Adopt per-thread high-water-mark CAS in URB receive loop — skip cross-thread CAS on ~99% of packets

[stanvx]

Parent

Architecture review (2026-06-21), "Latency tricks worth lifting wholesale". Lifted from Satellite's net/receiver.cpp:99-103. Independent of the deepening series.

What to build

In usbip-server's URB receive loop, the per-URB latency sample is currently posted to a shared atomic. The cross-thread CAS contention on a 1 kHz+ URB stream is measurable. Lift Satellite's per-thread high-water-mark pattern:

• Each worker thread keeps a thread_local HighWaterMark { local_peak_ns: u64, last_flushed_at: Instant }
• On each URB, the worker compares its local latency_ns to local_peak_ns using a relaxed load (no cross-thread CAS)
• Only when latency_ns > local_peak_ns does the worker perform the cross-thread atomic CAS on the global high-water-mark
• The CAS frequency drops from "every URB" to "only when a new local peak is set" — empirically ~1% of packets in steady state

This is a small, self-contained optimization with measurable latency-tail improvement under load. It does not change wire protocol, error handling, or test surface beyond one new benchmark.

Acceptance criteria

• HighWaterMark is a new private struct in usbip-server/src/urb_loop.rs (or wherever the receive loop lives)
• Each URB receive updates the thread-local high-water-mark with a relaxed load, then conditionally CASes the global atomic
• A new benches/urb_high_water_mark.rs benchmark compares the cross-thread CAS rate before and after (criterion-based, default sample size)
• cargo test --workspace --release passes; cargo clippy --workspace -- -D warnings clean
• No public API change; the LatencySample shape emitted on /api/events is unchanged
• Brief KDoc on the new struct explains the relaxed-load + conditional-CAS pattern with a one-line citation of Satellite's receiver

Blocked by

None — can start immediately.