@cannbot-plugin/cannbot-harness 0.0.0-stage → 0.1.5

package/assets/ops-direct-invoke/skills/ops-spec-gen/scripts/validate_spec.py

@@ -0,0 +1,1848 @@
+#!/usr/bin/env python3
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+
+"""spec.yaml validator — full 12-stage L0 校验.
+
+Stage 1: jsonschema validation against schemas/op-spec.json.
+Stage 2: category ↔ paradigm consistency, mutual exclusion,
+         paradigm internal constraints, and white-list checks for
+         primitives.op / invariants.kind / machine_check.kind / synthesize.patterns / error_type.
+Stage 3: shape_closure — solve outputs[].shape_rule via shape DSL.
+Stage 4: dtype_closure — cross-check dtype_rule with supported_combinations.
+Stage 5: broadcast_legality — validate broadcast semantics of operator computation.
+Stage 6: boundary_min_set — per-paradigm minimum case set.
+Stage 7: tolerance_coverage — per-dtype tolerance covers output dtypes + tightness.
+Stage 8: formula_smoke_eval — run formula on tiny tensors via numpy sandbox.
+Stage 9: oracle_reachable — real import framework + walk api attribute chain.
+Stage 10: formula_oracle_equiv — compare formula vs oracle outputs on special-value inputs (NaN/inf/large/signed-zero).
+
+Stage interface contract:
+  - status ∈ {PASS, FAIL, SKIP}
+  - findings: list of {severity, rule_id, field_path, message, suggested_fix}
+
+Usage:
+    python3 validate_spec.py path/to/spec.yaml
+    python3 validate_spec.py path/to/spec.yaml --json     # machine-readable output
+    python3 validate_spec.py path/to/spec.yaml --strict   # exit non-zero on warnings too
+"""
+
+from __future__ import annotations
+
+import argparse
+import ast
+import functools
+import json
+import re
+import sys
+from dataclasses import dataclass, field, asdict
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+try:
+    import jsonschema
+    from jsonschema import Draft202012Validator
+except ImportError:
+    print("ERROR: jsonschema not installed. Run: pip install jsonschema pyyaml", file=sys.stderr)
+    sys.exit(2)
+
+SKILL_DIR = Path(__file__).resolve().parent.parent
+REGISTRIES = SKILL_DIR / "registries"
+SCHEMA_PATH = SKILL_DIR / "schemas" / "op-spec.json"
+
+
+@dataclass
+class Finding:
+    severity: str       # error | warning | info
+    rule_id: str
+    field_path: str
+    message: str
+    suggested_fix: str | None = None
+
+
+@dataclass
+class StageResult:
+    stage_id: int
+    status: str         # PASS | FAIL | SKIP
+    findings: list[Finding] = field(default_factory=list)
+
+
+def _load_yaml(p: Path) -> dict:
+    """Load a YAML file. Raises yaml.YAMLError on parse error (caller decides exit code)."""
+    with open(p, "r", encoding="utf-8") as f:
+        return yaml.safe_load(f)
+
+
+def _get(d: dict, *path, default=None):
+    cur: Any = d
+    for k in path:
+        if isinstance(cur, dict) and k in cur:
+            cur = cur[k]
+        else:
+            return default
+    return cur
+
+
+def stage_1(spec: dict) -> StageResult:
+    """Validate against op-spec.json (Draft 2020-12)."""
+    res = StageResult(stage_id=1, status="PASS")
+    schema = _load_schema()
+    validator = Draft202012Validator(schema)
+    errors = sorted(validator.iter_errors(spec), key=lambda e: list(e.absolute_path))
+    for err in errors:
+        path = "/".join(str(p) for p in err.absolute_path) or "<root>"
+        res.findings.append(Finding(
+            severity="error",
+            rule_id="schema_static",
+            field_path=path,
+            message=err.message,
+            suggested_fix=None,
+        ))
+    if res.findings:
+        res.status = "FAIL"
+    return res
+
+
+@functools.lru_cache(maxsize=1)
+def _load_schema() -> dict:
+    """Cached JSON Schema; deserialized once per process."""
+    return json.loads(SCHEMA_PATH.read_text(encoding="utf-8"))
+
+
+def _output_count_attr_determined(spec: dict) -> bool:
+    """True 当算子声明 op.output_count_determined_by=attribute（split/chunk/unbind 类：
+    输出个数由属性决定，非 data-dependent）。用于豁免 VariableOutput 的 data_dependent_shape
+    与 nonzero 类 boundary case。"""
+    return (spec.get("op") or {}).get("output_count_determined_by", "data") == "attribute"
+
+
+def stage_2(spec: dict, registries: dict) -> StageResult:
+    """category ↔ paradigm consistency + paradigm internal constraints + whitelists.
+
+    Dispatches to 6 focused checks; each appends to res.findings. Overall FAIL if
+    any error-severity finding is produced. Sub-checks correspond to the design
+    doc's lettered sub-rule groups (A/B/C/D/E/F + the H 'hint' group).
+    """
+    res = StageResult(stage_id=2, status="PASS")
+    cat = _get(spec, "op", "category")
+    paradigms = set(_get(spec, "op", "paradigms", default=[]) or [])
+
+    _check_category_paradigm_consistency(cat, paradigms, registries, res.findings)
+    _check_mutually_exclusive_paradigms(paradigms, registries, res.findings)
+    _check_paradigm_internal_constraints(spec, paradigms, registries, res.findings)
+    _check_paradigm_groups(spec, paradigms, res.findings)
+    _check_invariants(spec, registries, res.findings)
+    _check_machine_check(spec, registries, res.findings)
+    _check_synthesize_patterns(spec, registries, res.findings)
+    _check_supported_chips(spec, registries, res.findings)
+    _check_anti_pattern_ids(spec, registries, res.findings)
+    _check_op_error_codes(spec, registries, res.findings)
+    _check_composition_hint(spec, paradigms, res.findings)
+    _check_broadcast_paradigm_hint(spec, paradigms, res.findings)
+    _check_format_variants(spec, res.findings)
+    _check_interface_contracts(spec, res.findings)
+    _check_data_distribution(spec, cat, paradigms, res.findings)
+
+    if any(f.severity == "error" for f in res.findings):
+        res.status = "FAIL"
+    return res
+
+
+# ---------- stage 2 sub-checks --------------------------------------------
+
+
+def _check_interface_contracts(spec: dict, findings: list[Finding]) -> None:
+    """Validate TensorList and conditional-interface additions.
+
+    JSON Schema owns field types and conditional requiredness.  This stage owns
+    cross-reference validation (list-length references and semantic cases) and
+    the deliberately small, non-executable guard/expression languages.
+    """
+    inputs = spec.get("inputs") or []
+    outputs = spec.get("outputs") or []
+    attributes = spec.get("attributes") or []
+    input_by_name = {item.get("name"): item for item in inputs}
+    output_by_name = {item.get("name"): item for item in outputs}
+    attr_names = {item.get("name") for item in attributes}
+    attrs_by_name = {item.get("name"): item for item in attributes}
+
+    def error(rule_id: str, path: str, message: str, fix: str) -> None:
+        findings.append(Finding("error", rule_id, path, message, fix))
+
+    def check_list_length(item: dict, path: str, default_role: str) -> None:
+        role = item.get("role", default_role)
+        length = item.get("list_length")
+        if role == "tensor_list" and not isinstance(length, dict):
+            error(
+                "interface_contract.tensor_list_length_missing", path,
+                "role=tensor_list 必须声明 list_length。",
+                "添加 list_length: {kind: unconstrained} 或来源支持的更强约束。",
+            )
+            return
+        if role != "tensor_list" and length is not None:
+            error(
+                "interface_contract.list_length_non_list", f"{path}.list_length",
+                "list_length 只允许用于 role=tensor_list。",
+                "删除 list_length，或将 role 改为 tensor_list。",
+            )
+            return
+        if not isinstance(length, dict):
+            return
+
+        kind = length.get("kind")
+        allowed_fields = {
+            "unconstrained": {"kind"},
+            "fixed": {"kind", "value"},
+            "range": {"kind", "min", "max"},
+            "same_as": {"kind", "ref"},
+            "expression": {"kind", "value"},
+        }
+        if kind in allowed_fields and (unexpected := set(length) - allowed_fields[kind]):
+            error(
+                "interface_contract.list_length_unexpected_field", f"{path}.list_length",
+                f"list_length.kind={kind!r} 不允许字段 {sorted(unexpected)}。",
+                f"仅保留 {sorted(allowed_fields[kind])}。",
+            )
+        if kind == "fixed" and isinstance(length.get("value"), int):
+            if length["value"] < 0:
+                error(
+                    "interface_contract.list_length_negative", f"{path}.list_length.value",
+                    "list_length.fixed value 必须非负。",
+                    "将 value 设置为大于等于 0 的整数。",
+                )
+        elif kind == "range" and isinstance(length.get("min"), int) and isinstance(length.get("max"), int):
+            if length["min"] < 0 or length["max"] < 0:
+                error(
+                    "interface_contract.list_length_negative", f"{path}.list_length",
+                    "list_length.range 的 min 和 max 必须非负。",
+                    "将 min 和 max 设置为大于等于 0 的整数。",
+                )
+            if length["min"] > length["max"]:
+                error(
+                    "interface_contract.list_length_invalid_range", f"{path}.list_length",
+                    "list_length.range 必须满足 min <= max。",
+                    "交换或修正 min/max。",
+                )
+        elif kind == "same_as":
+            ref = length.get("ref")
+            target = None
+            if isinstance(ref, str) and ref.startswith("input."):
+                target = input_by_name.get(ref.removeprefix("input."))
+            elif isinstance(ref, str) and ref.startswith("output."):
+                target = output_by_name.get(ref.removeprefix("output."))
