cann-workbench 0.0.0-stage → 0.0.2-alpha

package/skills/asc-tools/tool-ops-tune/ops-verdict/evals/smoke/install-hooks.sh

@@ -0,0 +1,65 @@
+#!/usr/bin/env bash
+# -----------------------------------------------------------------------------------------------------------
+# 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.
+# -----------------------------------------------------------------------------------------------------------
+# install-hooks.sh — 把 evals + smoke 挂成 pre-push 门禁（本地侧；CI 侧见 .workflow/）
+#
+# 用法：bash ops-verdict/evals/smoke/install-hooks.sh（在 skills/tool-ops-tune/ 下执行）
+set -euo pipefail
+# 脚本所在套件目录（三级上溯 = skills/tool-ops-tune）；git 仓库根与 git dir
+# 一律以 git 命令解析（直接把套件目录当仓库根会导致
+# hooks 写到不存在的 skills/tool-ops-tune/.git/hooks/，set -e 下安装直接失败）
+SUITE_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)"
+REPO_ROOT="$(git -C "$SUITE_ROOT" rev-parse --show-toplevel)" || {
+  echo "ERROR: 不在 git 仓库内（$SUITE_ROOT）——pre-push 门禁需要 git hooks 目录" >&2
+  exit 1
+}
+GIT_DIR="$(git -C "$SUITE_ROOT" rev-parse --absolute-git-dir)"
+HOOK="$GIT_DIR/hooks/pre-push"
+mkdir -p "$GIT_DIR/hooks"
+
+# 套件在仓内的真实路径（hook 以 git toplevel 为根解析；
+# 此前两套路径互相矛盾且都缺 tool-ops-tune 一级 → 每次 push 必被误拦）
+SUITE_DIR="skills/tool-ops-tune"
+cat > "$HOOK" <<HOOKEOF
+#!/usr/bin/env bash
+# OpsTune pre-push 门禁（install-hooks.sh 生成
+set -uo pipefail
+REPO_ROOT="\$(git rev-parse --show-toplevel)"
+SUITE="\$REPO_ROOT/${SUITE_DIR}"
+FAIL=0
+
+EVALS_PY="\$SUITE/ops-verdict/evals/run_evals.py"
+SMOKE_SH="\$SUITE/ops-verdict/evals/smoke/run_smoke.sh"
+for f in "\$EVALS_PY" "\$SMOKE_SH"; do
+  if [ ! -f "\$f" ]; then
+    echo "[pre-push] ERROR: 门禁脚本缺失: \$f（套件未完整检出？）" >&2
+    exit 1
+  fi
+done
+
+echo "[pre-push] evals（判定回归 38 用例）…"
+python3 "\$EVALS_PY" >/dev/null || {
+  echo "[pre-push] FAIL: evals 未全绿 —— 规则/阈值变更未过回归，禁止推送" >&2
+  FAIL=1
+}
+
+echo "[pre-push] smoke（离线干跑冒烟）…"
+bash "\$SMOKE_SH" >/dev/null || {
+  echo "[pre-push] FAIL: smoke 冒烟未通过 —— 禁止推送" >&2
+  FAIL=1
+}
+
+if [[ \$FAIL -ne 0 ]]; then
+  exit 1
+fi
+echo "[pre-push] 门禁全绿，放行"
+HOOKEOF
+chmod +x "$HOOK"
+echo "installed: $HOOK（pre-push 时自动跑 evals + smoke；跳过用 git push --no-verify）"

package/skills/asc-tools/tool-ops-tune/ops-verdict/evals/smoke/run_smoke.sh

@@ -0,0 +1,535 @@
+#!/usr/bin/env bash
+# -----------------------------------------------------------------------------------------------------------
+# 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.
