Function dispatch_event 

pub fn dispatch_event(
    conn: &Connection,
    event: &str,
    memory_id: &str,
    namespace: &str,
    agent_id: Option<&str>,
    db_path: &Path,
)

Fire an event to all matching subscribers. Each dispatch is fire-and-forget — the caller is NOT blocked on delivery. Errors are logged and counted in the DB via failure_count.

PERF-3 (FX-10, 2026-05-26): pre-fix posture was std::thread::spawn + per-worker Connection::open(...) per matching subscriber — 1000 subscribers minted 1000 OS threads + 1000 SQLite handles per store event. Post-fix posture uses a bounded tokio::sync::Semaphore (default 32 permits, tunable via AI_MEMORY_WEBHOOK_DISPATCH_CONCURRENCY) + tokio::task::spawn_blocking against the existing daemon Tokio runtime so the in-flight delivery count is capped. The pre-fix std::thread::spawn path is preserved as a fallback for legacy unit tests that run outside any runtime.

Caller owns the connection. Dispatch workers still re-open the connection per delivery to update counters (cheap — SQLite connections are process-shared via WAL — and the new concurrency bound keeps the simultaneous handle count bounded too).

P5 (G9): convenience wrapper for the historical no-details case (used by memory_store). New event types should call dispatch_event_with_details and pass the matching *EventDetails struct serialised to JSON.