Add stress/sleep ML train harness and reproduce pack

[Newbbsss]

Summary

• Lean stress tip + sleep Rest/stage ML reproduce pack only: training Tools, golden labels/manifests, factor docs, and docs/STRESS_SLEEP_ML_REPRO.md.
• Lets someone recreate Gilbert tip@clock / Rest evaluation without Fold live ADB, UI polish, APKs, or the large MG handoff branch.
• Desk CPU train run recorded; accuracy_valid stays false until documented N gates (honest underfit with 1 tip@clock pair).

Reproduce stress + sleep

python Tools\stress_sleep_train.py --as-of 2026-07-14
python Tools\ml_engine_train.py --labels pairing-logs\whoop-app-labels.jsonl --as-of 2026-07-14
python Tools\sleep_stage_eval.py --synthetic

See docs/STRESS_SLEEP_ML_REPRO.md.

Included (lean)

• Tools/stress_sleep_train.py, ml_engine_train.py, sleep_stage_eval.py, calibrate_whoop_noop.py, ingest_export.ps1
• Docs: STRESS_SLEEP_ML_REPRO.md, screenshot-compare playbook, stress/sleep factor + dataset notes
• Golden: pairing-logs/whoop-app-labels.jsonl, export manifests, weights + status JSON

Expected results (this pack)

• Stress tip@clock: WHOOP 1.4 MEDIUM vs NOOP 1.8 @ 22:14 on 2026-07-13 (δ +0.4)
• Sleep: WHOOP Sleep 64% vs NOOP Rest 73 (Deep 2% / REM 0% — REST-WEIGHT / staging notes in factor docs)
• Runtime RestScorer / DaytimeStress / SleepStager live on main; this PR adds the train + label + repro layer on top (algo formula edits stay out to keep the diff lean)

Explicitly excluded

• UI Fable polish, APKs, store publish, Menstrudel, android-sdk, pairing-log dumps, screenshot JPG trees, PR feat: MG debug line — cycle/WHOOP compare, Sleep polish, calibration (full scrubbed handoff) #331 MG mega-handoff

Notes

• Fold was plugged in but powered OFF this session: no bank pull or live tip@clock verification.
• Supersedes the mistaken same-fork PR: Add stress/sleep ML train harness and reproduce pack Newbbsss/noop#1

AI assistance

Prepared with Cursor agent assistance.