+# -----------------------------------------------------------------------------------------------------------
+# run_smoke.sh — 离线干跑冒烟
+#
+# 目的：「bash -n 查不出、只有真跑才暴露」类 bug 在无 NPU 环境即暴露。
+# 手段：msprof shim（按参数产出预置 CSV 目录的假可执行文件）+ 三模式
+# collect_board.sh 干跑 + bash 语法检查 + isolation from-prof 干跑 +
+# 规则包加载校验 + 族表资产断言。
+#
+# 用法：bash evals/smoke/run_smoke.sh   （在仓库根执行）
+# 退出码：0 全绿 / 1 有失败（CI 门禁语义，与 evals/run_evals.py 同级）。
+
+set -euo pipefail
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../../.." && pwd)"  # skills/tool-ops-tune/
+FAILURES=0
+
+pass() { echo "[PASS] $1"; }
+fail() { echo "[FAIL] $1"; FAILURES=$((FAILURES + 1)); }
+check() { if [[ $? -eq 0 ]]; then pass "$1"; else fail "$1"; fi }
+
+echo "=== smoke 1: bash 语法检查（collect_board.sh）==="
+bash -n "$REPO_ROOT/ops-collect/scripts/collect_board.sh" 2>/dev/null
+check "bash -n collect_board.sh"
+
+echo "=== smoke 2: msprof shim 三模式干跑 ==="
+SMOKE_TMP="$(mktemp -d)"
+trap 'rm -rf "$SMOKE_TMP"' EXIT
+SHIM_BIN="$SMOKE_TMP/bin"; mkdir -p "$SHIM_BIN"
+SHIM_OUT="$SMOKE_TMP/shim_out"; mkdir -p "$SHIM_OUT"
+
+# shim：解析 --output= 参数，产出预置 op_summary CSV（Task Duration + pipe 占比）
+cat > "$SHIM_BIN/msprof" <<'SHIM'
+#!/usr/bin/env bash
+OUT=""
+for a in "$@"; do
+  case "$a" in --output=*) OUT="${a#--output=}" ;; esac
+done
+mkdir -p "$OUT/mindstudio_profiler_output"
+cat > "$OUT/mindstudio_profiler_output/op_summary_1.csv" <<'CSV'
+Task Duration(us),aic_time(us),aiv_time(us),aiv_vec_ratio,aic_cube_ratio,aic_mte2_ratio,aic_mte1_ratio,aic_fixpipe_ratio,ai_scalar_ratio,aic_mte3_ratio,ai_icache_miss_rate
+100,0,100.5,0.32,0.00,0.55,0.02,0.01,0.22,0.03,0.9
+100,0,99.8,0.31,0.00,0.56,0.02,0.01,0.21,0.03,0.9
+CSV
+SHIM
+chmod +x "$SHIM_BIN/msprof"
+
+for MODE in standard quick mc2; do
+  OUT_DIR="$SMOKE_TMP/prof_$MODE"
+  if PATH="$SHIM_BIN:$PATH" bash "$REPO_ROOT/ops-collect/scripts/collect_board.sh" \
+      --exe /bin/true --mode "$MODE" --output "$OUT_DIR" >/dev/null 2>&1 \
+      && python3 -c "
+import json, sys
+m = json.load(open('$OUT_DIR/collection_manifest.json'))
+assert m['mode'] == '$MODE'
+assert m['blocks'] and all(b['status'] == 'ok' for b in m['blocks']), m['blocks']
+" 2>/dev/null; then
+    pass "collect_board.sh --mode $MODE（shim 干跑 + manifest 全 ok）"
+  else
+    fail "collect_board.sh --mode $MODE"
+  fi
+done
+
+# ranks 守卫：--mode standard --ranks>1 必须拒绝（exit 非 0）并
+# 引导 --mode mc2——fork 多 rank 走非 task-based 采集数据不可靠。
+# 子 shell 必须 `|| true` 兜底（守卫生效时 collect_board.sh 以 exit 4 终止，
+# set -e 会连带杀掉 smoke 本体）；退出码在兜底前先行捕获。
+GUARD_RC=0
+GUARD_ERR=$(PATH="$SHIM_BIN:$PATH" bash "$REPO_ROOT/ops-collect/scripts/collect_board.sh" \
+    --exe /bin/true --mode standard --ranks 4 --output "$SMOKE_TMP/prof_guard" 2>&1) || GUARD_RC=$?
+if [[ $GUARD_RC -ne 0 ]] && grep -q -- "--mode mc2" <<< "$GUARD_ERR"; then
+  pass "collect_board.sh --mode standard --ranks 4 拒绝并引导 mc2（rc=$GUARD_RC）"
+else
+  fail "collect_board.sh ranks 守卫（rc=$GUARD_RC，stderr 未含引导）"
+fi
+
+echo "=== smoke 3: isolation from-prof 干跑（R 值档位从规则 YAML 派生）==="
+if [ ! -f "$REPO_ROOT/ops-collect/scripts/isolation_test.py" ]; then
+  echo "[SKIP] isolation_test.py 未交付（随多卡域 PR），smoke 3 暂跳过——交付后本段自动生效"
+else
+ISO_TMP="$SMOKE_TMP/iso"
+mkdir -p "$ISO_TMP/comm/PROF_x/mindstudio_profiler_output" "$ISO_TMP/compute/PROF_y/mindstudio_profiler_output"
