bop O(N) — keep loops linear

When adding behavior to a per-card loop (dispatcher, merge-gate, reap, status), never introduce per-item syscalls. The loop body should do O(1) memory work per card; all I/O flushes once per iteration.

Approved patterns

1. Collect then flush

let mut events: Vec<Event> = Vec::new();

for card in cards {
    process(card);                   // existing work
    events.push(build_event(card));  // O(1) memory, no I/O
}

flush_events(&events);  // one write, one POST

2. Opt-in gate

let enabled = lineage::is_enabled();  // check once

for card in cards {
    process(card);
    if enabled {
        events.push(...);  // zero work when disabled
    }
}

3. Defer to background

let (tx, rx) = mpsc::channel();
std::thread::spawn(move || { for ev in rx { write(ev); } });

for card in cards {
    process(card);
    let _ = tx.send(build_event(card));  // non-blocking
}

Anti-patterns — reject these in review

Where this applies in bop

• run_dispatcher() — processes up to max_workers cards per iteration
• run_merge_gate() — processes all done/ cards per iteration
• reap_orphans() — scans all running/ cards
• print_state_group() — reads all cards in a state dir for bop status
• cmd_doctor() acceptance criteria lint — iterates pending/ cards

COW note

fs::rename is metadata-only on APFS/ext4/btrfs — no bytes copied, COW reflinks preserved. It's already O(1). Don't wrap it in anything that adds O(N) I/O.