Known per-item query fan-outs

This is the current list of confirmed database round-trip loops. It is not an audit journal: resolved entries and rejected findings belong in version history. A benchmark query budget is a tripwire against regression, while this file records debt that the benchmark dataset may not expose.

When fixing an entry, delete it in the same change and record the measured before/after statement count. A lower benchmark budget needs an explanation too; dataset changes can lower counts without improving code.

Entry	Severity	Bound	Why it remains
`migration-process-move-per-local-blocker`	high	unbounded local blockers	Each carried relationship has distinct federation and moderation effects.
`migration-process-move-per-local-follower`	high	unbounded local followers	Each relationship produces distinct follow/unfollow work.
`ingest-store-remote-attachments`	medium	100 per Note	Inbound attachment rows are inserted individually.
`inbox-flag-report-per-target`	low	50 targets	Each local target receives a distinct report; instance reports may also enqueue webhooks.
`ingest-thread-ancestor-backfill`	low	5 ancestors	Each missing ancestor requires guarded network fetch and full ingest.

Details

Move replay

process_move in crates/server/src/migration.rs loops over local followers and local accounts blocking the source. Both loops are unbounded and run from the inbound Move request path. Each item performs relationship checks, mutations, notifications, and/or distinct ActivityPub delivery. No benchmark currently exercises this path with a SQL query budget.

Remote attachments

store_remote_attachments in crates/server/src/ingest.rs calls media::create_remote per attachment. Input is capped at MAX_REMOTE_ATTACHMENTS = 100, so this is bounded but a hostile Note can still cause 100 serial inserts. The inline-image fallback is deduplicated separately.

Flag targets

The instance and group handle_flag paths in crates/server/src/routes/inbox.rs file one report per deduplicated local target. MAX_FLAG_OBJECTS caps the set at 50. Instance-level reports also run the configured report webhook path; group-scoped reports do not.

Thread ancestor backfill

resolve_thread_parent in crates/server/src/ingest.rs walks and then ingests up to MAX_THREAD_ANCESTORS = 5. Each step includes a guarded remote fetch and a complete remote-note ingest. This is fixed-bound rather than a traditional unbounded N+1, but its body is expensive enough to track here.

Single SQL statements containing correlated subqueries are query-plan costs, not extra round trips, and do not belong in this inventory. Investigate those with EXPLAIN and performance benchmarks instead.