+            if target is None:
+                error(
+                    "interface_contract.list_length_unknown_ref", f"{path}.list_length.ref",
+                    f"list_length.same_as 引用了不存在的 {ref!r}。",
+                    "引用已声明的 input.<name> 或 output.<name>。",
+                )
+            elif target.get("role", "tensor") != "tensor_list":
+                error(
+                    "interface_contract.list_length_ref_not_list", f"{path}.list_length.ref",
+                    f"{ref!r} 不是 role=tensor_list，不能作为列表长度引用。",
+                    "改为引用 tensor_list，或为目标声明 role: tensor_list。",
+                )
+        elif kind == "expression":
+            expression = length.get("value")
+            if isinstance(expression, str):
+                for message in _validate_list_length_expression(expression, attrs_by_name, set(input_by_name)):
+                    error(
+                        "interface_contract.list_length_expression", f"{path}.list_length.value",
+                        message,
+                        "只使用 attr.<name>、len(attr.<list_attr>)、shape.<input>[<常量轴>]、整数、+、-、* 和括号。",
+                    )
+
+    for index, item in enumerate(inputs):
+        check_list_length(item, f"inputs[{index}]", "tensor")
+    for index, item in enumerate(outputs):
+        check_list_length(item, f"outputs[{index}]", "tensor")
+    _check_list_length_cycles(inputs, outputs, findings)
+
+    seen_ids: set[str] = set()
+    signatures: dict[str, tuple[set[str], set[str], set[str], set[str]]] = {}
+    for index, case in enumerate(spec.get("semantic_cases") or []):
+        path = f"semantic_cases[{index}]"
+        case_id = case.get("id")
+        if case_id in seen_ids:
+            error(
+                "interface_contract.semantic_case_duplicate_id", f"{path}.id",
+                f"semantic case id {case_id!r} 重复。",
+                "为每个 semantic case 使用唯一 id。",
+            )
+        seen_ids.add(case_id)
+
+        when = case.get("when")
+        if isinstance(when, str):
+            for message in _validate_semantic_guard(when, attr_names, set(input_by_name)):
+                error(
+                    "interface_contract.semantic_case_invalid_guard", f"{path}.when", message,
+                    "guard 只可引用 attr.<name>、input.<name>.dtype、input.<name>.is_present，"
+                    "并使用 ==/!=/in/not in/and/or/not。",
+                )
+
+        case_inputs = case.get("inputs") or {}
+        case_outputs = case.get("outputs") or {}
+        required = set(case_inputs.get("required") or [])
+        forbidden = set(case_inputs.get("forbidden") or [])
+        present = set(case_outputs.get("present") or [])
+        absent = set(case_outputs.get("absent") or [])
+        _check_case_name_set(required, forbidden, input_by_name, "input", path, findings)
+        _check_case_name_set(present, absent, output_by_name, "output", path, findings)
+
+        for name in required | forbidden:
+            if name in input_by_name and not input_by_name[name].get("optional", False):
+                error(
+                    "interface_contract.semantic_case_input_not_optional", f"{path}.inputs",
+                    f"{name!r} 不是 optional input，不能由 semantic_cases 改变存在性。",
+                    "在 input 上添加 optional: true，或删除该 case 约束。",
+                )
+        for name in present | absent:
+            if name in output_by_name and not output_by_name[name].get("optional", False):
+                error(
+                    "interface_contract.semantic_case_output_not_optional", f"{path}.outputs",
+                    f"{name!r} 不是 optional output，不能由 semantic_cases 改变存在性。",
+                    "在 output 上添加 optional: true，或删除该 case 约束。",
+                )
+
+        signature = (required, forbidden, present, absent)
+        if isinstance(when, str) and when in signatures and signatures[when] != signature:
+            error(
+                "interface_contract.semantic_case_conflicting_same_guard", path,
+                f"guard {when!r} 与先前 case 的 input/output 存在性约束冲突。",
+                "合并 case，或使用互斥的 when guard。",
+            )
+        elif isinstance(when, str):
+            signatures[when] = signature
+
+    _check_layout_contract(
+        spec.get("layout_contract"), input_by_name, output_by_name, attr_names, findings,
+    )
+
+
+def _check_layout_contract(
+    contract: Any, input_by_name: dict, output_by_name: dict,
+    attr_names: set[str], findings: list[Finding],
+) -> None:
+    """Validate cross references in the optional logical-layout contract.
+
+    Schema validation owns the tree shape, identifiers and duplicate axis names.
+    This function owns references to the declared interface, guard DSL reuse and
+    the one rank fact available in the current interface model: input.rank_range.
+    """
+    if not isinstance(contract, dict):
+        return
+
+    def error(rule_id: str, path: str, message: str, fix: str) -> None:
+        findings.append(Finding("error", rule_id, path, message, fix))
+
+    seen_ids: set[str] = set()
+    for index, variant in enumerate(contract.get("variants") or []):
+        if not isinstance(variant, dict):
+            continue
+        path = f"layout_contract.variants[{index}]"
+        variant_id = variant.get("id")
+        if variant_id in seen_ids:
+            error(
+                "interface_contract.layout_contract_duplicate_variant_id", f"{path}.id",
+                f"layout variant id {variant_id!r} 重复。",
+                "为每个 layout variant 使用唯一 id。",
+            )
+        seen_ids.add(variant_id)
+
+        when = variant.get("when")
+        if isinstance(when, str):
+            for message in _validate_semantic_guard(when, attr_names, set(input_by_name)):
+                error(
+                    "interface_contract.layout_contract_invalid_guard", f"{path}.when", message,
+                    "guard 只可引用 attr.<name>、input.<name>.dtype、input.<name>.is_present，"
+                    "并使用 ==/!=/in/not in/and/or/not。",
+                )
+
+        seen_refs: set[str] = set()
+        for tensor_index, tensor in enumerate(variant.get("tensors") or []):
+            if not isinstance(tensor, dict):
+                continue
+            tensor_path = f"{path}.tensors[{tensor_index}]"
+            ref = tensor.get("ref")
+            if ref in seen_refs:
+                error(
+                    "interface_contract.layout_contract_duplicate_tensor_ref", f"{tensor_path}.ref",
+                    f"{ref!r} 在同一 layout variant 中重复。",
+                    "每个 Tensor 在一个 variant 中只声明一次 logical_axes。",
+                )
+            seen_refs.add(ref)
+
+            target = None
+            is_input = False
+            if isinstance(ref, str) and ref.startswith("input."):
+                target = input_by_name.get(ref.removeprefix("input."))
+                is_input = True
+            elif isinstance(ref, str) and ref.startswith("output."):
+                target = output_by_name.get(ref.removeprefix("output."))
+            if target is None:
+                error(
+                    "interface_contract.layout_contract_unknown_ref", f"{tensor_path}.ref",
+                    f"layout_contract 引用了未声明的 Tensor {ref!r}。",
+                    "引用已声明的 input.<name> 或 output.<name>。",
+                )
+                continue
+            if target.get("role", "tensor") != "tensor":
+                error(
+                    "interface_contract.layout_contract_ref_not_tensor", f"{tensor_path}.ref",
+                    f"{ref!r} 的 role={target.get('role', 'tensor')!r}，不能声明单 Tensor 的 logical_axes。",
+                    "只引用 role=tensor 的 input/output；TensorList 的元素布局需单独建模。",
+                )
+                continue
+
+            # Outputs currently have no rank_range. Do not infer rank from a
+            # shape formula; only validate an explicit, exact input rank.
+            rank_range = target.get("rank_range") if is_input else None
+            axes = tensor.get("logical_axes")
+            if (
+                isinstance(rank_range, list) and len(rank_range) == 2
+                and rank_range[0] == rank_range[1]
+                and isinstance(rank_range[0], int)
+                and isinstance(axes, list) and len(axes) != rank_range[0]
+            ):
+                error(
+                    "interface_contract.layout_contract_rank_mismatch", f"{tensor_path}.logical_axes",
+                    f"{ref!r} 的固定 rank 为 {rank_range[0]}，但 logical_axes 有 {len(axes)} 项。",
+                    "使 logical_axes 项数与该 input 的固定 rank 一致，或修正 rank_range。",
+                )
+
+
+def _check_case_name_set(
+    positive: set[str], negative: set[str], declared: dict, kind: str,
+    path: str, findings: list[Finding],
+) -> None:
+    if overlap := positive & negative:
+        findings.append(Finding(
+            "error", "interface_contract.semantic_case_conflicting_members", path,
+            f"同一 semantic case 中 {kind} 同时出现在相反集合：{sorted(overlap)}。",
+            "将每个名称仅保留在 required/forbidden 或 present/absent 的一侧。",
+        ))
+    for name in sorted((positive | negative) - set(declared)):
+        findings.append(Finding(
+            "error", "interface_contract.semantic_case_unknown_member", path,
+            f"semantic case 引用了未声明的 {kind} {name!r}。",
+            f"改为已声明的 {kind} 名称，或先在接口中声明它。",
+        ))
+
+
+def _check_list_length_cycles(inputs: list[dict], outputs: list[dict], findings: list[Finding]) -> None:
+    """same_as may follow input/output references, but it must not form a cycle."""
+    refs: dict[str, str] = {}
+    for prefix, items in (("input", inputs), ("output", outputs)):
+        for item in items:
+            length = item.get("list_length") or {}
+            if length.get("kind") == "same_as" and isinstance(length.get("ref"), str):
+                refs[f"{prefix}.{item.get('name')}"] = length["ref"]
+
+    reported_cycles: set[frozenset[str]] = set()
+    for start in refs:
+        seen: list[str] = []
+        current = start
+        while current in refs:
+            if current in seen:
+                cycle = seen[seen.index(current):] + [current]
+                cycle_nodes = frozenset(cycle[:-1])
+                if cycle_nodes not in reported_cycles:
+                    findings.append(Finding(
+                        "error", "interface_contract.list_length_cycle", "list_length.ref",
+                        f"list_length.same_as 存在循环引用：{' -> '.join(cycle)}。",
+                        "让至少一个 TensorList 使用 unconstrained/fixed/range/expression 作为长度根。",
+                    ))
+                    reported_cycles.add(cycle_nodes)
+                break
+            seen.append(current)
+            current = refs[current]
+def _validate_semantic_guard(guard: str, attr_names: set[str], input_names: set[str]) -> list[str]:
+    """Return validation errors for the non-executable semantic-case guard DSL."""
