ActStress-Bench: Activity-Conditioned, Missing-Modality-Robust Wearable Stress Recognition -- Code and Experimental Results (Anonymous Review Artifact)

Anonymous double-blind review artifact for ActStress-Bench: an activity-conditioned, missing-modality-robust wearable stress recognition pipeline evaluated under cross-subject (Leave-One-Subject-Out) and cross-dataset (WESAD ↔ PhysioNet Wearable Stress and Structured Exercise) shifts. Contents: the full actstress Python package (dataset parsers, subject-safe splitting, classical and deep models, activity-conditioned fusion, calibration, ONNX/LiteRT export), its unit test suite, all experiment run directories (configs, seeds, per-window predictions, per-subject/per-fold metrics, aggregate metrics) backing every quantitative claim in the accompanying paper, on-device (phone and Wear OS watch) replay-benchmark logs, exported deployment models with desktop-vs-on-device parity reports, and the Wear OS benchmark app source (Kotlin). See README.md and EVIDENCE_MAP.md inside the archive for a claim-by-claim map to the exact supporting file. Raw WESAD and PhysioNet sensor data are third-party, academic-use-licensed releases and are not redistributed here; see the README for official sources and citation requirements. No clinical, battery-life, or live-sensing-accuracy claims are made. Code is released under MIT; results/configs/exported models under CC-BY-4.0 (see LICENSE files in the archive).