feat: density-based activity classifier and data import/export

[JanJetze]

Summary

• Density-based walking detection: replaces the fragile 36-consecutive-window requirement with a rolling 5-minute density check (5+ windows with hz ≥ 1.5). Tolerates accelerometer throttling gaps that previously prevented walking detection entirely.
• Sticky cycling exit: requires hz ≥ 2.0 for 8 consecutive windows to exit CYCLING, filtering false step events from road/bike vibration (0.4–1.8 Hz).
• 30-second evaluation interval: aligned with the step frequency tracker's 30s sliding window. Each reading is now independent (was 83% overlapping at 5s). Reduces battery usage ~6x.
• DB migration v2→v3: consolidates existing 5s sensor windows into 30s by keeping highest-variance window per slot.
• JSON data export/import: complete export including sensor windows, with import for data migration between app variants.

Validated against real sensor data

Activity	Expected	Detected
Walk (Mar 5)	~12:40–13:00	12:43–13:00 (17 min)
Bike ride (Mar 5)	~15:45–16:45	16:00–16:59 (59 min)
Chores/indoor	No detection	None (zero false positives)

Test plan

• All 748 unit tests pass
• Debug APK builds successfully
• Installed on device, migration ran, activities display correctly
• Verify walking detection on a new real walk
• Verify cycling detection on a new real ride
• Verify no false walking from indoor chores over a full day

🤖 Generated with Claude Code