+printf 'Task Duration(us),aic_mte2_ratio\n101,4\n102,4\n103,5\n104,5\n105,5\n106,5\n107,5\n108,5\n109,5\n110,5\n' \
+  > "$ISO_TMP/comm/PROF_x/mindstudio_profiler_output/op_summary_a.csv"
+printf 'Task Duration(us),aic_mte2_ratio\n251,40\n252,40\n253,41\n254,41\n255,41\n256,41\n257,41\n258,41\n259,41\n260,41\n' \
+  > "$ISO_TMP/compute/PROF_y/mindstudio_profiler_output/op_summary_b.csv"
+if python3 "$REPO_ROOT/ops-collect/scripts/isolation_test.py" \
+    --exe /bin/true --switch-mode from-prof \
+    --comm-prof "$ISO_TMP/comm" --compute-prof "$ISO_TMP/compute" --warmup 5 \
+    --output "$SMOKE_TMP/iso_result.json" >/dev/null 2>&1 \
+    && python3 -c "
+import json
+d = json.load(open('$SMOKE_TMP/iso_result.json'))
+assert d['bound_hint'] == 'L3_MC2_STRONG_COMPUTE', d['bound_hint']
+assert d['mte2_ratio_pure'] is not None
+assert 'mte2_ratio_pure' in d['verdict_scalar']
+" 2>/dev/null; then
+  pass "isolation_test from-prof（R≈2.4 → STRONG_COMPUTE + mte2_ratio_pure 产出）"
+else
+  fail "isolation_test from-prof"
+fi
+
+# 档界值双侧断言（检视 P1-4 守护）：开闭语义由上界叶子算子推导——
+# R=0.9/1.1（BALANCED 闭上界两侧）与 R=0.5/2.0（开上界档界值，归下一档）
+# 任一翻转即 _has_closed_upper 类硬编码回归。
+if python3 -c "
+import sys
+sys.path.insert(0, '$REPO_ROOT/ops-collect/scripts')
+from isolation_test import r_bands_from_rules, classify_r
+bands = r_bands_from_rules('$REPO_ROOT/ops-verdict/rules/L3-inter-card.yaml')
+cases = {
+    0.5:  'L3_MC2_COMM_BOUND',     # STRONG_COMM 上界 lt=开 → r==0.5 归 COMM_BOUND
+    0.9:  'L3_MC2_BALANCED',       # BALANCED 下界 gte=闭
+    1.1:  'L3_MC2_BALANCED',       # BALANCED 上界 between=闭
+    2.0:  'L3_MC2_COMPUTE_BOUND',  # STRONG_COMPUTE 下界 gt=开 → r==2.0 归 COMPUTE_BOUND
+}
+for r, want in cases.items():
+    got, _ = classify_r(r, bands)
+    assert got == want, f'R={r}: {got} != {want}'
+" 2>/dev/null; then
+  pass "isolation_test 档界值双侧（0.5/0.9/1.1/2.0 开闭语义正确）"
+else
+  fail "isolation_test 档界值双侧（开闭语义回归）"
+fi
+fi
+
+echo "=== smoke 4: 规则包加载校验（四包零误报）==="
+if python3 "$REPO_ROOT/ops-verdict/scripts/verdict.py" \
+    --rules "$REPO_ROOT/ops-verdict/rules/L3-inter-card.yaml,$REPO_ROOT/ops-verdict/rules/L2-inter-core.yaml,$REPO_ROOT/ops-verdict/rules/L1-single-core.yaml,$REPO_ROOT/ops-verdict/rules/L0-instruction.yaml" \
+    --soc a5 --scalar t_comm_us=100,t_compute_us=250 \
+    -o "$SMOKE_TMP/v.json" >/dev/null 2>&1; then
+  pass "四规则包全瀑布加载校验通过（V1-V4 零误报）"
+else
+  fail "四规则包全瀑布加载校验"
+fi
+
+echo "=== smoke 5: 防回归自证——顶层 local 注入必须被冒烟抓住 ==="
+INJECT_TMP="$SMOKE_TMP/inject.sh"
+cp "$REPO_ROOT/ops-collect/scripts/collect_board.sh" "$INJECT_TMP"
+# 在 mc2 case 前注入一行顶层 local（复刻「local 函数外使用」bug 形态）
+python3 - "$INJECT_TMP" <<'PYEOF'
+import sys
+from pathlib import Path
+p = Path(sys.argv[1])
+t = p.read_text(encoding="utf-8")
+t = t.replace("  mc2)\n    run_msprof_mc2\n", "  mc2)\n    local injected=1\n    run_msprof_mc2\n")
+p.write_text(t, encoding="utf-8")
+PYEOF
+if bash "$INJECT_TMP" --exe /bin/true --mode mc2 --output "$SMOKE_TMP/prof_inject" >/dev/null 2>&1; then
+  fail "顶层 local 注入未被抓住（冒烟网失效）"
+else
+  pass "顶层 local 注入 → mc2 模式退出非 0（bash -n 查不出类 bug 冒烟可抓）"
+fi
+
+echo "=== smoke 6: 族表资产静态断言（防 SKILL.md 数据漂移）==="
+if [ -f "$REPO_ROOT/ops-family/scripts/family_asset_audit.py" ]; then
+  python3 "$REPO_ROOT/ops-family/scripts/family_asset_audit.py" >/dev/null 2>&1
+  check "family_asset_audit.py（族表模板数与实测一致）"