+    try:
+        tree = ast.parse(guard, mode="eval")
+    except SyntaxError as exc:
+        return [f"guard 不是合法表达式：{exc.msg}。"]
+
+    errors: list[str] = []
+    parents = {
+        id(child): parent
+        for parent in ast.walk(tree)
+        for child in ast.iter_child_nodes(parent)
+    }
+
+    def ref(node: ast.AST) -> tuple[str, str, str] | None:
+        if not isinstance(node, ast.Attribute):
+            return None
+        if isinstance(node.value, ast.Name) and node.value.id == "attr":
+            return ("attr", node.attr, "")
+        if isinstance(node.value, ast.Name) and node.value.id == "input":
+            # `input.x` 是 `input.x.dtype/is_present` 的中间 AST 节点。
+            return ("input_base", node.attr, "")
+        if isinstance(node.value, ast.Attribute) and isinstance(node.value.value, ast.Name) and node.value.value.id == "input":
+            return ("input", node.value.attr, node.attr)
+        return None
+
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Call):
+            errors.append("guard 不允许函数调用。")
+        elif isinstance(node, ast.BinOp):
+            errors.append("guard 不允许算术运算。")
+        elif isinstance(node, ast.Compare) and any(not isinstance(op, (ast.Eq, ast.NotEq, ast.In, ast.NotIn)) for op in node.ops):
+            errors.append("guard 只允许 ==、!=、in、not in 比较。")
+        elif isinstance(node, ast.Attribute):
+            parsed = ref(node)
+            if parsed is None:
+                errors.append("guard 引用必须是 attr.<name>、input.<name>.dtype 或 input.<name>.is_present。")
+                continue
+            kind, name, leaf = parsed
+            if kind == "attr" and name not in attr_names:
+                errors.append(f"guard 引用了未声明属性 attr.{name}。")
+            if kind in {"input", "input_base"}:
+                if name not in input_names:
+                    errors.append(f"guard 引用了未声明输入 input.{name}。")
+                if kind == "input_base":
+                    parent = parents.get(id(node))
+                    if not (isinstance(parent, ast.Attribute) and parent.value is node):
+                        errors.append(
+                            f"guard 不支持裸 input.{name}；只允许 input.{name}.dtype 或 input.{name}.is_present。"
+                        )
+                if kind == "input" and leaf not in {"dtype", "is_present"}:
+                    errors.append(f"guard 不支持 input.{name}.{leaf}；只允许 dtype 或 is_present。")
+        elif isinstance(node, ast.Name) and node.id not in {"attr", "input", "true", "false"}:
+            errors.append(f"guard 不允许名称 {node.id!r}。")
+        elif isinstance(node, (ast.Expression, ast.Name, ast.BoolOp, ast.UnaryOp, ast.Compare, ast.Load,
+                               ast.And, ast.Or, ast.Not, ast.Eq, ast.NotEq, ast.In, ast.NotIn,
+                               ast.Constant, ast.List, ast.Tuple)):
+            continue
+        elif isinstance(node, ast.Attribute):
+            continue
+        else:
+            errors.append(f"guard 不允许语法节点 {type(node).__name__}。")
+    return list(dict.fromkeys(errors))
+
+
+def _validate_list_length_expression(
+    expression: str, attrs_by_name: dict[str, dict], input_names: set[str],
+) -> list[str]:
+    """Validate the intentionally small integer-expression language for list length."""
+    try:
+        tree = ast.parse(expression, mode="eval")
+    except SyntaxError as exc:
+        return [f"list_length expression 不是合法表达式：{exc.msg}。"]
+
+    errors: list[str] = []
+
+    def attribute_interval(name: str) -> tuple[int | None, int | None]:
+        """Return the statically known integer interval for an attribute."""
+        attr = attrs_by_name.get(name)
+        if not attr or attr.get("type") != "int64":
+            return None, None
+        constraint = attr.get("machine_constraint") or {}
+        lower = constraint.get("lower_inclusive")
+        upper = constraint.get("upper_inclusive")
+        if not isinstance(lower, int) or isinstance(lower, bool):
+            lower = None
+        if not isinstance(upper, int) or isinstance(upper, bool):
+            upper = None
+        return lower, upper
+
+    def interval(node: ast.AST) -> tuple[int | None, int | None]:
+        if isinstance(node, ast.Constant) and isinstance(node.value, int) and not isinstance(node.value, bool):
+            return node.value, node.value
+        if isinstance(node, ast.Attribute) and isinstance(node.value, ast.Name) and node.value.id == "attr":
+            return attribute_interval(node.attr)
+        if isinstance(node, ast.Call):
+            return 0, None
+        if isinstance(node, ast.Subscript):
+            return 0, None
+        if isinstance(node, ast.BinOp):
+            left_lower, left_upper = interval(node.left)
+            right_lower, right_upper = interval(node.right)
+            if isinstance(node.op, ast.Add):
+                lower = (
+                    left_lower + right_lower
+                    if left_lower is not None and right_lower is not None else None
+                )
+                upper = (
+                    left_upper + right_upper
+                    if left_upper is not None and right_upper is not None else None
+                )
+                return lower, upper
+            if isinstance(node.op, ast.Mult):
+                # A product is provably non-negative only when both operands
+                # are provably non-negative.
+                lower = (
+                    left_lower * right_lower
+                    if left_lower is not None and right_lower is not None
+                    and left_lower >= 0 and right_lower >= 0
+                    else None
+                )
+                if None not in (left_lower, left_upper, right_lower, right_upper):
+                    products = [
+                        left_lower * right_lower,
+                        left_lower * right_upper,
+                        left_upper * right_lower,
+                        left_upper * right_upper,
+                    ]
+                    return min(products), max(products)
+                return lower, None
+            if isinstance(node.op, ast.Sub):
+                lower = (
+                    left_lower - right_upper
+                    if left_lower is not None and right_upper is not None else None
+                )
+                upper = (
+                    left_upper - right_lower
+                    if left_upper is not None and right_lower is not None else None
+                )
+                return lower, upper
+        return None, None
+
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Name) and node.id not in {"attr", "shape", "len"}:
+            errors.append(f"list_length expression 不允许名称 {node.id!r}。")
+        elif isinstance(node, ast.Attribute):
+            if isinstance(node.value, ast.Name) and node.value.id == "attr":
+                if node.attr not in attrs_by_name:
+                    errors.append(f"list_length expression 引用了未声明属性 attr.{node.attr}。")
+            elif isinstance(node.value, ast.Name) and node.value.id == "shape":
+                if node.attr not in input_names:
+                    errors.append(f"list_length expression 引用了未声明输入 shape.{node.attr}。")
+            else:
+                errors.append("list_length expression 的属性引用只允许 attr.<name> 或 shape.<input>。")
+        elif isinstance(node, ast.Call):
+            valid_len_arg = (
+                len(node.args) == 1 and isinstance(node.args[0], ast.Attribute)
+                and isinstance(node.args[0].value, ast.Name) and node.args[0].value.id == "attr"
+            )
+            if not (isinstance(node.func, ast.Name) and node.func.id == "len" and valid_len_arg and not node.keywords):
+                errors.append("list_length expression 只允许 len(attr.<list_attr>) 调用。")
+            elif not str(attrs_by_name.get(node.args[0].attr, {}).get("type", "")).startswith("list["):
+                errors.append(f"len(attr.{node.args[0].attr}) 只能引用 list[...] 类型属性。")
+        elif isinstance(node, ast.Subscript):
+            if not (isinstance(node.value, ast.Attribute) and isinstance(node.value.value, ast.Name)
+                    and node.value.value.id == "shape" and isinstance(node.slice, ast.Constant)
+                    and isinstance(node.slice.value, int) and node.slice.value >= 0):
+                errors.append("list_length expression 下标只允许 shape.<input>[<非负常量轴>]。")
+        elif isinstance(node, ast.BinOp) and not isinstance(node.op, (ast.Add, ast.Sub, ast.Mult)):
+            errors.append("list_length expression 只允许 +、-、* 运算。")
+        elif isinstance(node, ast.Constant) and (not isinstance(node.value, int) or isinstance(node.value, bool) or node.value < 0):
+            errors.append("list_length expression 只允许非负整数常量。")
+        elif isinstance(node, (ast.Expression, ast.Name, ast.Attribute, ast.Call, ast.Subscript,
+                               ast.BinOp, ast.Add, ast.Sub, ast.Mult, ast.Load, ast.Constant)):
+            continue
+        else:
+            errors.append(f"list_length expression 不允许语法节点 {type(node).__name__}。")
+    bound, _ = interval(tree.body)
+    if bound is None:
+        errors.append(
+            "list_length expression 无法由已声明约束证明非负；使用非负 shape/len，"
+            "或为 int64 属性声明 machine_constraint.lower_inclusive >= 0；减法还需要 RHS upper_inclusive。"
+        )
+    elif bound < 0:
+        errors.append("list_length expression 的可证明下界小于 0。")
+    return list(dict.fromkeys(errors))
+
+
+def _check_category_paradigm_consistency(cat, paradigms, registries, findings):
+    """A. category 必含 paradigm 是否齐全；fused_composite 额外要求 ≥ 2 条基础 paradigm。"""
+    cmap = registries["category_map"]
+    required = set(cmap["category_requires_paradigms"].get(cat, []))
+    missing = required - paradigms
+    if missing:
+        findings.append(Finding(
+            severity="error",
+            rule_id="category_paradigm_consistency.required_paradigm_missing",
+            field_path="op.paradigms",
+            message=f"category={cat} 必含 paradigm {sorted(missing)}",
+            suggested_fix=f"在 op.paradigms 中加入 {sorted(missing)}",
+        ))
+
+    if cat == "FusedComposite":
+        basic = {"Broadcast", "Reduction", "Contraction",
+                 "ArgReduce", "LayoutTransform"}
+        if len(paradigms & basic) < 2:
