docs(audits): ReorderBuffer cache-behavior profiling plan at 1M files

[oferchen]

Summary

• Benchmark plan for cachegrind/perf profiling of BoundedReorderBuffer in crates/transfer/src/reorder_buffer.rs at 1M files.
• Documents BTreeMap node layout (B-tree with up to 11 KV per leaf, pointer-chasing per insert/lookup), references the ring-buffer alternative explored in Improve CLI binary lookup for tests #1853, defines three delivery workloads (in-order, mostly-in-order with 5% gap, pathological reversed-window), and pins L1d/LLC miss rate plus IPC as the decision metrics.
• Sets explicit migration criteria: ring wins only if mostly-in-order workload at window 64 hits >= 1.5x IPC or >= 2x lower L1d miss rate with end-to-end cycle ratio <= 0.7, and pathological workload does not regress beyond 10%.

Tracking task: oc-rsync #1854.

Test plan

• Verify doc renders cleanly on GitHub.
• Confirm the cited line ranges in crates/transfer/src/reorder_buffer.rs (struct 57-64, insert 129-146, drain 149-156) still match after any concurrent edits before merge.
• No source-code changes; CI fmt/clippy/nextest unaffected.