CURRENT_NOW unit auto-calibration: show the real current on mA-unit devices (#152 v2)

[almothafar]

Stacked on #169 (v1 floors) — review/merge that first; this PR's diff is just the calibration commit.

v1 stops the Kirin garbage from being displayed; this makes the readings correct there instead of hidden.

How it works

New CurrentUnitCalibrator, hooked into the single read point SystemService.getInstantaneousCurrentMicroAmps, so the rate tracker, the Current row, and ChargeSpeed all see the same corrected value:

• Persistently tracks the maximum |raw| CURRENT_NOW value ever observed (across sessions, charging and discharging alike).
• A genuine-µA device exceeds 20,000 (20 mA) within moments of awake use — screen-on draw alone is 100,000+. A mA-unit device physically can't reach it (that would be a sustained 20+ A).
• After 60 readings spaced ≥ 1 min apart (≥ 1 h cumulative observation) with the max still below the floor → conclude mA units → scale readings ×1000.

Safety properties:

• Observation records the raw value (pre-scaling), so scaled outputs can't feed back into the decision — no oscillation.
• A later large reading (e.g. firmware fixes the unit) voids the conclusion in the same call, immediately and permanently. That ordering also caps any scaled value at < 20,000,000 µA — no int overflow.
• Preference churn is bounded: the max plateaus in minutes, counting stops for good at the cap; the steady-state path writes nothing.
• Sentinels (Integer.MIN_VALUE/MAX_VALUE) and blank 0 pass through untouched.

Side benefit: once calibrated, source A of the %/h rate (current ÷ capacity) works automatically with scaled current on any such device where capacity is known.

Known transient: at the moment the conclusion first flips on, the rate window still holds pre-scaling samples for up to 10 minutes; on Kirin source A is off anyway (capacity 0), so this is harmless and self-heals.

Tests

Pure helpers fully covered (observe fold: max tracking, spacing throttle, count cap, unchanged-instance contract; isMilliAmpUnits boundaries at the floor and the sample cap; scaling incl. sign) plus Robolectric passthrough tests for the entry point. testDebugUnitTest + lintDebug pass.

On-device verification needed before merging (Mate 10 Pro)

• Confirm raw magnitudes match expectation (~200–2000 discharging, ~1000–2000 charging) — e.g. temporarily log getIntProperty(CURRENT_NOW)
• After ~1 h of use, the Current row/notification shows realistic values (hundreds of mA, not ±1)
• Charging tier + wattage look sane while charging; no false slow-charge warning

Closes #152

🤖 Generated with Claude Code