+            findings.append(Finding(
+                severity="error",
+                rule_id="category_paradigm_consistency.fused_composite_basics",
+                field_path="op.paradigms",
+                message="category=fused_composite 必须含 ≥ 2 条基础 paradigm（Elementwise 被其他范式吸收，不计入）",
+                suggested_fix=f"从 {sorted(basic)} 中至少选 2 条加入 paradigms",
+            ))
+
+
+def _check_broadcast_paradigm_hint(spec, paradigms, findings):
+    """G. broadcast.kind=numpy 时 paradigms 应含 Broadcast（交叉校验：shape_rule 含 broadcast_shapes）。"""
+    broadcast_kind = (spec.get("broadcast") or {}).get("kind", "none")
+    has_broadcast_shape_rule = any(
+        "broadcast_shapes" in (out.get("shape_rule") or "")
+        for out in (spec.get("outputs") or [])
+    )
+
+    if broadcast_kind == "numpy" and "Broadcast" not in paradigms:
+        findings.append(Finding(
+            severity="error",
+            rule_id="category_paradigm_consistency.broadcast_kind_requires_paradigm",
+            field_path="op.paradigms",
+            message=(
+                f"broadcast.kind=numpy 但 paradigms {sorted(paradigms)} 不含 Broadcast。"
+                "当算子计算需要 broadcast 数据时，paradigms 必须包含 Broadcast"
+            ),
+            suggested_fix="在 op.paradigms 中加入 Broadcast，并将 op.category 改为 Broadcast",
+        ))
+
+    if has_broadcast_shape_rule and "Broadcast" not in paradigms:
+        findings.append(Finding(
+            severity="error",
+            rule_id="category_paradigm_consistency.shape_rule_broadcast_requires_paradigm",
+            field_path="op.paradigms",
+            message=(
+                f"outputs[].shape_rule 使用了 np.broadcast_shapes 但 paradigms {sorted(paradigms)} 不含 Broadcast。"
+                "输出 shape 由 broadcast 推导的算子应使用 Broadcast paradigm"
+            ),
+            suggested_fix="在 op.paradigms 中加入 Broadcast，并将 op.category 改为 Broadcast",
+        ))
+
+
+def _check_mutually_exclusive_paradigms(paradigms, registries, findings):
+    """B. paradigm 互斥对（同时出现 ⇒ FAIL）。"""
+    cmap = registries["category_map"]
+    for pair in cmap.get("mutually_exclusive_paradigms", []):
+        if set(pair) <= paradigms:
+            findings.append(Finding(
+                severity="error",
+                rule_id="category_paradigm_consistency.mutually_exclusive",
+                field_path="op.paradigms",
+                message=f"paradigms 互斥：{pair}",
+                suggested_fix=f"删除 {pair} 之一；写冲突另成 atomic_update",
+            ))
+
+
+def _check_paradigm_internal_constraints(spec, paradigms, registries, findings):
+    """C. 每个 paradigm 的内部结构性要求。
+
+    本函数是该规则的**真值**（category_paradigm_map.yaml 的 paradigm_constraints
+    段已删除，因为通用 mini-DSL 表达力不足，且会让维护者误以为改 yaml 即可改逻辑）。
+    """
+    primitives_wl = set(registries["primitives"])
+
+    if "NumericalStable" in paradigms:
+        if not _get(spec, "numerical_stability", "required"):
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.numerical_stable",
+                field_path="numerical_stability.required",
+                message="paradigms 含 NumericalStable ⇒ numerical_stability.required 必须为 true",
+                suggested_fix="把 numerical_stability.required 改为 true 并补 techniques",
+            ))
+
+    if "FusedComposite" in paradigms:
+        _check_fused_composite(spec, primitives_wl, findings)
+
+    if "Reduction" in paradigms and "Recurrence" not in paradigms:
+        inputs = spec.get("inputs") or []
+        has_input_axis = any(
+            inp.get("name") in ("axes", "axis", "dim")
+            and set(inp.get("dtype_set") or []).issubset({"int32", "int64"})
+            for inp in inputs
+        )
+
+        attrs = spec.get("attributes") or []
+        axis_names_attr = {a.get("name") for a in attrs}
+        has_attr_axis = bool(axis_names_attr & {"axis", "dim", "axes"})
+
+        reduction = spec.get("reduction") or {}
+        has_fixed_axis = reduction.get("axis_source") == "fixed"
+        has_implicit_all = reduction.get("axis_source") == "implicit_all"
+
+        if not (has_input_axis or has_attr_axis or has_fixed_axis or has_implicit_all):
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.reduction_axis_missing",
+                field_path="inputs | attributes | reduction",
+                message=(
+                    "paradigms 含 Reduction（且不含 Recurrence）⇒ 必须通过以下方式之一指定规约轴："
+                    "(A) 名为 axes/axis/dim 且 dtype∈{int32,int64} 的输入张量；"
+                    "(B) 名为 axis/dim/axes 的属性；"
+                    "(C) reduction.axis_source: fixed 声明固定轴；"
+                    "(D) reduction.axis_source: implicit_all 声明隐式全轴归约"
+                ),
+                suggested_fix=(
+                    "添加 axes/axis/dim 输入、添加 axis/dim/axes 属性、"
+                    "声明 reduction: {axis_source: fixed, fixed_value: <N>}、"
+                    "或声明 reduction: {axis_source: implicit_all}"
+                ),
+            ))
+
+    if "ArgReduce" in paradigms:
+        # ArgReduce 输出必须是 int32 / int64；用 dtype_eval 求值后检查
+        try:
+            sys.path.insert(0, str(Path(__file__).parent))
+            from evaluators import dtype_eval as _de
+            from evaluators.types import DslError as _DslError
+        except ImportError:
+            _de = None  # type: ignore
+        for i, out in enumerate(spec.get("outputs", [])):
+            rule = (out.get("dtype_rule") or "").strip()
+            kind = out.get("dtype_rule_kind", "numpy_expr")
+            derived: str | None = None
+            if _de is not None and kind == "numpy_expr" and rule:
+                # 用 dtype_policy.supported_combinations[0] 作为试探输入；
+                # ArgReduce 通常 dtype_rule 是 `y.dtype = np.int32 / int64` 不依赖输入
+                combos = ((spec.get("dtype_policy") or {})
+                          .get("supported_combinations") or [])
+                in_dt = combos[0].get("inputs", {}) if combos else {}
+                try:
+                    derived = _de.evaluate_dtype_rule(
+                        rule, output_name=out.get("name"),
+                        input_dtypes=in_dt,
+                    )
+                except _DslError:
+                    derived = None
+            if derived not in ("int32", "int64"):
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="paradigm_constraint.argreduce_dtype",
+                    field_path=f"outputs[{i}].dtype_rule",
+                    message=f"ArgReduce 输出 dtype 必须为 int32 / int64，当前 dtype_rule={rule!r} 推得 {derived!r}",
+                    suggested_fix="把 dtype_rule 改为 `<output>.dtype = np.int32` 或 `np.int64`",
+                ))
+
+    if "Stateful" in paradigms:
+        has_state_input = any(inp.get("role") == "state" for inp in spec.get("inputs", []))
+        has_inplace_output = any(
+            re.match(r"^inplace_with\(", out.get("aliasing") or "")
+            for out in spec.get("outputs", [])
+        )
+        if not (has_state_input or has_inplace_output):
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.stateful_state_or_inplace",
+                field_path="inputs[*].role | outputs[*].aliasing",
+                message="Stateful 必须有 inputs.role=state 或 outputs.aliasing=inplace_with(<state_input>)",
+                suggested_fix="添加 role: state 的 input，或在某 output 加 aliasing: inplace_with(<state_input>)",
+            ))
+
+    if "Quantization" in paradigms:
+        attr_names = {a.get("name") for a in (spec.get("attributes") or [])}
+        input_names = {i.get("name") for i in (spec.get("inputs") or [])}
+        # 量化算子的 scale/zero_point 可以是 attribute 或 tensor input；接受 5 种合法形式：
+        #   1. {scale, zero_point}     —— 经典非对称量化（W8A16 标准 quantize）
+        #   2. {scale}                 —— 对称量化（CANN QuantMatmul: deqScale 无 zp）
+        #   3. {x1Scale, x2Scale}      —— 双输入分别量化（CANN FusedQuantMatmul）
+        #   4. {dequant_scale}         —— 反量化算子专用
+        #   5. {deqScale}              —— CANN 驼峰别名
+        #   6. 算子名含 "quant"         —— 兜底：名称明确标识为量化算子
+        all_names = attr_names | input_names
+        canonical_forms = [
+            {"scale", "zero_point"},
+            {"scale"},
+            {"x1Scale", "x2Scale"},
+            {"dequant_scale"},
+            {"deqScale"},
+        ]
+        op_name = (_get(spec, "op", "name") or "").lower()
+        name_match = "quant" in op_name
+        if not name_match and not any(form <= all_names for form in canonical_forms):
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.quantization_attrs",
+                field_path="attributes",
+                message=(
+                    "Quantization 必须含以下任一组量化参数（attribute 或 input 均可）："
+                    "{scale, zero_point} / {scale} / {x1Scale, x2Scale} / "
+                    "{dequant_scale} / {deqScale}，或算子名含 'quant'"
+                ),
+                suggested_fix="按算子语义选择对应形式；对称量化只需 scale；双输入分别量化用 x1Scale + x2Scale",
+            ))
+
+    if "VariableOutput" in paradigms:
+        # 属性决定输出个数（split/chunk/unbind 类）：输出个数与 shape 由属性决定，
+        # 非 data-dependent，豁免 data_dependent_shape 约束（仍走 attribute-driven shape_rule）。
+        _attr_determined = _output_count_attr_determined(spec)
+        if not _attr_determined:
+            for i, out in enumerate(spec.get("outputs", [])):
+                if not out.get("data_dependent_shape"):
+                    findings.append(Finding(
+                        severity="error",
+                        rule_id="paradigm_constraint.variable_output_flag",
+                        field_path=f"outputs[{i}].data_dependent_shape",
+                        message="VariableOutput 输出必须配 data_dependent_shape: true",
+                        suggested_fix="在该 output 加 data_dependent_shape: true，或若输出个数由属性决定则在 op 声明 output_count_determined_by: attribute",
+                    ))
+
+    if "RandomSampling" in paradigms:
+        attr_names = {a.get("name") for a in (spec.get("attributes") or [])}
+        if "seed" not in attr_names:
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.random_sampling_seed",
+                field_path="attributes",
+                message="RandomSampling 必须含 seed 属性",
+                suggested_fix="在 attributes 中添加 seed (int64)",
+            ))
+
+    if "CollectiveCommunication" in paradigms:
+        if not _get(spec, "op", "platform_constraints", "requires_hccl"):
