The orchestrator already has a performance-tuning stage (Stage 5), but in practice its gains are limited and inconsistent: the same tuning lessons are relearned from scratch on every operator, accepted "improvements" occasionally do not hold up, results lack reviewable measurement evidence, and there is no clear ending when no gain exists.
Systematically strengthen Stage 5 into a tuning stage that can be trusted: tuning knowledge persisted in-repo, accumulating across operators and reusable; only verified, genuine gains are accepted for delivery; when no gain exists, the original kernel is delivered honestly with a record of what was tried; the operator's public contract and precision are never compromised.
现状 / Current situation
pypto-pro-op-orchestrator 已有性能调优阶段（Stage 5），但实际使用中收益有限且不稳定：同样的调优教训每个算子都要重新踩一遍，经验不沉淀；接受的"优化"偶尔并不成立，事后也缺少可复核的测量证据；没有收益时的收尾也不明确。
The orchestrator already has a performance-tuning stage (Stage 5), but in practice its gains are limited and inconsistent: the same tuning lessons are relearned from scratch on every operator, accepted "improvements" occasionally do not hold up, results lack reviewable measurement evidence, and there is no clear ending when no gain exists.
诉求 / Request
系统性增强 Stage 5，使其成为可信赖的调优阶段：调优知识沉淀在仓库中并随算子积累、可复用；只有经过验证的真实收益才被接受交付；确实没有收益时如实交付原 kernel 并说明已尝试过什么；全程不改变算子对外契约、不引入精度回退。
Systematically strengthen Stage 5 into a tuning stage that can be trusted: tuning knowledge persisted in-repo, accumulating across operators and reusable; only verified, genuine gains are accepted for delivery; when no gain exists, the original kernel is delivered honestly with a record of what was tried; the operator's public contract and precision are never compromised.
验收 / Acceptance
代表性算子上相对现有 Stage 5 有可测量的性能提升（如 cann-bench 任务集分数），全部正确性用例保持通过。
调优知识随仓库沉淀，后续算子可直接复用。
每个交付附可追溯、可复核的测量证据，包括未采纳候选的处置记录。
Measurable gains over the current Stage 5 on representative operators (e.g., the cann-bench task set), with all correctness cases still passing.
Tuning knowledge persisted in-repo and directly reusable by later operators.
Every delivery carries traceable, reviewable measurement evidence, including the disposition of rejected candidates.