Correlates a network's reply (or terminal error) back to the caller that injected the request.

bloccs is otherwise fire-and-forget: Bloccs.Producer.push/3 returns once a message is admitted, not once it is processed. The collector restores request/response on top of that — without making the pipeline synchronous — by reusing the per-message trace_id (Bloccs.Lineage) as the correlation key:

1. Bloccs.call/4 mints a root lineage (trace_id T), registers T here (so the collector is expecting it before the request can fail or reply), then pushes the request.
2. The request flows through the graph; trace_id rides unchanged on every 1:1 / split / router hop.
3. A node declared reply = true reports its emitted payload here (an {:ok, payload} result); a node that fails terminally on the trace reports an {:error, %Bloccs.EffectError{}} result. Both are keyed by T.
4. The collector matches T and hands the result to the blocked caller.

A single process serves every running network; entries are keyed by {network_id, trace_id} so ids never collide across networks.

Register-before-push (no races, no fire-and-forget overhead)

call/4 / cast/4 register synchronously before pushing, so by the time the message can produce a reply or error, the collector is already expecting it. A result for a key that is not registered is dropped immediately (not buffered) — so a plain Bloccs.Producer.push/3 or a reply = true node used in a fire-and-forget pipeline costs the collector nothing beyond an O(1) drop. A result that arrives before the caller blocks on await/2 is held (:ready) until claimed. The timeout is enforced collector-side, so a late result can never crash a timed-out caller.

Limitations (current milestone)

• Aggregation is first-wins: the first result reported for a trace_id is the answer; later ones for the same id are dropped. Fan-out networks with several reply emits per request need an :all policy (later milestone).
• A [batch]/[join] on the request path mints a fresh trace_id, breaking correlation — including error correlation. Keep the reply node on the request's 1:1 trace.