Function step_scratch_store_observed 

pub fn step_scratch_store_observed<M, S, O>(
    model: &M,
    store: &mut S,
    walls: &[WallSegment],
    params: &M::Params,
    dt: f64,
    scratch: &mut Scratch,
    peds_buf: &mut Vec<Pedestrian>,
    observer: &mut O,
)
where
    M: CrowdStep,
    S: AgentStore<CrowdAgent>,
    O: CrowdObserver + ?Sized,

Observed variant of step_scratch_store: identical semantics, plus a post-writeback callback for every agent.

observer.observe(id, &ped) is invoked once per agent, in the same order as store.iter_ids(), with ped holding the post-tick state that was just written back. The observer sees the same state the next AgentStore::get call would return.

This is the production telemetry entry point: a closure that forwards into rustsim::TelemetryPipeline::push_row (or any other sink — CSV, Parquet, in-memory buffer, Prometheus counter) gets per-agent per-tick coverage with zero allocation on the hot path beyond what the sink itself may do.

Panic safety: if observer.observe panics the writeback loop unwinds and the store is left in a partially-updated state (rows before the panicking index are already committed; rows after it have not yet been observed but have been written back). Callers that cannot tolerate that should make observer.observe infallible or forward its errors via captured state.