A test harness that steps only the simulation is a different system from the one the frame loop runs

Fact

Driving a real-time app from a script usually means calling its update(dt) in a loop and rendering once at the end, because rendering is the expensive part. That is only equivalent if nothing integrates inside the render path — and in practice plenty does: camera smoothing, interpolated transforms, LOD selection, anything written as current += (target - current) * rate.

Batch 200 updates behind one render and such a value advances one step, not 200. The simulation is correct, the picture is of a state the app would never show, and the two disagree by exactly the amount that was skipped.

Why it matters

It presents as a bug in whatever lives in the render path. A camera that has integrated once looks like a camera that will not converge — so the search goes into the follow code, the lerp rate and the target computation, all of which are fine. Screenshots taken from the harness then keep confirming it.

The same shape catches anything else the frame does outside update: input sampling, overlay sync, visibility toggles driven per frame.

How to apply

• Have the harness call the app's own frame function, the whole one, N times. If rendering is too slow for that, factor the frame into advance(dt) (sim + every per-frame integrator) and draw(), and drive advance N times.
• The giveaway is a value that converges in the running app and stalls under the harness. Compare it after n harness steps against n real frames before blaming the value's own code.
• Driving the whole frame N times still does not move anything interpolated on the wall clock, such as remote entities drawn now - delay between network snapshots: the clock barely advances inside one synchronous batch. Moved by the test, such an entity stays drawn where it was, and a tap aimed at its drawn position misses. Wait in real time until the drawn position matches the authoritative one before reading the screen.
• The timestamp half of this is [[driven-game-loop-needs-accumulator-reset]].