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.collective_hccl",
+                field_path="op.platform_constraints.requires_hccl",
+                message="CollectiveCommunication 必须声明 op.platform_constraints.requires_hccl: true",
+                suggested_fix="在 op 下添加 platform_constraints: {requires_hccl: true}",
+            ))
+
+
+def _check_paradigm_groups(spec, paradigms, findings):
+    """paradigm_groups 校验：
+    - 所有组的 paradigms 并集必须是 op.paradigms 的子集（未分组的视为修饰符）
+    - fusion 组内 ≥ 2 基础范式（Elementwise 不计入）
+    - combination 组必须有 switch + when
+    - 当有 ≥ 2 个基础范式且 paradigm_groups 缺失时，提示应声明组合关系
+    - 当存在模式切换属性（string_in/enum_in）但 paradigm_groups 缺失时，提示检查
+    """
+    groups = _get(spec, "op", "paradigm_groups", default=[]) or []
+
+    # ── 通用信号检测：模式切换属性 ────────────────────────────────────────
+    # string_in 和 enum_in 语义等价，都表示属性在多个离散值之间切换。
+    # int_in_range 且 lower_inclusive=0 也视为模式切换信号（0 值通常对应
+    # none/identity/passthrough，如 reduction=0→none, mode=0→求值 等）。
+    _MODE_SWITCH_KINDS = {"string_in", "enum_in"}
+    attrs = spec.get("attributes") or []
+    mode_switch_attr = next(
+        (a for a in attrs
+         if (a.get("machine_constraint") or {}).get("kind") in _MODE_SWITCH_KINDS),
+        None,
+    )
+    # 扩展检测：int_in_range + lower=0 也暗示模式切换
+    if mode_switch_attr is None:
+        mode_switch_attr = next(
+            (a for a in attrs
+             if (a.get("machine_constraint") or {}).get("kind") == "int_in_range"
+             and (a.get("machine_constraint") or {}).get("lower_inclusive", 1) == 0),
+            None,
+        )
+
+    # ── combination_should_exist ──────────────────────────────────────────
+    # 修饰符范式（不参与计算主路径，不强制要求分组）
+    _MODIFIER_PARADIGMS = {
+        "NumericalStable", "FusedComposite", "Quantization",
+        "RandomSampling", "DynamicShape", "VariableOutput",
+        "Stateful", "Sparse", "CollectiveCommunication",
+    }
+    base_paradigms = paradigms - _MODIFIER_PARADIGMS
+
+    if not groups:
+        if len(base_paradigms) >= 2:
+            # 情况 A：≥2 基础范式 + 无 groups → 明确需要 paradigm_groups
+            findings.append(Finding(
+                severity="warning",
+                rule_id="paradigm_groups.combination_should_exist",
+                field_path="op.paradigm_groups",
+                message=(
+                    f"op.paradigms 包含 ≥ 2 个基础范式 {sorted(base_paradigms)} "
+                    f"但未声明 paradigm_groups。"
+                    + (f"属性 {mode_switch_attr.get('name')!r} 是模式切换属性"
+                       f"（{mode_switch_attr.get('machine_constraint', {}).get('kind')}），"
+                       f"强烈暗示横向组合（combination）。"
+                       if mode_switch_attr else "")
+                    + " 建议声明 paradigm_groups 明确范式之间的关系"
+                      "（combination=横向组合 / fusion=纵向融合）"
+                ),
+                suggested_fix=(
+                    "添加 paradigm_groups，参考 stage-rules.md 中的 mse_loss 横向组合示例"
+                ),
+            ))
+        elif mode_switch_attr and "Elementwise" not in paradigms:
+            # 情况 B：只有 1 个基础范式，但有模式切换属性，且 Elementwise 未在 paradigms 中
+            # 典型场景：mse_loss 的 reduction=none 使算子退化为 Elementwise，
+            # 但 spec 作者只看到了 Reduction 范式
+            mc = mode_switch_attr.get("machine_constraint") or {}
+            mc_kind = mc.get("kind", "")
+            values = mc.get("values") or []
+            none_like = [v for v in values if str(v).lower() in ("none", "identity", "passthrough")]
+            # int_in_range + lower=0：值 0 通常对应 none/identity 模式
+            if mc_kind == "int_in_range" and mc.get("lower_inclusive", 1) == 0:
+                none_like = [0]
+            if none_like:
+                findings.append(Finding(
+                    severity="warning",
+                    rule_id="paradigm_groups.mode_switch_may_need_combination",
+                    field_path="op.paradigm_groups",
+                    message=(
+                        f"属性 {mode_switch_attr.get('name')!r} 存在模式切换值 "
+                        f"{none_like}（kind={mc.get('kind')}），"
+                        f"该值可能使算子退化为 Elementwise 等非归约行为，"
+                        f"但 paradigms 中未包含对应范式且无 paradigm_groups。"
+                        f" 建议检查该属性值下的计算模式，必要时添加 Elementwise 范式"
+                        f"并声明 paradigm_groups combination"
+                    ),
+                    suggested_fix=(
+                        f"分析 {mode_switch_attr.get('name')}={none_like[0]} 时的计算行为；"
+                        f"若为逐元素输出，添加 Elementwise 到 paradigms 并声明 "
+                        f"paradigm_groups combination（参考 stage-rules.md mse_loss 示例）"
+                    ),
+                ))
+
+    if not groups:
+        return
+
+    basic = {"Broadcast", "Reduction", "Contraction", "ArgReduce", "LayoutTransform"}
+    all_grouped = set()
+
+    for i, group in enumerate(groups):
+        kind = group.get("kind")
+        group_paradigms = set(group.get("paradigms") or [])
+        all_grouped.update(group_paradigms)
+
+        if kind == "fusion":
+            basic_count = len(group_paradigms & basic)
+            if basic_count < 2:
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="paradigm_groups.fusion_min_basic",
+                    field_path=f"op.paradigm_groups[{i}].paradigms",
+                    message=f"fusion 组必须含 ≥ 2 条基础范式（Elementwise 不计入），当前 {basic_count} 条",
+                    suggested_fix=f"从 {sorted(basic)} 中至少选 2 条加入该组 paradigms",
+                ))
+        elif kind == "combination":
+            if not group.get("switch"):
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="paradigm_groups.combination_missing_switch",
+                    field_path=f"op.paradigm_groups[{i}].switch",
+                    message="combination 组必须有 switch 字段",
+                    suggested_fix="添加 switch: <属性名>",
+                ))
+            if "when" not in group:
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="paradigm_groups.combination_missing_when",
+                    field_path=f"op.paradigm_groups[{i}].when",
+                    message="combination 组必须有 when 字段",
+                    suggested_fix="添加 when: <属性值>",
+                ))
+
+    extra = all_grouped - paradigms
+    if extra:
+        findings.append(Finding(
+            severity="error",
+            rule_id="paradigm_groups.mismatch",
+            field_path="op.paradigm_groups",
+            message=f"paradigm_groups 包含 op.paradigms 中没有的 {sorted(extra)}",
+            suggested_fix="确保 paradigm_groups 中的 paradigms 都在 op.paradigms 中",
+        ))
+
+
+def _check_fused_composite(spec, primitives_wl, findings):
+    """FusedComposite 子项：composition 必填、primitives ≥ 2、白名单、不泄漏、闭合。"""
+    comp = _get(spec, "math_semantics", "composition")
+    if comp is None:
+        findings.append(Finding(
+            severity="error",
+            rule_id="paradigm_constraint.fused_composite_composition_missing",
+            field_path="math_semantics.composition",
+            message="paradigms 含 FusedComposite ⇒ math_semantics.composition 必填",
+            suggested_fix="添加 composition.primitives (≥2 条) 与 dataflow",
+        ))
+        return
+
+    prims = comp.get("primitives", []) or []
+    if len(prims) < 2:
+        findings.append(Finding(
+            severity="error",
+            rule_id="paradigm_constraint.fused_composite_min_primitives",
+            field_path="math_semantics.composition.primitives",
+            message=f"FusedComposite 要求 primitives 长度 ≥ 2（当前 {len(prims)}）",
+            suggested_fix="拆分成至少 2 条白名单内原语",
+        ))
+    for i, prim in enumerate(prims):
+        if prim.get("op") not in primitives_wl:
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.primitive_not_whitelisted",
+                field_path=f"math_semantics.composition.primitives[{i}].op",
+                message=f"原语 op={prim.get('op')!r} 不在 PRIMITIVE_WHITELIST",
+                suggested_fix=f"改为白名单内之一: {sorted(primitives_wl)}",
+            ))
+
+    df = comp.get("dataflow") or {}
+    if df.get("no_leak"):
+        output_names = {o["name"] for o in spec.get("outputs", [])}
+        leaked = set(df.get("intermediates") or []) & output_names
+        if leaked:
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.intermediate_leak",
+                field_path="math_semantics.composition.dataflow.intermediates",
+                message=f"中间 tensor 泄漏到 outputs: {sorted(leaked)}",
+                suggested_fix="把这些名字从 intermediates 移除或重命名 outputs",
+            ))
+
+    # dataflow 闭合：每个 prim.input 要么是 spec.inputs，要么是先前 prim.outputs
+    produced = {inp["name"] for inp in spec.get("inputs", [])}
+    for i, prim in enumerate(prims):
+        for x in (prim.get("inputs") or []):
+            if x not in produced:
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="paradigm_constraint.dataflow_unclosed",
+                    field_path=f"math_semantics.composition.primitives[{i}].inputs",
+                    message=f"未定义的中间 tensor: {x!r}",
+                    suggested_fix=f"先在前序 primitive 的 outputs 中产出 {x}",
+                ))
+        produced.update(prim.get("outputs") or [])
+
+
+def _check_invariants(spec, registries, findings):
+    """D. invariants[].kind 白名单 + required_fields + tolerance_inherit 约束。"""
+    invariant_kinds = registries["invariant_kinds"]
+    invariants = _get(spec, "math_semantics", "invariants", default=[]) or []
+    for i, inv in enumerate(invariants):
+        kind = inv.get("kind")
+        info = invariant_kinds.get(kind)
+        if info is None:
+            findings.append(Finding(
+                severity="error",
+                rule_id="invariant_kind_resolved.unknown_kind",
+                field_path=f"math_semantics.invariants[{i}].kind",
+                message=f"未知 invariant kind: {kind!r}",
+                suggested_fix="使用 registries/invariant_kind_registry.yaml 中的 kind",
+            ))
+            continue
+        for f in info.get("required_fields", []):
+            if f not in inv:
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="invariant_kind_resolved.missing_field",
+                    field_path=f"math_semantics.invariants[{i}].{f}",
+                    message=f"invariant kind={kind} 缺必填字段 {f!r}",
+                    suggested_fix=f"补充 {f}",
+                ))
+        if info.get("forbid_tolerance_inherit") and inv.get("tolerance_inherit") is True:
