Parametrize DEVICE in issue658_fit_predictors.py (CPU pin hit #763/#812)

Overview / Motivation

Companion code fix from the 2026-07-02 serial-compute transcript audit (user-approved: "Make all seven changes"). The workflow-side rule change (artifact-reuse throughput item (i)) is filed separately; this task fixes the concrete offender at the source.

Goal

Parametrize the hardcoded DEVICE = "cpu" module constant in scripts/issue658_fit_predictors.py (line ~143) so downstream reusers get cuda-when-available (or an explicit env/CLI override) instead of silently inheriting a CPU pin.

Problem

scripts/issue658_fit_predictors.py:143 pins DEVICE = "cpu" at module scope. Two downstream tasks imported/forked this module and inherited the pin: #763's fit ran vectorized-but-CPU (~67 min/behavior; the EPM_FIT_DEVICE=cuda cutover landed the whole fit in ~2h vs ~24h serial) and #812's pooling fit projected ~17h on an e2-standard-8 for the same reason. The batched helpers themselves (vectorized_mlp_skill.py) are device-parametrized; the pin lives in this caller module and propagates via reuse.

Proposed change

• Replace the module constant with a resolved default: EPM_FIT_DEVICE env var if set, else "cuda" if torch.cuda.is_available() else "cpu", exposed as a function/CLI flag so shard launchers can pin per-process.
• Audit direct importers/forks (grep -rn 'issue658_fit_predictors' scripts/ src/ — at least the #763/#812 forks) and thread the parameter through rather than re-pinning.
• Keep behavior reproducible: log the resolved device in the run metadata.
• Smoke: one small fit on CPU and (if available) GPU produces matching outputs within the existing equivalence tolerance.

Constraints

• No behavior change to the fitted numbers (device only changes placement); ruff passes; existing tests pass.
• This is analysis/experiment-shared code — normal kind: infra code-change pipeline (implementer + code-review + test-verdict), no clean-result.

Provenance

Origin: 2026-07-02 transcript audit — incidents #763 (facet 2: "correctly VECTORIZED but device-pinned to CPU: it imports the hardcoded DEVICE = 'cpu' from issue658_fit_predictors.py:143") and #812 ("the same CPU-pinned-batched-fit shape #763 just escaped").