fast-sim spends 60–70% of its wall time in RKF45 continuous integration because Generator hardcodes dtMax=1e-3, capping every step at 1 ms (~1.85M derivative evals/run). That tiny dtMax is load-bearing: block-boundary crossings are detected purely by step granularity, so raising it breaks the simulation. This converts the two train-movement event detections to kDisco's new waitCrossing { guard } root-finding — step 1 of the issue's proposed direction.

Depends on kDisco PR #68 (waitCrossing); interlockSim CI cannot compile until that is merged and published.

Changes (sim/Train.kt)

• Block-boundary crossing (Site.actions): waitUntil { position.state + dtMin >= nextLength } → root-found state event. The dtMin slack is retained to avoid an asymptotic-stop hang when a train decelerates to a STOP semaphore sitting on the boundary and settles ~1e-8 short.
• Tail-entry crossing (Train.actions): waitUntil { front.totalDistance >= length } → root-found state event with the same dtMin slack.
• Re-added import cz.ksimulantenbande.kdisco.dtMin.

// before
waitUntil { position.state + dtMin >= nextLength }

// after — crossing time located by root-finding, independent of step size
val boundaryThreshold = nextLength - dtMin
if (position.state < boundaryThreshold) {
    waitCrossing { boundaryThreshold - position.state }
}

Generator.kt

• dtMax left at 1e-3. See below.

Not yet addressed — dtMax stays at 1e-3

Raising dtMax still collapses the run (5 → 2 trains), so the perf win is deferred. Two root causes were isolated; both need a Motor redesign, golden re-baselining, and expert sign-off under the conservative sim/ rules:

• Asymptotic Motor stop — deceleration a = -v²/2s stops via a velocity-clamp, undershooting the boundary by ~O(dtMax²) (1.1e-8 @ 1e-3, 2.2e-5 @ 1e-2), so a fixed tolerance cannot absorb it as dtMax grows. Fix direction: waitCrossing { distanceToSemaphore() } to stop exactly at s=0.
• Front/tail velocity-clamp drift — Motor.derivatives() mutates the shared velocity.state in place; front and tail then integrate against different velocities, drifting |front − tail − length| to 0.0246 @ dtMax=1.0 (> maxAbsError=1e-2, LengthChecker FATAL). Fix direction: an order-independent clamp.

Validation notes

• Behavior is identical to the golden baseline at dtMax=1e-3 (5 trains, 75 TRAIN_EVENTS lines, 213.1 s sim) — the conversion introduces no drift.
• Scope is limited to Train.kt and Generator.kt; Motor, velocity-clamp, and LengthChecker are unchanged from the original.
• At dtMax=1e-3 this is groundwork with no runtime speedup on its own; the win lands once the Motor state-event work above is complete. :core:heavyTest and re-baselining still pending.