+            findings.append(Finding(
+                severity="error",
+                rule_id="invariant_kind_resolved.tolerance_inherit_forbidden",
+                field_path=f"math_semantics.invariants[{i}].tolerance_inherit",
+                message=f"kind={kind} 不允许 tolerance_inherit: true（结构性/离散）",
+                suggested_fix="改为 false 或删除该字段",
+            ))
+        if info.get("require_tolerance_inherit") and not inv.get("tolerance_inherit"):
+            findings.append(Finding(
+                severity="warning",
+                rule_id="invariant_kind_resolved.tolerance_inherit_required",
+                field_path=f"math_semantics.invariants[{i}].tolerance_inherit",
+                message=f"kind={kind} 建议 tolerance_inherit: true（浮点累加）",
+                suggested_fix="添加 tolerance_inherit: true",
+            ))
+
+
+def _check_machine_check(spec, registries, findings):
+    """E. boundary/extreme 的 machine_check.kind 白名单 + raises_error.error_type 枚举。"""
+    machine_check_kinds = set(registries["machine_check_kinds"])
+    error_codes = set(registries["error_codes"])
+    for sect in ("boundary_conditions", "extreme_inputs"):
+        for i, c in enumerate(spec.get(sect) or []):
+            mc = c.get("machine_check") or {}
+            kind = mc.get("kind")
+            if kind not in machine_check_kinds:
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="machine_check_kind_unknown",
+                    field_path=f"{sect}[{i}].machine_check.kind",
+                    message=f"未知 machine_check.kind: {kind!r}",
+                    suggested_fix="使用 registries/machine_check_kind_registry.yaml 中的 kind",
+                ))
+            elif kind == "raises_error":
+                etype = mc.get("error_type")
+                if etype not in error_codes:
+                    findings.append(Finding(
+                        severity="error",
+                        rule_id="error_type_unknown",
+                        field_path=f"{sect}[{i}].machine_check.error_type",
+                        message=f"未知 error_type: {etype!r}",
+                        suggested_fix=f"使用 registries/error_code_enum.yaml 中的枚举值",
+                    ))
+
+
+_PATTERN_RE = re.compile(r"^(?P<name>[a-z_]+)(\([^()]*\))?$")
+
+
+def _check_synthesize_patterns(spec, registries, findings):
+    """F. extreme_inputs.synthesize.patterns[].pattern 白名单（兼容旧 pattern: 顶层写法）。"""
+    pattern_names = set(registries["patterns"])
+    for i, c in enumerate(spec.get("extreme_inputs") or []):
+        synth = c.get("synthesize") or {}
+        plist: list[tuple[int | None, str | None]] = []
+        if "patterns" in synth and isinstance(synth["patterns"], list):
+            plist = [(idx, p.get("pattern")) for idx, p in enumerate(synth["patterns"])]
+        elif "pattern" in synth:
+            plist = [(None, synth["pattern"])]
+            findings.append(Finding(
+                severity="warning",
+                rule_id="synthesize_legacy_format",
+                field_path=f"extreme_inputs[{i}].synthesize",
+                message="旧 'pattern:' 顶层写法已过时（v1 兼容期 90 天），建议迁移到 'patterns: [{pattern, target}]'",
+                suggested_fix="改为 patterns: [{pattern: <name>, target: <input_name>}, ...]",
+            ))
+        for idx, pname in plist:
+            if not pname:
+                continue
+            m = _PATTERN_RE.match(pname)
+            if not m or m.group("name") not in pattern_names:
+                fp = (f"extreme_inputs[{i}].synthesize.patterns[{idx}].pattern"
+                      if idx is not None else f"extreme_inputs[{i}].synthesize.pattern")
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="synthesize_pattern_unknown",
+                    field_path=fp,
+                    message=f"未知 pattern: {pname!r}",
+                    suggested_fix=f"使用 registries/synthesize_pattern_registry.yaml 中的 name",
+                ))
+
+
+def _check_supported_chips(spec, registries, findings):
+    """G. 平台兼容性校验。
+
+    若 spec 声明 supported_chips（可选），其中任一芯片必须支持所有 inputs.dtype_set
+    + supported_combinations 涉及的 dtype。
+    """
+    chip_ids = _get(spec, "op", "platform_constraints", "supported_chips", default=None)
+    if not chip_ids:
+        return
+    chips = registries.get("chips") or {}
+    if not chips:
+        return  # registry 缺失走 schema-only 校验，跳过此项
+
+    # 收集 spec 涉及的所有 dtype
+    declared_dtypes: set = set()
+    for inp in spec.get("inputs") or []:
+        declared_dtypes.update(inp.get("dtype_set") or [])
+    for combo in (spec.get("dtype_policy") or {}).get("supported_combinations") or []:
+        for d in (combo.get("inputs") or {}).values():
+            declared_dtypes.add(d)
+        for d in (combo.get("outputs") or {}).values():
+            declared_dtypes.add(d)
+
+    for chip_id in chip_ids:
+        chip = chips.get(chip_id)
+        if chip is None:
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.unknown_chip",
+                field_path="op.platform_constraints.supported_chips",
+                message=f"未注册的芯片 id: {chip_id!r}",
+                suggested_fix="在 registries/chip_registry.yaml 登记，或修正拼写",
+            ))
+            continue
+        unsupported = declared_dtypes - chip["dtypes"]
+        if unsupported:
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.dtype_chip_mismatch",
+                field_path="op.platform_constraints.supported_chips",
+                message=(
+                    f"芯片 {chip_id!r} 不支持以下 dtype: {sorted(unsupported)}；"
+                    f"已知该芯片支持: {sorted(chip['dtypes'])}"
+                ),
+                suggested_fix=(
+                    "从 supported_chips 移除不兼容的型号，或从 dtype_set / "
+                    "supported_combinations 移除该芯片不支持的 dtype"
+                ),
+            ))
+
+
+def _check_anti_pattern_ids(spec, registries, findings):
+    """I. numerical_stability.techniques[].anti_pattern_id 必须在 anti_pattern_registry 中。
+
+    schema 已用 ^AP-\\d{3}$ 卡格式；本检查锁定**有效编号集合**，避免 AP-999 / AP-001
+    （未分配）混入 spec。
+    """
+    valid = registries.get("anti_patterns") or set()
+    if not valid:
+        return  # registry 缺失走 schema-only 校验，跳过此项
+    techniques = _get(spec, "numerical_stability", "techniques", default=[]) or []
+    for i, tech in enumerate(techniques):
+        ap_id = tech.get("anti_pattern_id")
+        if ap_id is None:
+            continue
+        if ap_id not in valid:
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.unknown_anti_pattern",
+                field_path=f"numerical_stability.techniques[{i}].anti_pattern_id",
+                message=(
+                    f"未注册的 anti_pattern_id: {ap_id!r}；已知 IDs={sorted(valid)}"
+                ),
+                suggested_fix=(
+                    "在 registries/anti_pattern_registry.yaml 中登记新编号，或修正拼写"
+                ),
+            ))
+
+
+def _check_op_error_codes(spec, registries, findings):
+    """J. boundary_conditions[].machine_check.error_type 必须在 op.error_codes 全集内。
+
+    若 spec 没声明 op.error_codes 但在 boundary/extreme 中使用了 raises_error，
+    报 error 强制声明（error_codes 是算子契约的一部分，不是可选描述）。
+    """
+    declared_raw = _get(spec, "op", "error_codes", default=None)
+    declared_set = set(declared_raw or [])
+
+    raises_cases: list[tuple[str, int, str]] = []
+    for sect in ("boundary_conditions", "extreme_inputs"):
+        for i, case in enumerate(spec.get(sect) or []):
+            mc = case.get("machine_check") or {}
+            if mc.get("kind") != "raises_error":
+                continue
+            etype = mc.get("error_type")
+            if etype:
+                raises_cases.append((sect, i, etype))
+
+    if not declared_set and raises_cases:
+        etypes_used = sorted({e for _, _, e in raises_cases})
+        findings.append(Finding(
+            severity="error",
+            rule_id="paradigm_constraint.error_codes_undeclared",
+            field_path="op.error_codes",
+            message=(
+                f"boundary/extreme 中使用了 raises_error (error_type={etypes_used}) "
+                f"但 op.error_codes 未声明；error_codes 是算子契约，必须显式列全集"
+            ),
+            suggested_fix=f"在 op 块下添加 error_codes: {etypes_used}",
+        ))
+        return
+
+    for sect, i, etype in raises_cases:
+        if etype not in declared_set:
+            findings.append(Finding(
+                severity="error",
+                rule_id="paradigm_constraint.error_code_not_declared",
+                field_path=f"{sect}[{i}].machine_check.error_type",
+                message=(
+                    f"error_type={etype!r} 不在 op.error_codes 声明的全集 "
+                    f"{sorted(declared_set)} 中"
+                ),
+                suggested_fix="把该 error_type 加到 op.error_codes，或修正本 case",
+            ))
+
+
+def _check_composition_hint(spec, paradigms, findings):
+    """H. composition 存在但 paradigms 缺 FusedComposite ⇒ WARN（不算 FAIL）。"""
+    has_comp = _get(spec, "math_semantics", "composition") is not None
+    if not has_comp or "FusedComposite" in paradigms:
+        return
+    prims = _get(spec, "math_semantics", "composition", "primitives", default=[]) or []
+    if len(prims) >= 2:
+        findings.append(Finding(
+            severity="warning",
+            rule_id="composition_without_fused_composite",
+            field_path="op.paradigms",
+            message="composition.primitives ≥ 2 但 paradigms 不含 FusedComposite，建议补 FusedComposite",
+            suggested_fix="在 op.paradigms 添加 FusedComposite",
+        ))
+    elif len(prims) == 1:
+        findings.append(Finding(
+            severity="warning",
+            rule_id="composition_single_primitive",
+            field_path="math_semantics.composition",
+            message="composition.primitives 只有 1 条，建议删除 composition 段",
+            suggested_fix="删除 math_semantics.composition",
+        ))
+
+
+def _check_format_variants(spec, findings):
+    """format_variants 合法性：rank 范围、reduction_axes 合法、oracle_kwargs 与 axes 对齐。"""
+    variants = _get(spec, "math_semantics", "format_variants")
+    if not variants:
+        return
+
+    # 从 shape_constraints.symbols 推算最大 rank
+    symbols = _get(spec, "shape_constraints", "symbols", default={}) or {}
+    # inputs 的 rank_range 是更准确的来源
+    inputs = _get(spec, "inputs", default=[]) or []
+    max_rank = 0
