feat: bulk_delete_data on Storage

[sejori]

Changes:

• src/daemon/mod.rs — Added email and completion_window labels to per-user Prometheus metrics (fusillade_user_requests_in_flight gauge, fusillade_user_requests_completed_total counter). Email falls back to user_id if not present in batch metadata.
• src/manager/mod.rs — Added bulk_delete_data(creator_id, batch_size) to the Storage trait for chunked soft-deletion of a creator's batches and files with metadata nullification.
• src/manager/postgres.rs — Implemented bulk_delete_data with two-stage chunked UPDATE using FOR UPDATE SKIP LOCKED: soft-deletes batches (nullifies metadata, cancels active ones) then soft-deletes files. Existing purge daemon handles child row cleanup.

[Copilot]

Pull request overview

Adds completion_window segmentation to per-user Prometheus metrics so dashboards can compare user throughput across SLA windows (e.g., 1h vs 24h).

Changes:

• Extract completion_window from batch_metadata when spawning request processing tasks.
• Add completion_window label to fusillade_user_requests_in_flight gauge (inc/dec) and fusillade_user_requests_completed_total counter (success/failed/cancelled paths).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

completion_window is pulled from request.data.batch_metadata, which is populated based on DaemonConfig.batch_metadata_fields (and can be empty, as in several tests). In that configuration this label becomes an empty string via unwrap_or_default(), so the new dashboard split-by-completion_window won’t work and metrics will silently aggregate under a blank value. Consider sourcing completion_window from an always-present field (e.g., carry batch_completion_window on RequestData when claiming) or at least emit a distinct fallback like "unknown" and log/warn when the metadata key is missing.

Suggested change

	let completion_window = request.data.batch_metadata.get("completion_window").cloned().unwrap_or_default();
	let completion_window = match request.data.batch_metadata.get("completion_window") {
	Some(value) if !value.is_empty() => value.clone(),
	_ => {
	tracing::warn!(
	request_id = %request_id,
	batch_id = %batch_id,
	"Missing `completion_window` in batch_metadata; using \"unknown\" label for metrics"
	);
	"unknown".to_string()
	}
	};

[Copilot]

Adding raw completion_window as a label to high-volume per-user metrics (fusillade_user_requests_in_flight / fusillade_user_requests_completed_total) risks very high label cardinality because completion_window is user-provided and not normalized (e.g., 60m, 1h, 1h0m, etc. all become distinct series). This can significantly increase Prometheus/OTel metric storage and query cost. Consider normalizing to a canonical representation (e.g., seconds or a fixed set of buckets) before using it as a label, or restricting allowed values.

Suggested change

	gauge!("fusillade_user_requests_in_flight", "user" => user_id.clone(), "completion_window" => completion_window.clone())
	gauge!("fusillade_user_requests_in_flight", "user" => user_id.clone())