[Medium] Trace accumulation is O(steps^2) on the generic and stale-async execution paths

[gratus907]

Severity: Medium (performance)

Trace event accumulation is O(steps²) on the generic sync and stale-async execution paths, because each step copies the entire event tuple.

Locations — src/variopt/study/execution.py:646, src/variopt/study/stale_async.py:394

trace = trace.append(event)  # Trace.append copies the full tuple each call

Scenario

A 10k-step, batch_size=1 run performs ~5·10^7 tuple-element copies purely for tracing. Profiling of cheap objectives on the generic path shows trace append dominating wall time.

Notably, the recently added cheap-scalar-sequential fast path (execution.py:217,301) already avoids this by buffering trace_events in a plain list and constructing the Trace once at report assembly — the general and stale-async paths still pay the quadratic cost this fast path was specifically built to avoid.

Fix direction

Accumulate TraceEvents in a list during the run loop and construct Trace once at report-assembly time, as the fast path already does.