V1.4.2 — Readiness measurement closure

V1.4.2

V1.4.2 closes the existing V1.4 readiness and measurement contracts and makes system-level optimization decisions more explicit.

Changes

• New optimization Targets require combined readiness with a two-sample smoke.
• Readiness and baseline exactly validate metric names, units, constraint sets, and requested sample counts.
• Command and measurement failures preserve actionable codes, field differences, bounded output, return codes, and cleanup status.
• Profile evidence is checked against Target, Variant, request slice, phase, coverage, system attribution, and the idealized end-to-end ceiling before Candidate and formal-target decisions.
• Shared-host performance evidence requires continuous, time-aligned resource observation; missing evidence makes performance inconclusive without discarding independent correctness facts.
• Added live-workload cost disclosure, bounded subagent use, and complete Handoff requirements.

Frozen legacy separate Targets remain readable by the evaluator, but new optimization readiness is combined-only until a versioned multi-probe evidence contract exists.

Validation

• Focused conformance: Candidate 7/7 passed; Original failed 6 of the 7 cases.
• Full suite: 242 passed, 2 skipped on Python 3.9.6.
• GitHub CI passed on Python 3.10 and 3.12.
• Four clean behavioral trials confirmed safe decisions; the Candidate made applicability, resource, and Handoff closure explicit.

This release makes no GPU performance or production-workload speed claim and contains no private workload, trace, host address, model path, or internal data.

V1.4.1 — Evidence-driven project evolution

V1.4.1 adds a lightweight, evidence-driven process for improving the project from real use. The V1.4 runtime, installed skill, knowledge base, and model-led optimization workflow are unchanged.

• Adds public templates for Case Snapshot, Evaluation Definition, Evaluation Result, and Release Decision.
• Documents how public cases, private practice, repository revisions, evaluation evidence, and maintainer decisions remain separate.
• Publishes the first retrospective replay for Profiler evidence-object validation, with an explicit claim limit: it confirms two deterministic validation behaviors, not GPU performance or broad generality.
• Updates the contribution guide, pull request checks, documentation navigation, and structural tests.
• Does not automatically upload private material, admit knowledge, modify the running skill, submit pull requests, merge changes, or publish releases.

Validation: 230 CPU tests passed on both Python 3.10 and Python 3.12 in GitHub Actions; staged-installation smoke, Python compilation, and skill self-check also passed.

V1.4.0 — Model-led GPU optimization

V1.4.0 将优化决策收回 ChatGPT，随 skill 安装的工具只执行一次明确操作并记录可复核事实。

• 删除自动规划、全局流程状态和重复执行入口，不保留旧流程兼容层。
• 生产代码收敛为 17 个模块，以 Target、Variant、Experiment、Invocation 和 Champion 作为唯一持久模型。
• 候选不会自动晋级；正确性、短版初筛、profiler 和正式比较由 ChatGPT 根据当前证据显式启动。
• NCU、Nsys、PyTorch Profiler、编译产物和 SASS 只返回身份绑定的事实；未知版本、关键字段和单位 fail closed。
• 离线知识保持 advisory；没有匹配不会阻断源码分析、profiling 或新的优化假设。
• 补齐长任务心跳、启动期限、本机进程组清理和安装后公开请求格式。

本次发布在 Python 3.10 与 3.12 上各通过 225 项自动化测试；RTX 5090 物理验收通过 5 项检查。完整验证范围见 docs/validation.md。

这是不向后兼容的架构收敛版本。旧 Controller、自动投资控制和旧 schema 不再提供入口。

V1.3.0 — Knowledge-backed direction analysis

V1.3.0 强化了从当前 workload 证据中提出可证伪优化方向的能力。\n\n- 加入受环境、源码和 workload 身份约束的离线知识检索。\n- 原始 profile 缺少机制语义时，可先执行一次低成本只读检查。\n- 知识库没有匹配时，不再阻断模型根据已封存 profile、执行路径和源码提出方向。\n- 知识候选保持 advisory，不产生收益事实、代码修改权限或 promotion。\n- RTX 5090 保留案例回归中，V1.3 命中 3/4 个已推广机制，并将 profiler 建议从 4 次降为 0 次。\n\n限制：上述指标来自已知案例回归，不代表新 workload 的命中率；收益结论仍需当前 workload 的正确性与成对性能证据。

V1.2.0 — Adaptive investment control

V1.2.0 adds adaptive investment control to the performance decision engine introduced in V1.1. The Controller now decides whether the next stage is worth running from measured headroom, evidence quality, expected cost, scope, risk, and user authorization instead of treating the time ceiling as a target.

Highlights:

• Adds run-level grants for execution time, modification scope, risk, and the furthest permitted validation stage.
• Freezes one ChangeSet and advances it through static review, correctness, short paired timing, targeted profiling when needed, and formal workload validation.
• Preserves completed stages and candidates across interruption; authorization pauses keep the working state, while rejection or explicit abandonment restores the original implementation.
• Prevents failed or low-value candidates from starting unnecessary expensive stages.
• Keeps search and external AI review advisory and bound to locally verifiable evidence.
• Reworks the Chinese README as the primary user guide and clarifies test workload and correctness-check requirements in both languages.

