Sampling without draining: a wearable's power budget

The wearable I'm building is a smartwatch-shaped safety device for known spaces — corridors, work floors — built on an ESP32 with a GSR sensor, a MAX heart-rate module, and SpO₂ monitoring. Its whole reason to exist is vigilance: it has to notice when a human being is in trouble. Vigilance, it turns out, is a power budget problem.

Sample everything continuously and the battery dies in hours — a safety device that's dead by afternoon is worse than no device, because someone trusted it. Sample too lazily and you miss the event you exist to catch. Every sensor has a different honest minimum: heart rate needs enough samples to catch an anomaly's shape, GSR drifts and needs baseline tracking, SpO₂ is expensive to read but changes slowly. The ESP32's sleep states help, but every wake-sample-sleep cycle has its own overhead tax, and the radio is a spendthrift on top.

I don't have a good answer yet. The current direction is tiered vigilance — cheap sensors gating expensive ones, so the device spends power in proportion to suspicion — but I haven't proven the tiers catch what they must never miss. This entry is the problem statement; the answer, when I earn it, gets its own.