+    for inp in inputs:
+        if isinstance(inp, dict):
+            rr = inp.get("rank_range")
+            if rr and isinstance(rr, list) and len(rr) == 2:
+                max_rank = max(max_rank, rr[1])
+    # 回退：按 symbols 数量推算
+    if max_rank == 0 and symbols:
+        max_rank = len(symbols)
+
+    for i, v in enumerate(variants):
+        if not isinstance(v, dict):
+            continue
+        prefix = f"math_semantics.format_variants[{i}]"
+
+        # 1. reduction_axes 合法性：非负且 < min(ranks)
+        #    同一 variant 的 reduction_axes 必须对所有适用 rank 都合法
+        axes = v.get("reduction_axes", [])
+        ranks = v.get("rank", [])
+        variant_max_rank = min(ranks) if ranks else max_rank
+
+        for ax in axes:
+            if not isinstance(ax, int):
+                continue
+            if ax < 0:
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="format_variants.reduction_axes_negative",
+                    field_path=f"{prefix}.reduction_axes",
+                    message=f"reduction_axes 含负值 {ax}（{v.get('format', '?')} 格式）",
+                    suggested_fix="使用非负轴索引",
+                ))
+            elif variant_max_rank > 0 and ax >= variant_max_rank:
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="format_variants.reduction_axes_out_of_rank",
+                    field_path=f"{prefix}.reduction_axes",
+                    message=(f"reduction_axes 含 {ax}，超出 {v.get('format', '?')} "
+                             f"格式的 rank={variant_max_rank}"),
+                    suggested_fix=f"确保所有 reduction_axes < {variant_max_rank}",
+                ))
+
+        # 2. oracle_kwargs.dim 与 reduction_axes 对齐
+        oracle_kw = v.get("oracle_kwargs")
+        if oracle_kw and isinstance(oracle_kw, dict):
+            dim_val = oracle_kw.get("dim")
+            if dim_val is not None and isinstance(dim_val, list):
+                if sorted(dim_val) != sorted(axes):
+                    findings.append(Finding(
+                        severity="warning",
+                        rule_id="format_variants.oracle_kwargs_dim_mismatch",
+                        field_path=f"{prefix}.oracle_kwargs",
+                        message=(f"{v.get('format', '?')} 格式：oracle_kwargs.dim={dim_val} "
+                                 f"与 reduction_axes={axes} 不一致"),
+                        suggested_fix=f"将 oracle_kwargs.dim 改为 {axes}",
+                    ))
+
+
+# 涉及累加的 category —— 这些算子的输入数据分布必须是正态分布（normal），
+# 因为均匀分布在累加链路上容易掩盖精度问题（值域窄、无大数吃小数），正态分布
+# 能暴露累加顺序敏感的精度缺陷。
+_ACCUMULATION_CATEGORIES = {
+    "Reduction", "ReductionComposite",
+    "Recurrence", "AtomicUpdate", "Histogram",
+}
+
+
+def _check_data_distribution(spec, cat, paradigms, findings):
+    """data_distribution 约束：
+      - 累加类算子（Reduction/ReductionComposite/Recurrence/AtomicUpdate/Histogram）⇒
+        tensor role 的数据输入必须声明 data_distribution: normal
+        （dtype_set 全为整型的索引/轴输入跳过，不适用正态分布）
+        注：Contraction（matmul）不计入累加类——矩阵乘法精度取决于条件数而非累加顺序
+      - 非累加类算子 ⇒ 不应声明 data_distribution（留空 / 不写）
+    """
+    is_accum = cat in _ACCUMULATION_CATEGORIES
+    inputs = spec.get("inputs") or []
+    _integer_dtypes = {"int4", "int8", "int16", "int32", "int64",
+                       "uint1", "uint4", "uint8", "uint16", "uint32", "uint64", "bool"}
+
+    for i, inp in enumerate(inputs):
+        if not isinstance(inp, dict):
+            continue
+        # 只检查 tensor role 的输入（跳过 scalar / attribute_alias / state）
+        if inp.get("role") != "tensor":
+            continue
+        # 跳过整型索引/轴输入（如 axis/dim/axes/index）—— 正态分布不适用于索引值
+        dtype_set = set(inp.get("dtype_set") or [])
+        if dtype_set and dtype_set.issubset(_integer_dtypes):
+            continue
+        dist = inp.get("data_distribution")
+        prefix = f"inputs[{i}]({inp.get('name', '?')}).data_distribution"
+
+        if is_accum:
+            if dist != "normal":
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="data_distribution.accumulation_requires_normal",
+                    field_path=prefix,
+                    message=(
+                        f"category={cat} 属累加类算子，tensor 数据输入必须声明 "
+                        f"data_distribution: normal（当前为 {dist!r}）。"
+                        "正态分布能暴露累加顺序敏感的精度缺陷，均匀分布会掩盖"
+                    ),
+                    suggested_fix="添加 data_distribution: normal",
+                ))
+        else:
+            if dist is not None:
+                findings.append(Finding(
+                    severity="error",
+                    rule_id="data_distribution.non_accumulation_must_omit",
+                    field_path=prefix,
+                    message=(
+                        f"category={cat} 非累加类算子，不应声明 data_distribution"
+                        f"（当前为 {dist!r}）。data_distribution 仅累加类算子填写"
+                    ),
+                    suggested_fix="删除该字段的 data_distribution",
+                ))
+
+
+def stage_6(spec: dict, registries: dict) -> StageResult:
+    """boundary_min_set — 按 paradigms 检查 spec 是否覆盖各范式必含的最低 case 集。
+
+    数据源：registries/boundary_min_cases.yaml
+    匹配方式（双轨）：
+      1. 显式 tag 优先：spec 的 case 可声明 `tags: [<requirement.id>, ...]`，命中
+         即覆盖。新写 spec 推荐做法，避免子串歧义。
+      2. 关键词子串兜底：req.match_any 任一关键词出现在 case 描述里也算覆盖。保留兼容。
+    结构性约束（如 DynamicShape 的 shape_set ≥ 3）走 special_check。
+    """
+    res = StageResult(stage_id=6, status="PASS")
+    paradigms = set(_get(spec, "op", "paradigms", default=[]) or [])
+    requirements = (registries.get("boundary_min_cases") or {}).get("paradigm_requirements", {})
+
+    boundary_cases = spec.get("boundary_conditions") or []
+    extreme_cases = spec.get("extreme_inputs") or []
+
+    def _section_data(section: str) -> tuple[str, set]:
+        cases = boundary_cases if section == "boundary_conditions" else extreme_cases
+        text = " | ".join(str(c.get("case", "")) for c in cases)
+        tags: set = set()
+        for c in cases:
+            for t in (c.get("tags") or []):
+                tags.add(str(t))
+        return text, tags
+
+    broadcast_kind = (spec.get("broadcast") or {}).get("kind", "none")
+
+    _attr_determined_output = _output_count_attr_determined(spec)
+    for paradigm in sorted(paradigms):
+        # Broadcast paradigm 的 boundary cases 仅适用于输入间广播（kind != none）；
+        # 单输入或内部计算广播（kind=none）不需要这些 case。
+        if paradigm == "Broadcast" and broadcast_kind == "none":
+            continue
+        # 属性决定输出个数（split 类）：VariableOutput 的 empty/full/sparse case
+        # 是为 nonzero/where 这类数据决定输出个数的算子设计的，不适用于属性决定。
+        if paradigm == "VariableOutput" and _attr_determined_output:
+            continue
+
+        for req in requirements.get(paradigm, []) or []:
+            req_id = req.get("id")
+            summary = req.get("summary", req_id)
+
+            if "special_check" in req:
+                # 历史遗留特殊校验已随弃用字段（test_matrix 等）移除；
+                # registry 里若还残留 special_check，直接跳过，不再产生 finding。
+                continue
+
+            section = req.get("section", "boundary_conditions")
+            text, tags = _section_data(section)
+            keywords = req.get("match_any", []) or []
+
+            if req_id in tags:
+                continue   # 显式 tag 命中
+            if any(kw in text for kw in keywords):
+                continue   # 子串兜底命中
+
+            res.findings.append(Finding(
+                severity="error",
+                rule_id="boundary_min_set.missing_required_case",
+                field_path=f"{section}",
+                message=f"paradigm={paradigm} 缺必含 case: {summary} (id={req_id})",
+                suggested_fix=(
+                    f"在 {section} 中添加一条 case，加 `tags: [{req_id}]` "
+                    f"或描述包含以下任一关键词：{keywords}"
+                ),
+            ))
+
+    if any(f.severity == "error" for f in res.findings):
+        res.status = "FAIL"
+    return res
+
+
+# 容差紧度阈值从 registries/tolerance_defaults.yaml 加载（与 generator 共享数据源）
+@functools.lru_cache(maxsize=1)
+def _tolerance_tight_threshold() -> dict[str, float]:
+    data = yaml.safe_load((REGISTRIES / "tolerance_defaults.yaml").read_text(encoding="utf-8"))
+    return data.get("tightness_threshold") or {}
+
+
+def stage_7(spec: dict) -> StageResult:
+    """tolerance_coverage — numerical_tolerance.per_dtype 是否覆盖所有声明的输出 dtype。
+
+    与 stage 4 (dtype DSL 求解) 完整版的差别：本 stage 用 dtype_policy.supported_combinations
+    显式枚举的输出 dtype 作为代理，stage 4 实现后会改为用 DSL 推导结果。两者通常一致。
+
+    额外做"容差过紧"启发式检查 (WARN)：rtol 显著低于 dtype 单步舍入精度时提示。
+    """
+    res = StageResult(stage_id=7, status="PASS")
+
+    declared_output_dtypes: set[str] = set()
+    for combo in _get(spec, "dtype_policy", "supported_combinations", default=[]) or []:
+        for d in (combo.get("outputs") or {}).values():
+            declared_output_dtypes.add(d)
+
+    per_dtype = _get(spec, "numerical_tolerance", "per_dtype", default={}) or {}
+    declared_tol_dtypes = set(per_dtype.keys())
+
+    uncovered = declared_output_dtypes - declared_tol_dtypes
+    for d in sorted(uncovered):
+        res.findings.append(Finding(
+            severity="error",
+            rule_id="tolerance_coverage.uncovered_output_dtype",
+            field_path=f"numerical_tolerance.per_dtype.{d}",
+            message=f"输出 dtype {d!r} 未在 numerical_tolerance.per_dtype 中声明容差",
+            suggested_fix=f"为 {d} 补充 {{rtol, atol, metric}}",
+        ))
+
+    # 紧度提示（启发式 WARN）
+    thresholds = _tolerance_tight_threshold()
+    for d, entry in per_dtype.items():
+        threshold = thresholds.get(d)
+        if threshold is None:
+            continue
+        rtol = entry.get("rtol")
+        if isinstance(rtol, (int, float)) and 0 < rtol < threshold:
+            res.findings.append(Finding(
+                severity="warning",
+                rule_id="tolerance_coverage.tolerance_too_tight",
+                field_path=f"numerical_tolerance.per_dtype.{d}.rtol",
+                message=(
+                    f"{d} rtol={rtol:.1e} 低于该 dtype 单步舍入量级 ({threshold:.0e})；"
+                    "大量累加后实际不可能达到，可能导致 PR 误报"
+                ),
+                suggested_fix=f"放宽 {d} 的 rtol 到至少 {threshold:.0e}",
+            ))
+
+    if any(f.severity == "error" for f in res.findings):
+        res.status = "FAIL"
+    return res
+
+
+@functools.lru_cache(maxsize=1)
+def load_registries() -> dict:
+    """Load all registries from disk; cached per-process."""