Validation:

• 1,299 automated tests: 1,289 passed and 10 physical-GPU opt-in tests were skipped in the CPU/static suite.
• Python 3.10 and Python 3.12 CI passed for the released source.
• Installable-skill tests, staged-installation smoke, source compilation, and self_check passed.
• Physical GPU coverage and known profiler-permission limits remain documented in docs/validation.md.

Upgrade note:

• Unfinished runs created before the V1.2 run-grant protocol are not migrated automatically. Start a new V1.2 run for controlled execution.

---

V1.2.0 在 V1.1 性能决策引擎的基础上加入自适应投入控制。Controller 会根据可优化空间、证据质量、后续成本、修改范围、风险和用户授权，判断下一阶段是否值得执行，不再把时间上限当成需要用完的预算。

主要变化：

• 以运行级授权统一约束执行时间、修改范围、风险和最远验证阶段。
• 每次冻结一个 ChangeSet，依次完成静态检查、正确性校验、短版成对测试、必要的定向 profile 和正式 workload 验证。
• 中断后保留已完成阶段和候选现场；授权不足时暂停，候选被否决或明确放弃后恢复原实现。
• 前一阶段失败或预期收益不足时，不再启动后续昂贵验证。
• 外部搜索和第三方 AI 只提供方向挑战，最终结论仍由本地正确性与性能证据决定。
• 以中文 README 作为主要使用文档，并统一说明测试 workload 和正确性校验要求。

验证情况：

• 自动化测试共 1,299 项，其中 1,289 项通过，10 项物理 GPU 可选测试在 CPU/static 测试中跳过。
• Python 3.10 和 Python 3.12 的 GitHub CI 均通过。
• 安装包测试、临时安装验证、源码编译检查和 self_check 均通过。
• 物理 GPU 覆盖范围和 profiler 权限限制记录在 docs/validation.md。

升级说明：

• V1.2 之前尚未结束的运行不会自动迁移到新的运行级授权协议，受控执行请新建 V1.2 任务。

V1.1.0 — Performance decision engine

V1.1.0 adds an evidence-gated performance decision engine for CUDA, CUTLASS, Triton, and complete GPU workloads.

Highlights:

• Builds a deterministic critical-path performance model from admitted evidence.
• Keeps active diagnosis to three competing mechanisms and one lowest-cost next action.
• Produces an early investment brief before expensive profiling.
• Returns MEASURE, PURSUE, REVIEW_REQUIRED, or STOP from explicit evidence and project thresholds.
• Prevents closed mechanisms from being reopened through renamed hypotheses or overlapping scopes without fresh, relevant evidence.
• Records external AI challenges without treating unrelated workload evidence as an answer.
• Adds four RTX 5090 Controller evidence-admission scenarios and a separately authorized disposable NCU smoke path.

Validation:

• 1,190 automated tests: 1,180 passed and 10 physical-GPU opt-in tests skipped in the CPU/static suite.
• Python 3.10 and 3.12 CI passed for commit 095c872.
• The physical RTX 5090 V1.1 lane passed 24 of 24 checks; the final four hardened Controller scenarios were rerun successfully.

V1.1.0 为 CUDA、CUTLASS、Triton 与完整 GPU workload 增加了基于证据的性能决策引擎。它会先建立关键路径模型，再从少量竞争机制中选择最低成本的下一步；在进入昂贵 profile 前给出投入判断，并阻止已关闭方向通过改名或调整范围重复消耗优化轮次。

V1.0.1

This patch completes the standalone release boundary.

• Ships LICENSE and NOTICE with subdirectory skill installs.
• Makes the RTX 5090 acceptance lane portable and rejects writable artifact mounts that overlap the repository.
• Applies the hard deadline to open-iter before process launch and during profiling, with durable checkpoint and terminal evidence.
• Moves unversioned schema identities to the standalone repository while retaining versioned pre-V1 protocol IDs.
• Clarifies tested Python and POSIX support, release installation, and release-versus-protocol terminology.

Validation: 1,122 project tests, 14 installed-skill tests, Python 3.10/3.12 CI, staged installation smoke, self-check, compileall, and shellcheck.

v1.0.0 — First standalone release

English

First standalone release of CUDA Kernel Optimizer.

• Environment readiness and resumable active diagnosis
• Evidence-gated CUDA, CUTLASS, Triton, and complete-workload optimization
• Deterministic long-run recovery and staged cost controls
• 1,111 local CPU/static tests; nine physical RTX 5090 tests remain opt-in
• GitHub CI verified on Python 3.10 and 3.12

Pre-v1 development history remains in troycheng/cuda-optimized-skill and the internal mlsys/cuda-optimized-skill repository.

简体中文

CUDA Kernel Optimizer 首个独立公开版本。

• 环境准备与可恢复的主动诊断
• 以证据门禁控制 CUDA、CUTLASS、Triton 和完整 workload 优化
• 确定性的长任务恢复与分阶段成本控制
• 本地 CPU/static 测试共 1,111 项；9 项物理 RTX 5090 测试仍为 opt-in
• GitHub CI 已在 Python 3.10 和 3.12 上通过

V1 之前的完整开发历史保留在 troycheng/cuda-optimized-skill 和内网 mlsys/cuda-optimized-skill。