Compact per-batch engagement blobs for a date into a single file.

Fixes the aggregator 30s-timeout from 2026-06-12 — at ~4000 batches/day, enumerating + reading the per-batch dir blew past Cloud Run's request budget. The compacted single file is one O(1) GCS read.

Defaults to yesterday (UTC). Pass ?date=YYYY-MM-DD to target a specific date — useful for backfilling broken days.

Idempotent — running again on an already-compacted day folds in any late-arriving per-batch blobs and re-emits the single file. Safe to schedule nightly via Cloud Scheduler.

Gated like /admin/aggregate: admin token, or a Cloud Scheduler OIDC identity token (src/serve/scheduler_auth.py, PD-388).