcome for each of the four command-environment buckets, including the results dispatch case.