Deadline-scoped fan-outs

A deadline-scoped fan-out composes three bounds:

1. Per-branch: each remote call gets orTimeout — one slow dependency degrades to ITS fallback, not the whole request's.
2. Aggregate: allOf(...).orTimeout(total, ...) bounds the sum — branches can be individually fine while the composition (join order, queueing) still overruns.
3. Join with recovery: collect with handle per future, then join in submission order — a map of branch→result where every entry is present and every failure is a value ("timeout", "error"), never a thrown exception escaping the aggregator.

List<CompletableFuture<String>> branches = ids.stream()
    .map(id -> supplyAsync(() -> call(id), exec)
        .orTimeout(50, MILLISECONDS)
        .handle((r, e) -> e == null ? r : "down"))
    .toList();
allOf(branches).join();           // bounded aggregate
return branches.stream().map(CompletableFuture::join).toList();

The result: partial success is a structured outcome the caller can reason about, not a lottery between one caller's timeout policy and another's.

One JDK subtlety the aggregator must survive: a task's own exception reaches handle wrapped in CompletionException, while orTimeout's TimeoutException arrives raw. Robust code unwraps one level before classifying:

Throwable ex = (e instanceof CompletionException && e.getCause() != null)
    ? e.getCause() : e;