docs: the handoff and the check task say what they do

AGENTS.md gains the continuation handoff so a fresh session inherits the real
baseline instead of re-deriving it, README drops a paragraph that restated its
own previous sentence, and `mise run check` runs the gates itself rather than
depending on three tasks whose ordering it could not guarantee.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: malys

AGENTS.md

@@ -50,3 +50,19 @@ the complementary `emulator-car` (Automotive API 33) and `emulator-screen` (API
 1920×720) profiles; `mise run run` installs and launches the debug APK. JDK 17 and Android
 SDK 36 are used for compilation; the deployed minimum remains API 28. Emulator checks do
 not replace a vehicle pass on each supported firmware generation.
+
+## Continuation handoff
+
+- The implemented baseline is the passive dashboard, nullable shared energy snapshot,
+  parked-only process-local trip recording, bounded atomic trip history, crash diagnostics,
+  stable/unstable variants, CI gates and the two emulator profiles documented in `README.md`.
+- Arrival-SOC prediction, charger routing, background trip recording and energy attribution
+  are not implemented. Treat each as a separately scoped milestone with explicit evidence
+  for every input; do not fill missing vehicle signals with inferred measurements.
+- Start app work by checking `git status` and the EVHardware submodule branch. Any shared
+  telemetry model, property, unit, fallback or integration change starts in the standalone
+  `../EVHardware` repository, is tested and pushed there, then reaches this app through a
+  submodule pointer update. Never develop library changes inside the nested checkout.
+- Before handing off a change, run `mise run check` and `mise run build`. For UI/lifecycle
+  work, also run the API 28 screen emulator at 1920×720; for vehicle interfaces, record the
+  remaining on-vehicle validation requirement rather than claiming emulator coverage.

README.md

@@ -67,10 +67,8 @@ Recording currently samples only while the dashboard is visible and survives act
 recreation only while the process remains alive. Persistent background recording is a later,
 separately reviewed milestone. A trip's reported duration is therefore the time actually
 covered by usable samples, not wall clock: hiding the dashboard for ten minutes adds nothing
-to the duration, the distance or the energy, so the averages divide comparable numbers. Because of that, a trip reports the time its samples actually
-cover, not wall clock: hide the dashboard for ten minutes and the recorded time does not
-move, so the average consumption stays the quotient of a distance and an energy that were
-measured over the same interval.
+to the duration, distance or energy, so consumption averages compare values measured over
+the same interval.
 
 ## Install
 

mise.toml

@@ -61,7 +61,14 @@ run = "bash .github/security/check-permissions.sh"
 
 [tasks.check]
 description = "What CI runs: permissions + lint + tests"
-depends = ["permissions", "lint", "test"]
+run = '''
+set -e
+bash .github/security/check-permissions.sh
+./gradlew -Dorg.gradle.java.home=$JAVA_HOME \
+  :evhardware:testDebugUnitTest \
+  testStableDebugUnitTest testUnstableDebugUnitTest \
+  lintStableDebug lintUnstableDebug
+'''
 
 [tasks.release]
 description = "Build the minified stable release APK"