+else
+  echo "[SKIP] family_asset_audit.py 未交付（随算子族 PR），smoke 6 暂跳过——交付后本段自动生效"
+fi
+
+echo "=== smoke 7: facts_bridge 真实导出干跑（--scalar 标量生产者）==="
+BRIDGE_TMP="$SMOKE_TMP/bridge"
+mkdir -p "$BRIDGE_TMP/fixtures"
+# 用 fixture 形态（本仓 6 个 fixtures 同款逐核明细 + 顶层 7 组导出）构造最小产物集
+printf 'block_id,sub_block_id,aic_time(us),aiv_time(us),aiv_vec_ratio,aic_cube_ratio,aic_mte2_ratio,aic_mte1_ratio,aic_fixpipe_ratio,ai_scalar_ratio,aic_mte3_ratio,ai_icache_miss_rate,aiv_mte2_active_bw(GB/s)\n0,Vec0,0,100.5,0.32,0.00,0.55,0.02,0.01,0.22,0.03,1.2,1200.0\n1,Vec0,0,99.8,0.31,0.00,0.56,0.02,0.01,0.21,0.03,1.1,1240.0\n' > "$BRIDGE_TMP/fixtures/PipeUtilization.csv"
+printf 'block_id,sub_block_id,aiv_total_hit_rate(%%)\n0,Vec0,96.1\n1,Vec0,95.7\n' > "$BRIDGE_TMP/fixtures/L2Cache.csv"
+printf 'block_id,sub_block_id,aiv_vec_fp32_ratio\n0,Vec0,0.61\n1,Vec0,0.63\n' > "$BRIDGE_TMP/fixtures/ArithmeticUtilization.csv"
+printf 'block_id,sub_block_id,aiv_vec_total_cflt_ratio\n0,Vec0,0.08\n1,Vec0,0.07\n' > "$BRIDGE_TMP/fixtures/ResourceConflictRatio.csv"
+printf 'Op Name,Op Type,Task Duration(us),Block Dim\nop866214,vector,14.7,8\n' > "$BRIDGE_TMP/fixtures/OpBasicInfo.csv"
+if python3 "$REPO_ROOT/ops-collect/scripts/facts_bridge.py" \
+    --manifest "$BRIDGE_TMP/fixtures" --soc a2 --output "$BRIDGE_TMP/scalars.json" >/dev/null 2>&1 \
+    && python3 -c "
+import json, sys
+d = json.load(open('$BRIDGE_TMP/scalars.json'))
+s = d['scalars']
+assert abs(s['l2_hit_rate'] - 95.9) < 0.01, s
+assert abs(s['icache_miss_rate'] - 1.15) < 0.01, s
+assert abs(s['vec_fp32_ratio'] - 62.0) < 0.01, s          # 0~1 比例 ×100
+assert abs(s['vec_total_cflt_ratio'] - 8.0) < 0.01, s      # 0~1 比例 ×100（mean 0.075 → round 0.08 ×100）
+assert abs(s['mte2_active_bw_pct_of_peak'] - 76.25) < 0.01, s  # 1220/1600 GB/s
+assert abs(s['task_duration_us'] - 14.7) < 0.01, s
+# physical_cores 为 AIV 口径（Vec 行集与 row_count 同维；a2 vector_core_cnt=48）
+assert s['block_dim'] == 8 and s['physical_cores'] == 48, s
+assert d['missing'] == [], d['missing']
+" 2>/dev/null; then
+  pass "facts_bridge 标量提取与峰值换算（7 标量 + 换算正确）"
+else
+  fail "facts_bridge 标量提取"
+fi
+
+# physical_cores 分母按行集核型选择：row_count 是活跃子块行数
+# （verdict.time_metrics_from_rows 不滤 sub_block_id，Vec/Cube 行都计入），分母须与
+# 实际行集同维——纯 Cube 行集按 ai_core_cnt，其余按 vector_core_cnt（缺列/不可判
+# 回退 AIV 默认）。
+BRIDGE_TMP2="$SMOKE_TMP/bridge_cube"
+mkdir -p "$BRIDGE_TMP2/fixtures"
+# Cube 行形态（serial_no_overlap_a2.csv 同款）：aic_time 有值、aiv_time=0
+printf 'block_id,sub_block_id,aic_time(us),aiv_time(us),aiv_vec_ratio,aic_cube_ratio,aic_mte2_ratio,aic_mte1_ratio,aic_fixpipe_ratio,ai_scalar_ratio,aic_mte3_ratio,ai_icache_miss_rate,aiv_mte2_active_bw(GB/s)\n0,Cube0,612.4,0,0.01,0.44,0.42,0.18,0.02,0.06,0.01,0.8,1200.0\n1,Cube0,609.8,0,0.01,0.45,0.41,0.18,0.02,0.06,0.01,0.8,1240.0\n' > "$BRIDGE_TMP2/fixtures/PipeUtilization.csv"
+cp "$BRIDGE_TMP/fixtures/L2Cache.csv" "$BRIDGE_TMP2/fixtures/L2Cache.csv"
+cp "$BRIDGE_TMP/fixtures/ArithmeticUtilization.csv" "$BRIDGE_TMP2/fixtures/ArithmeticUtilization.csv"
+cp "$BRIDGE_TMP/fixtures/ResourceConflictRatio.csv" "$BRIDGE_TMP2/fixtures/ResourceConflictRatio.csv"
+cp "$BRIDGE_TMP/fixtures/OpBasicInfo.csv" "$BRIDGE_TMP2/fixtures/OpBasicInfo.csv"
+if python3 "$REPO_ROOT/ops-collect/scripts/facts_bridge.py" \
