dd7075df79
Verified via benchmark: 0.076µs per call (list append), 0.0044% of RTT. The load computation (sum/variance/σ) is deferred to _compute_load() which is ONLY called from log messages — never in make_request or the retry loop. - _track_rtt() now returns None, appends only (was: appended + computed load) - _compute_load() extracted from _track_rtt for lazy load estimation - _current_load() delegates to _compute_load() - Removed round() from hot path — raw float append, rounding deferred - Removed round() from make_request _track_rtt call - Tests: 79 passed (added TestHotPathZeroOverhead: 4 tests)