perf: investigate prewarmed latency variance and backend-dominant generation cost

[PowerBeef]

Performance follow-up should isolate the remaining latency variance and backend-dominant cost without blaming the wrong subsystem.

Source report: #22

Problem

Prewarmed latency remains noisy enough to obscure regression interpretation, while benchmark evidence still shows backend generation dominating wall time.

Evidence

• April 18, 2026 validation report: Dev validation report: April 18, 2026 #22
• clone_regression.json classified the slowdown source as overlay_refactor
• current_runtime_current_helper_total_s = 30.988
• current_runtime_old_helper_total_s = 26.3212
• perf_results.json identified total_backend, generation, and collect_generation as the top bottlenecks
• pro_custom_prewarmed_samples.json recorded wall-time CV 0.6238

Current ownership

• scripts/harness_lib/bench_runner.py
• Sources/QwenVoiceNativeRuntime/NativeMLXMacEngine.swift
• Sources/QwenVoiceNative/XPCNativeEngineClient.swift

Acceptance

The team can either show a stable reproduced explanation for the remaining variance or a measurable improvement in the smallest relevant benchmark category touched by the fix.

Focused rerun

• python3 scripts/harness.py validate
• python3 scripts/harness.py bench --category latency --runs 3 --output-dir <dir>
• python3 scripts/harness.py bench --category clone_regression --runs 3 --output-dir <dir> only if helper-overlay work is touched