+    --manifest "$BRIDGE_TMP2/fixtures" --soc a2 --output "$BRIDGE_TMP2/scalars.json" >/dev/null 2>&1 \
+    && python3 -c "
+import json
+d = json.load(open('$BRIDGE_TMP2/scalars.json'))
+s = d['scalars']
+# cube 主导行集 → 分母按行型选 ai_core_cnt（a2=24；非 48 的 AIV 值）
+assert s['physical_cores'] == 24, s
+" 2>/dev/null; then
+  pass "facts_bridge cube 行集分母按行型选择（physical_cores=ai_core_cnt 24）"
+else
+  fail "facts_bridge cube 行集分母（行型错维回归）"
+fi
+
+
+
+echo "=== smoke 8: timing 模式干跑（event 计时 harness 归档）==="
+TIMING_TMP="$SMOKE_TMP/timing"
+mkdir -p "$TIMING_TMP"
+cat > "$TIMING_TMP/fake_harness.sh" <<'HARNESS'
+#!/usr/bin/env bash
+echo '{"schema":"opstune.timing/1.0","kernel_avg_us":14.2,"warmup":5,"repeat":20}' > result.json
+HARNESS
+chmod +x "$TIMING_TMP/fake_harness.sh"
+( cd "$TIMING_TMP" && PATH="$SHIM_BIN:$PATH" bash "$REPO_ROOT/ops-collect/scripts/collect_board.sh" \
+    --exe "$TIMING_TMP/fake_harness.sh" --mode timing --output "$TIMING_TMP/prof" ) >/dev/null 2>&1
+if python3 -c "
+import json
+m = json.load(open('$TIMING_TMP/prof/collection_manifest.json'))
+assert m['mode'] == 'timing', m['mode']
+assert m['blocks'][0]['name'] == 'timing_harness' and m['blocks'][0]['status'] == 'ok', m['blocks']
+r = json.load(open('$TIMING_TMP/prof/result.json'))
+assert r['kernel_avg_us'] == 14.2
+" 2>/dev/null; then
+  pass "collect_board.sh --mode timing（harness result.json 归档 + manifest ok）"
+else
+  fail "collect_board.sh --mode timing"
+fi
+
+# 失败分支：harness 非零退出必须仍落 manifest（failed 块）——
+# 修复前 set -e 在 timeout 行直接杀脚本，failed 分支不可达、manifest 不落盘
+# （失败无审计记录，且该目录因无 manifest 可被重复采集，防重采闸门失效）。
+# 子 shell 必须 `|| true` 兜底：修复前的 bug 形态下 collect_board.sh 以
+# harness 退出码终止（非 0），set -e 会连带杀掉 smoke 本体。
+TIMING_TMP2="$SMOKE_TMP/timing_fail"
+mkdir -p "$TIMING_TMP2"
+cat > "$TIMING_TMP2/fail_harness.sh" <<'FAILH'
+#!/usr/bin/env bash
+echo "boom" >&2
+exit 3
+FAILH
+chmod +x "$TIMING_TMP2/fail_harness.sh"
+( cd "$TIMING_TMP2" && PATH="$SHIM_BIN:$PATH" bash "$REPO_ROOT/ops-collect/scripts/collect_board.sh" \
+    --exe "$TIMING_TMP2/fail_harness.sh" --mode timing --output "$TIMING_TMP2/prof" ) >/dev/null 2>&1 || true
+if python3 -c "
+import json
+m = json.load(open('$TIMING_TMP2/prof/collection_manifest.json'))
+assert m['blocks'][0]['name'] == 'timing_harness', m['blocks']
+assert m['blocks'][0]['status'] == 'failed', m['blocks']
+assert 'rc=3' in m['blocks'][0]['note'], m['blocks']
+" 2>/dev/null; then
+  pass "collect_board.sh --mode timing 失败分支（manifest 落盘 + failed 块 + rc 留痕）"
+else
+  fail "collect_board.sh --mode timing 失败分支（manifest 丢失=死代码回归）"
+fi
+
+
+echo "=== smoke 9: tiling_candidates 四族干跑（候选生成 + 容量校验）==="
+if [ ! -f "$REPO_ROOT/ops-family/scripts/tiling_candidates.py" ]; then
+  echo "[SKIP] tiling_candidates.py 未交付（随算子族 PR），smoke 9 暂跳过——交付后本段自动生效"
+else
+TILING_TMP="$SMOKE_TMP/tiling"; mkdir -p "$TILING_TMP"
+if python3 "$REPO_ROOT/ops-family/scripts/tiling_candidates.py" \
+    --family elementwise --shape 8448,4096 --dtype fp16 --soc a5 \
+    --output "$TILING_TMP/cand.json" >/dev/null 2>&1 \
+    && python3 "$REPO_ROOT/ops-family/scripts/tiling_candidates.py" \
+    --family reduction --shape 32,4096 --dtype fp16 --soc a5 --axis -1 \
+    --output "$TILING_TMP/red.json" >/dev/null 2>&1 \
+    && python3 -c "
+import json
+c = json.load(open('$TILING_TMP/cand.json'))