+    inv = {}
+    inv_data = yaml.safe_load((REGISTRIES / "invariant_kind_registry.yaml").read_text(encoding="utf-8"))
+    for group in ("value", "algebraic", "structural"):
+        for entry in inv_data.get(group, []) or []:
+            inv[entry["kind"]] = entry
+
+    return {
+        "primitives": yaml.safe_load((REGISTRIES / "primitive_whitelist.yaml").read_text(encoding="utf-8"))["primitives"],
+        "invariant_kinds": inv,
+        "machine_check_kinds": [
+            x["kind"] for x in
+            yaml.safe_load((REGISTRIES / "machine_check_kind_registry.yaml").read_text(encoding="utf-8"))["machine_check_kinds"]
+        ],
+        "patterns": [
+            x["name"] for x in
+            yaml.safe_load((REGISTRIES / "synthesize_pattern_registry.yaml").read_text(encoding="utf-8"))["patterns"]
+        ],
+        "error_codes": yaml.safe_load((REGISTRIES / "error_code_enum.yaml").read_text(encoding="utf-8"))["error_codes"],
+        "category_map": yaml.safe_load((REGISTRIES / "category_paradigm_map.yaml").read_text(encoding="utf-8")),
+        "boundary_min_cases": yaml.safe_load((REGISTRIES / "boundary_min_cases.yaml").read_text(encoding="utf-8")),
+        "chips": _load_chip_registry(),
+        "anti_patterns": _load_anti_pattern_registry(),
+    }
+
+
+def _load_anti_pattern_registry() -> set:
+    """加载 anti_pattern_registry.yaml；返回 {AP-XXX, ...}。"""
+    path = REGISTRIES / "anti_pattern_registry.yaml"
+    if not path.exists():
+        return set()
+    data = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
+    return {x["id"] for x in data.get("anti_patterns") or []}
+
+
+def _load_chip_registry() -> dict:
+    """加载 chip_registry.yaml；返回 {chip_id: {dtypes: set}}。"""
+    path = REGISTRIES / "chip_registry.yaml"
+    if not path.exists():
+        return {}
+    data = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
+    out = {}
+    for c in data.get("chips") or []:
+        out[c["id"]] = {
+            "dtypes": set(c.get("supported_dtypes") or []),
+        }
+    return out
+
+
+def _run_eval_stages(spec: dict) -> tuple[StageResult, ...]:
+    """Lazily import the evaluators package; if it fails, return SKIP stages.
+
+    Returns eight StageResult objects in order: stage 3, 4, 5, 8, 9, 10, 11, 12.
+    Stage 8 / 9 / 10 / 11 / 12 may individually SKIP based on env (numpy/framework not
+    installed) even when the evaluators package itself imports fine.
+    """
+    try:
+        from evaluators import stages as eval_stages
+        from evaluators import (formula_eval, oracle_check, formula_oracle_equiv,
+                                invariant_exec, extreme_inputs_check)
+    except ImportError as e:
+        skip_finding = Finding(
+            severity="info",
+            rule_id="stage_skipped",
+            field_path="<evaluators import>",
+            message=f"evaluators 子包不可用，stage 3-5/8/9/10/11/12 跳过：{e}",
+            suggested_fix="确认 scripts/evaluators/ 子包文件齐全",
+        )
+        return tuple(
+            StageResult(stage_id=sid, status="SKIP", findings=[skip_finding])
+            for sid in (3, 4, 5, 8, 9, 10, 11, 12)
+        )
+
+    pipeline = (
+        (3, eval_stages.stage_3),
+        (4, eval_stages.stage_4),
+        (5, eval_stages.stage_5),
+        (8, formula_eval.stage_8),
+        (9, oracle_check.stage_9),
+        (10, formula_oracle_equiv.stage_10),
+        (11, invariant_exec.stage_11),
+        (12, extreme_inputs_check.stage_12),
+    )
+    results: list[StageResult] = []
+    for stage_id, fn in pipeline:
+        try:
+            status, raw_findings = fn(spec)
+        except Exception as ex:  # 防御性 — DSL 内部 bug 不应整体打断
+            results.append(StageResult(
+                stage_id=stage_id,
+                status="FAIL",
+                findings=[Finding(
+                    severity="error",
+                    rule_id=f"stage_{stage_id}_internal_error",
+                    field_path="<internal>",
+                    message=f"stage {stage_id} 内部异常：{type(ex).__name__}: {ex}",
+                    suggested_fix="向 ops-spec-gen 维护者反馈完整堆栈",
+                )]
+            ))
+            continue
+        findings = [Finding(**f) for f in raw_findings]
+        results.append(StageResult(stage_id=stage_id, status=status, findings=findings))
+    return tuple(results)
+
+
+def render_text(stages: list[StageResult], *, quiet: bool = False) -> str:
+    """Render stage results as text. quiet=True 时只打 FAIL 的 stage（PASS/SKIP 折叠为一行汇总）。"""
+    lines: list[str] = []
+    icons = {"PASS": "✓", "FAIL": "✗", "SKIP": "↷"}
+
+    if quiet:
+        # 只列 FAIL，PASS/SKIP 计数
+        passed = [s for s in stages if s.status == "PASS"]
+        skipped = [s for s in stages if s.status == "SKIP"]
+        failed = [s for s in stages if s.status == "FAIL"]
+        if passed:
+            lines.append(f"  {icons['PASS']}  {len(passed)} 个 stage PASS: "
+                         + ", ".join(f"stage {s.stage_id}" for s in passed))
+        if skipped:
+            lines.append(f"  {icons['SKIP']}  {len(skipped)} 个 stage SKIP: "
+                         + ", ".join(f"stage {s.stage_id}" for s in skipped))
+        for st in failed:
+            lines.append(f"\n[stage {st.stage_id}] {icons[st.status]} {st.status}")
+            for f in st.findings:
+                if f.severity not in ("error", "warning"):
+                    continue
+                sev = {"error": "ERR ", "warning": "WARN"}[f.severity]
+                lines.append(f"  {sev}  {f.rule_id}")
+                lines.append(f"        path: {f.field_path}")
+                lines.append(f"        msg : {f.message}")
+                if f.suggested_fix:
+                    lines.append(f"        fix : {f.suggested_fix}")
+    else:
+        for st in stages:
+            lines.append(f"\n[stage {st.stage_id}] {icons[st.status]} {st.status}")
+            for f in st.findings:
+                sev = {"error": "ERR ", "warning": "WARN", "info": "INFO"}[f.severity]
+                lines.append(f"  {sev}  {f.rule_id}")
+                lines.append(f"        path: {f.field_path}")
+                lines.append(f"        msg : {f.message}")
+                if f.suggested_fix:
+                    lines.append(f"        fix : {f.suggested_fix}")
+
+    overall = (
+        "PASS"
+        if all(s.status in ("PASS", "SKIP") for s in stages)
+        else "FAIL"
+    )
+    lines.append(f"\n=== overall: {overall} ===")
+    return "\n".join(lines)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description="Validate spec.yaml (full 12-stage).")
+    ap.add_argument("spec_path")
+    ap.add_argument("--json", action="store_true", help="Emit JSON instead of text")
+    ap.add_argument("--strict", action="store_true",
+                    help="Exit non-zero on warnings as well as errors")
+    ap.add_argument("--stage", action="append", type=int, choices=range(1, 13),
+                    metavar="N",
+                    help="只跑指定 stage（可多次：--stage 1 --stage 2）；省略时跑全部 12 个")
+    ap.add_argument("--quiet", action="store_true",
+                    help="只打 FAIL 的 stage（仍输出 overall 行）")
+    args = ap.parse_args()
+
+    spec_path = Path(args.spec_path)
+    if not spec_path.exists():
+        print(f"ERROR: spec file not found: {spec_path}", file=sys.stderr)
+        return 2
+    try:
+        spec = _load_yaml(spec_path)
+    except yaml.YAMLError as e:
+        print(f"ERROR: YAML 解析失败 {spec_path}", file=sys.stderr)
+        print(f"  {e}", file=sys.stderr)
+        return 2
+    if not isinstance(spec, dict):
+        print(f"ERROR: {args.spec_path} did not parse as a YAML mapping", file=sys.stderr)
+        return 2
+
+    selected = set(args.stage) if args.stage else None
+    registries = load_registries()
+
+    # 按需跑 stage —— 跳过的 stage 不出现在 stages 列表中
+    stages: list[StageResult] = []
+    if selected is None or 1 in selected:
+        stages.append(stage_1(spec))
+    if selected is None or 2 in selected:
+        stages.append(stage_2(spec, registries))
+    # stage 3/4/5/8/9/10/11 共享 DSL 子包 lazy import；选中其一即触发
+    dsl_needed = selected is None or selected & {3, 4, 5, 8, 9, 10, 11, 12}
+    if dsl_needed:
+        eval_srs = _run_eval_stages(spec)
+        for sid, sr in zip((3, 4, 5, 8, 9, 10, 11, 12), eval_srs):
+            if selected is None or sid in selected:
+                stages.append(sr)
+    if selected is None or 6 in selected:
+        stages.append(stage_6(spec, registries))
+    if selected is None or 7 in selected:
+        stages.append(stage_7(spec))
+
+    # 按 stage_id 排序，确保打印顺序与 1..10 一致
+    stages.sort(key=lambda s: s.stage_id)
+
+    if args.json:
+        print(json.dumps({
+            "spec": str(args.spec_path),
+            "stages": [{"stage_id": s.stage_id, "status": s.status,
+                        "findings": [asdict(f) for f in s.findings]}
+                       for s in stages],
+        }, ensure_ascii=False, indent=2))
+    else:
+        print(render_text(stages, quiet=args.quiet))
+
+    has_error = any(f.severity == "error" for s in stages for f in s.findings)
+    has_warning = any(f.severity == "warning" for s in stages for f in s.findings)
+    if has_error:
+        return 1
+    if args.strict and has_warning:
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())