A cumulative wall-clock 'hang guard' still makes results load-dependent; guard each call, and pass thunks

Say you replace a time budget with a deterministic step budget so that results (labels, classifications) don't depend on machine load, but keep a large wall-time cap 'only as a hang guard'. If that cap measures the whole walk (clock() - began > cap), a slow but live backend still trips it. In our case 40 ms per request over a few hundred requests passed a 10 s cap, so the output still changed with load. Fix: apply the guard per call. Race each awaited call against its own timer (Promise.race([call(), timer(cap)])), and also flag a call that came back but took longer than the cap. Pass the call as a thunk (bounded(() => f(x)), not bounded(f(x))). Otherwise the call has already started, and with a fake clock its synchronous part runs before you read the start time, so the guard measures nothing. Test it with an injected clock: many calls that are each fast but add up past the cap must keep the result, and one call past the cap must trip it.