+assert c['soc_caps']['core_num'] == 28 and c['soc_caps']['ub_kb'] == 248, c['soc_caps']
+fc = next(x for x in c['candidates'] if x['variant'] == 'full-cores')
+assert fc['block_dim'] == 28 and fc['aligned'], fc
+assert fc['block_former'] % 512 == 0, fc        # 多核 512 元素对齐
+assert fc['capacity_kb'] <= 248, fc             # 容量预算不超限
+r = json.load(open('$TILING_TMP/red.json'))
+assert r['pruned'] and r['pruned'][0]['variant'] == 'full-load', r  # 整行驻留不可行被剪枝
+assert r['candidates'][0]['block_dim'] == 28, r
+" 2>/dev/null; then
+  pass "tiling_candidates（elementwise 满核对齐 + reduction full-load 剪枝记录）"
+else
+  fail "tiling_candidates 四族干跑"
+fi
+
+fi
+
+
+echo "=== smoke 10: collect_cases → batch 直连端到端（A1 铁律一条命令）==="
+P1_TMP="$SMOKE_TMP/p1"; mkdir -p "$P1_TMP/bin"
+cat > "$P1_TMP/cases.csv" <<'CASES'
+case_id,shape,dtype,ref_impl
+c01,8448x4096,fp16,aclnnGelu
+c02,4096x4096,fp16,aclnnGelu
+c03,1024x1024,bf16,aclnnGelu
+CASES
+cat > "$P1_TMP/bin/msprof" <<'P1SHIM'
+#!/usr/bin/env bash
+OUT=""
+for a in "$@"; do
+  case "$a" in --output=*) OUT="${a#--output=}" ;; esac
+done
+mkdir -p "$OUT/mindstudio_profiler_output"
+printf 'Task Duration(us),aiv_vec_ratio,aic_mte2_ratio\n100.5,0.32,0.55\n101.2,0.31,0.56\n' \
+  > "$OUT/mindstudio_profiler_output/op_summary_1.csv"
+P1SHIM
+chmod +x "$P1_TMP/bin/msprof"
+# scope.json（12 HARD 字段齐）
+cat > "$P1_TMP/scope.json" <<'SCOPE'
+{"schema":"opstune.scope/1.0","op":"gelu","soc":"a5","mode":"board","timing_method":"msprof_task_duration",
+ "baseline_kind":"system","cann_version":"8.3.0","device_id":"0","shape":"8448x4096","dtype":"fp16",
+ "format":"ND","warmup":5,"repeat":10}
+SCOPE
+( cd "$P1_TMP" && PATH="$P1_TMP/bin:$PATH" python3 \
+    "$REPO_ROOT/ops-collect/scripts/collect_cases.py" \
+    --cases "$P1_TMP/cases.csv" --exe /bin/true --mode quick \
+    --output "$P1_TMP/collect_out" ) >/dev/null 2>&1
+# 模拟标杆侧数据（system 行：80/82/40us）
+printf 'c01,system,80.0,pass\nc02,system,82.0,pass\nc03,system,40.0,pass\n' \
+  >> "$P1_TMP/collect_out/perf_data.csv"
+if python3 "$REPO_ROOT/ops-verdict/scripts/batch_pipeline.py" \
+    --from-collect "$P1_TMP/collect_out" --cases "$P1_TMP/cases.csv" \
+    --baseline-variant system --after-variants custom \
+    --scope "$P1_TMP/scope.json" --output-dir "$P1_TMP/matrix" >/dev/null 2>&1 \
+    && python3 -c "
+import csv, json
+rows = list(csv.DictReader(open('$P1_TMP/collect_out/perf_data.csv')))
+assert len(rows) == 6, rows                       # 3 case × 2 variant（custom 采集 + system 追加）
+assert {(r['case_id'], r['variant']) for r in rows if r['variant']=='custom'} == \
+       {('c01','custom'),('c02','custom'),('c03','custom')}
+md = open('$P1_TMP/matrix/comparison_matrix.md').read()
+assert 'geomean speedup' in md and 'c03' in md    # 矩阵产出且逐 case 覆盖（A1）
+assert '0.793x' in md or 'REVERT' in md           # vs 标杆读数成立（100.85/80 → 劣化如实呈现）
+s = json.load(open('$P1_TMP/collect_out/collect_summary.json'))
+assert s['ok'] == 3 and s['failed_or_partial'] == [], s
+" 2>/dev/null; then
+  pass "collect_cases → batch --from-collect（cases.csv 到对比矩阵零手工中间产物）"
+else
+  fail "P0-1 直连端到端"
+fi
+
+echo "=== smoke 11: csv_to_db + facts_bridge DB 通道端到端（Roadmap #4）==="
+DB_TMP="$SMOKE_TMP/db_channel"; mkdir -p "$DB_TMP/fixtures"
+# fixture：npu-compute 解包形态（Section CSV 顶层平铺 + summary.jsonl）
