Runtime performance work from the perf audit — Tiers A, B, C, all visual-safe. No pixels change (the one exception is a deliberate cost thousands-separator carried over from the prior PR's CHANGELOG entry). Build was already modern (Kotlin 2.3.20 strong-skipping, release minify/shrink), so these target hot paths.

Tier A — tiny change, big gain

• collectAsState → collectAsStateWithLifecycle app-wide (21 screens): stop recomposing while backgrounded/covered.
• Dashboard stops its 5s status poll when off screen (LifecycleStartEffect + resume/pause) — no off-screen network/CPU/battery.
• StatsRepository.getStats runs off Dispatchers.IO and fans its 3 independent groups out with async instead of ~50 sequential DAO queries on the main thread.
• contentType on the charges/drives/trips/mileage lists — row-node reuse during fling.
• ChargingNotificationWorker no longer refetches the full car list once per car.

Tier B — per-frame / allocation wins

• Hoisted chart Paints in ChartDrawUtils (built once via remember, not per draw) — the 800ms entrance animation redraws ~48 frames, so this removes thousands of transient Paint/Color/String allocations. Same pixels.
• PowerSocOverlayChart pre-sorts each curve's points once instead of re-sorting inside the draw loop and twice per crosshair-drag frame.
• Mileage memoizes its chart-data derivation + BarChartData mapping; BarChartData/BarSegment marked @Immutable.

Tier C — data layer

• Comparison repos (findComparable) move the full-history load + per-row Haversine + sort to Dispatchers.Default (off the main thread).
• TripRepository.getTrips runs off-main, and the beta-era SHA-256 duplicate cleanup now runs once per car per session, not every open.
• SyncRepository.syncChargeDetails preloads summaries once instead of a per-charge DB read in the batch loop.
• RouteSimplifier rewritten iteratively (keep[] bitset + explicit stack) — same output, no per-level list allocation, bounded stack.

Deliberately skipped / deferred (engineering judgment, flagged not silently dropped)

• Widget progress-bar bitmap cache (Tier B): each 3-min update carries a new battery level, so a cache almost never hits — churn without gain.
• AppSettings cache (Tier B): negligible gain over the already in-memory-cached DataStore, and a stale server URL/token would silently break API calls.
• Comparison bounding-box SQL prefilter (Tier C): the off-main move removes the jank; a too-tight SQL box could silently drop valid matches, so it needs a new indexed query + margin care — deferred rather than risk correctness.
• Baseline Profile (the biggest cold-start lever): not in this PR — it needs a :baselineprofile macrobenchmark module (half-day of scaffolding). Happy to do it as a follow-up. (F-Droid note: the generated baseline-prof.txt ships fine as a source file; only regeneration is local, not in F-Droid CI.)

Verification

• ./gradlew testDebugUnitTest lintDebug assembleDebug — all green.
• Installed on the test device.

🤖 Generated with Claude Code