[hmljy2020]

Summary

为 issue #361 建立正式 ACIR v2 资源与阶段基础。后续 Rule/Method 和 link 需要一致地回答：资源由谁分配、输出由谁提供、导出是否复制，以及实例是否已闭合。本 PR 将这些规则放入正式 Dialect，复用实验分支的结构验证模式；不增加另一条 frontend。

linked 仅表示资源/实例结构闭合，不代表事务可执行。Python 默认入口、Runtime 和后端生成未切换。

Change type

• Feature / behavior change
• Tests / CI
• Documentation / configuration

Related issues

Related to #361。PR-01，无前置实现 PR，不关闭整个 issue。feat/transactional-acir-gfsim 是复用来源，不是语义权威。

Changes

• 区分 primitive<T>、state<T>、queue<T> 和 module_ref<@M>。State/Queue 显式分配并保留身份；Queue 深度属于分配配置；实例引用不创建实例。
• output_kinds 区分外部提供的 bound 和导出已有资源的 export，打印明确显示 inputs(...) bound(...)。
• 独立 v2 验证器检查阶段、接口和资源/实例闭包；旧 source/link、Routed/final、anchor 和两个 emitter 拒绝新阶段。
• 按用户指导精简测试：新增 lit 从 1139 行减至 137 行，保留资源身份、绑定/导出、payload、重读和阶段边界；56 个反例减至 3 个核心检查。移除的 53 个案例是覆盖范围缩减，不计为当前 PASS，也不把临时实现限制固化为语言要求。
• 更新普通 Instance 打印的测试适配器，保留原数值、时序和实例数量断言；未重写 DUT。修复这些 Python 改动的 Black 格式问题。
• AGENTS.md 增加两项长期要求：每次提交/更新 PR 前精简内容，测试验证核心意图；提交前运行 CI changed-file Ruff/Black 等检查，格式化修改后复跑，不能以语法或功能测试替代。

主要修改所有者为 IR/Dialect、旧路径接纳边界、测试适配器和文档。编译器逻辑在本轮精简中未改变。PR 总新增行数从 2022 减至 1019；保留契约文档 docs/development/transactional-resource-ir.md。

Validation

当前提交：b69202aa8f63a89d81f23f0839e317009850d852；基线：b986bb0ad721bc0834a71fc710dddca4bde339f4。使用本 checkout 构建的 Clang/LLVM/MLIR 22.1.8 工具链，通过 pyc6/scripts/run.sh 执行：

范围	实际结果
精简后的新增 lit	5/5 PASS，含 3 个核心拒绝案例
本轮 lit gate + 原生 anchor 测试	PASS，退出码 0
changed-file pre-commit（与 CI 相同 hook 选择）	PASS，包含 Ruff、Black、Markdown 检查
暂存文档快照 mkdocs build --strict	PASS；工作目录未提交 RFC 草稿的失效链接单独保留为本地失败记录
独立精简审查	clear；16 个格式化 Python 文件 AST 不变，编译器逻辑不变
git diff --cached --check	PASS

本轮运行：ctest --test-dir .pycircuit_out/local-clang22/build -R '^(ACIRLitGateTests|ACIRAnchorArtifactTests)$' --output-on-failure --no-tests=error；SKIP=pyc-api-hygiene pre-commit run --files <PR changed files>。

精简前同一编译器实现已通过完整 API gate 25/25 和新消费者目录 example gate 13/13，包括既有 C++/RTL/四态回归；本轮未重复整个硬件矩阵。旧版本远端 Python Checks 的 Black 失败已修复，当前远端 CI 以 checks 为准。

NOT RUN：新事务契约执行仿真（本批无执行 lowering）、nightly/长 oracle/变异矩阵、非 gate 历史 helper 执行。原始日志与两个版本的审查记录保留在本地忽略目录 docs/gates/logs/20261027-pr01-resources/；上表是远端可读摘要，不将本地路径冒充公开 artifact。

• Executed checks and outcomes are recorded above
• Skipped, deferred or unavailable checks and reasons are recorded above

Review checklist

• Scope and affected contracts are identified; new proposals are distinguished from approved behavior
• Tests cover relevant supported behavior and rejection/rollback boundaries
• Both backends preserve common-IR semantics where affected; no fixture-specific admission or fallback shim
• Active documentation and remaining callers are updated where needed
• No generated build output, temporary evidence logs or secrets are committed
• Required CI and independent reviews are complete for this candidate

Reviewer notes

本地独立审查已完成，最后一项等待当前远端 CI。实现与测试由不同实例负责；独立架构/审查使用 gpt-6-astra，其余实例继承会话模型/effort，未显式 override。本轮独立审查确认核心意图保留，并按用户最新要求移除边缘矩阵，未以减少覆盖宣称额外能力。

后续：PR-02 Rule/Method；PR-07 primitive 具体化、实际绑定和默认 Queue；PR-09 泛型 Module 要求与跨文件 link。无自动 merge。

[zhoubot]

Reviewed PR #363 head b69202a with its full diff and all review/comment/thread surfaces. Independent resource review and independent core-boundary tests found no remaining actionable defect.

Fixed an integration regression in prerequisite #366 (merged as c220222): the new normal instance printer caused three void-module tests from #364 to count zero children. Only two text filters changed; exact callee, instance-count, result and execution checks remain.

Fresh verification of the combined candidate 19fcf70b3a00e32ac47dbb7e723dfd438cdb0aac, tree 55cbb4999cdd2bd5aad5df99a34b08969e61d6e2, built and installed from its own checkout with LLVM/MLIR 22.1.8 and lit 18.1.8:

• bash tools/run_api_tests.sh --tier gate: PASS 25/25.
• bash tools/run_examples.sh --tier gate: PASS 13/13.
• python -m pytest -q tests/system/test_void_module_results.py: PASS 30/30 (previously three reproduced failures).
• New transactional lit: PASS 5/5; anchor tests PASS.
• Independent checks: forwarded bound/export aliases, distinct allocations under CSE, reload stability, ownerless alias cycle and bound replacement rejection, both emitter boundary rejections PASS.
• Changed-file pre-commit (CI hook selection), mkdocs build --strict, git diff --check: PASS.

No transaction execution/lowering capability is claimed. Nightly/long mutation matrices were not run. Implementation/integration, independent resource testing and independent review used separate instances; children inherited session model/effort without override. Raw receipts remain locally under docs/gates/logs/20261027-review-pr363/; this review supplies the remotely readable outcome summary.