+printf 'block_id,sub_block_id,aic_time(us),aiv_time(us),aiv_vec_ratio,aiv_mte2_active_bw(GB/s)\n0,vector0,NA,100.5,0.32,1200.0\n1,vector0,NA,99.8,0.31,1240.0\n' > "$DB_TMP/fixtures/PipeUtilization.csv"
+printf 'block_id,sub_block_id,aiv_read_hit_rate(%%)\n0,vector0,96.1\n1,vector0,95.7\n' > "$DB_TMP/fixtures/L2Cache.csv"
+printf 'block_id,sub_block_id,aiv_vec_stu_cflt_ratio,aiv_vec_ldu_cflt_ratio,aiv_vec_sfu_cflt_ratio\n0,vector0,0.01,0.03,0.00\n1,vector0,0.01,0.05,0.00\n' > "$DB_TMP/fixtures/ResourceConflictRatio.csv"
+printf 'block_id,sub_block_id,aiv_vec_vf_ratio\n0,vector0,0.61\n1,vector0,0.63\n' > "$DB_TMP/fixtures/ArithmeticUtilization.csv"
+printf '{"category":"OpInfoSummary","Op Name":"vec_demo","Op Type":"vector","Task Duration(us)":51.7,"Block Dim":8}\n' > "$DB_TMP/fixtures/summary.jsonl"
+if python3 "$REPO_ROOT/ops-collect/scripts/csv_to_db.py" \
+    --csv-dir "$DB_TMP/fixtures" --db "$DB_TMP/fixtures/summary.db" --quiet >/dev/null 2>&1 \
+  && ! python3 "$REPO_ROOT/ops-collect/scripts/csv_to_db.py" \
+    --csv-dir "$DB_TMP/fixtures" --db "$DB_TMP/fixtures/summary.db" --quiet >/dev/null 2>&1 \
+  && python3 -c "
+import sqlite3
+# 注入 PcSampling 行（DB 独有表——真源 DDL 同构，冒烟只验证消费链不验证采集）
+db = sqlite3.connect('$DB_TMP/fixtures/summary.db')
+db.execute('''CREATE TABLE PcSampling(address INTEGER NOT NULL PRIMARY KEY,
+    ibuf_empty INTEGER NOT NULL, nop_cycles INTEGER NOT NULL,
+    scoreboard_not_ready INTEGER NOT NULL, register_bank_conflict INTEGER NOT NULL,
+    resource_conflict INTEGER NOT NULL, warp_level_sync INTEGER NOT NULL,
+    divergence_stack_spill INTEGER NOT NULL, others INTEGER NOT NULL)''')
+db.execute('INSERT INTO PcSampling VALUES(616,0,210,16320,0,0,0,0,0)')
+db.commit(); db.close()
+" 2>/dev/null \
+  && python3 "$REPO_ROOT/ops-collect/scripts/facts_bridge.py" \
+    --manifest "$DB_TMP/fixtures" --soc a5 --no-db \
+    --output "$DB_TMP/scalars_csv.json" >/dev/null 2>&1 \
+  && python3 "$REPO_ROOT/ops-collect/scripts/facts_bridge.py" \
+    --manifest "$DB_TMP/fixtures" --soc a5 \
+    --output "$DB_TMP/scalars_db.json" >/dev/null 2>&1 \
+  && python3 -c "
+import json
+d_csv = json.load(open('$DB_TMP/scalars_csv.json'))
+d_db = json.load(open('$DB_TMP/scalars_db.json'))
+assert d_csv['source_tool'] == 'npu-compute' and d_db['source_tool'] == 'npu-compute'
+s_db = d_db['scalars']
+# DB 通道产出（PcSampling/OpInfoSummary/列候选扩充三类新能力全断言）：
+assert abs(s_db['top_stall_pct'] - 98.73) < 0.01, s_db     # PcSampling → L0 生产者
+assert abs(s_db['task_duration_us'] - 51.7) < 0.01, s_db  # Summary.jsonl OpInfoSummary
+assert abs(s_db['block_dim'] - 8.0) < 0.01, s_db
+assert abs(s_db['l2_hit_rate'] - 95.9) < 0.01, s_db        # read_hit_rate 口径补位
+assert abs(s_db['vec_total_cflt_ratio'] - 5.0) < 0.01, s_db  # stu+ldu+sfu 分量和 ×100
+assert d_db['pcsampling'] and d_db['pcsampling']['top_stall'] == 'scoreboard_not_ready'
+# 双通道等价（同数据两格式，标量必须一致）：
+common = set(d_csv['scalars']) - {'top_stall_pct'}   # top_stall 为 DB 独有
+assert common == set(d_db['scalars']) - {'top_stall_pct'}, (d_csv['scalars'], d_db['scalars'])
+for k in common:
+    assert abs(d_csv['scalars'][k] - d_db['scalars'][k]) < 0.02, (k, d_csv['scalars'][k], d_db['scalars'][k])
+" 2>/dev/null; then
+  pass "csv_to_db 转换（覆盖保护：二次转换拒覆盖）+ facts_bridge DB 通道（top_stall/OpInfoSummary/双通道等价）"
+else
+  fail "facts_bridge DB 通道端到端"
+fi
+
