Arduino millis() rollover-safe timing patterns without core helpers

Context

The Arduino millis() function returns a 32-bit unsigned counter that wraps to zero after 4,294,967,295 milliseconds (~49.7 days). The Arduino reference documents this rollover but the core library (through at least ArduinoCore-API 1.0.x) does not ship a standard rollover-safe elapsed-time utility such as elapsedMillis (available in Teensyduino and some third-party cores).

Goal

Design a reusable timing helper that works correctly across the rollover boundary for interval checks, timeouts, and scheduled tasks, using only the standard Arduino API.

Constraints and uncertainty

• Must compile on AVR (Uno, Nano), SAMD (Zero, MKR), and ESP32 cores without core-specific extensions.
• Must handle intervals up to several hours (well under the 49.7-day wrap period).
• Must not rely on signed arithmetic tricks that invoke undefined behavior on some toolchains.
• Third-party libraries in the same project may implement their own timing logic inconsistently.

Questions

1. What is the minimal, standards-compliant C++ pattern for if (elapsed_since(start) >= interval) that is guaranteed correct across the rollover on all supported architectures?
2. Should the helper be a header-only template, a macro, or a small class to avoid code bloat and ensure inlining?
3. Are there any documented cases in the Arduino core source (e.g., wiring.c timer0 ISR) where the rollover behavior differs between architectures that would affect the helper's correctness?