+echo "=== smoke 12: db_bench 等价性门禁（W1-W3 双通道数值一致）==="
+if python3 "$REPO_ROOT/ops-collect/scripts/db_bench.py" \
+    --dir "$DB_TMP/fixtures" --repeat 3 \
+    -o "$DB_TMP/db_bench_report.json" >/dev/null 2>&1 \
+  && python3 -c "
+import json
+r = json.load(open('$DB_TMP/db_bench_report.json'))
+assert r['schema'] == 'opstune.db-bench/1.0'
+assert r['all_equivalent'] is True, r
+for w in ('W1-facts-scalar-extract', 'W2-verdict-full-load', 'W3-analytical-batch'):
+    assert r['workloads'][w]['equivalent'] is True
+    assert r['workloads'][w]['csv_ms'] > 0 and r['workloads'][w]['db_ms'] > 0
+# 不断言具体倍数（机器相关）；W4 PcSampling 独有断言：
+assert r['workloads']['W4-pcsampling-dist']['db_value'] is not None
+" 2>/dev/null; then
+  pass "db_bench 等价性门禁（提速比如实呈现不作门禁）"
+else
+  fail "db_bench 等价性"
+fi
+
+echo "=== smoke 13: collect_board --mode npu-compute shim 干跑（DB 格式采集编排）==="
+NPC_TMP="$SMOKE_TMP/npc"; mkdir -p "$NPC_TMP/bin" "$NPC_TMP/fixtures/unpacked"
+# fixture：解包形态目录（采集 shim 的 --import 复制源）
+printf 'block_id,sub_block_id,aic_time(us),aiv_time(us),aiv_vec_ratio\n0,vector0,NA,100.5,0.32\n' > "$NPC_TMP/fixtures/unpacked/PipeUtilization.csv"
+cp "$NPC_TMP/fixtures/unpacked/PipeUtilization.csv" "$NPC_TMP/fixtures/unpacked/L2Cache.csv"
+# shim npu-compute：--list-sets/--list-sections 查询门禁 + --export 落 rep 副本 +
+# --import 复制 fixture 目录并打印 unpacked= 路径（真实工具的终端契约）
+cat > "$NPC_TMP/bin/npu-compute" <<'NPCSHIM'
+#!/usr/bin/env bash
+FIXTURE="$(dirname "$0")/../fixtures"
+IMPORT=""; EXPORT=""
+for ((i=1; i<=$#; i++)); do
+  arg="${!i}"
+  case "$arg" in
+    --list-sets) echo "basic: Pipeline PipeUtilization Memory MemoryL0 MemoryUB L2Cache ArithmeticUtilization"; echo "full: + ResourceConflictRatio"; exit 0 ;;
+    --list-sections) printf 'PipeUtilization\nMemory\nMemoryL0\nMemoryUB\nL2Cache\nPipeline\nArithmeticUtilization\nResourceConflictRatio\n'; exit 0 ;;
+    --import) n=$((i+1)); IMPORT="${!n}" ;;
+    --export) n=$((i+1)); EXPORT="${!n}" ;;
+  esac
+done
+if [[ -n "$IMPORT" ]]; then
+  DIR="$EXPORT/unpacked_1"; mkdir -p "$DIR"
+  cp -r "$FIXTURE/unpacked/." "$DIR/"
+  echo "npu-compute: unpacked= $DIR"; exit 0
+fi
+printf 'npu-rep-container-fixture' > "$EXPORT"
+echo "npu-compute: report= $EXPORT"; exit 0
+NPCSHIM
+chmod +x "$NPC_TMP/bin/npu-compute"
+if PATH="$NPC_TMP/bin:$PATH" bash "$REPO_ROOT/ops-collect/scripts/collect_board.sh" \
+    --exe /bin/true --mode npu-compute --sections PmSampling,BogusSection \
+    --output "$NPC_TMP/npc_out" >/dev/null 2>&1 \
+  && python3 -c "
+import json
+m = json.load(open('$NPC_TMP/npc_out/collection_manifest.json'))
+assert m['mode'] == 'npu-compute'
+b = m['blocks'][0]
+assert b['name'] == 'npu-compute' and b['status'] == 'ok', m
+assert 'unpacked=' in b['note'], b
+# 产物断言（伪 Green 防护路径）：解包目录主输入在位
+# （note 含中文括号注记——用 (\S+) 在 ASCII 路径后截断）
+import pathlib, re
+unpacked = re.search(r'unpacked=(\S+)', b['note']).group(1)
+unpacked = unpacked.rstrip('（无含')
+assert (pathlib.Path(unpacked) / 'PipeUtilization.csv').exists(), unpacked
+" 2>/dev/null; then
+  pass "collect_board --mode npu-compute（探测门禁 + manifest + 解包产物断言）"
+else
+  fail "npu-compute 采集模式干跑"
+fi
+
+echo "=== smoke 14: trace_bubble 合成 trace 干跑（检视 C-5：唯一无回归覆盖脚本）==="
+# 合成 Chrome Trace（手算期望，内联生成）：VECTOR [0,40) / MTE2 [20,80)，
+# span=80 → bubble=50%；overlap=20/union=80 → 25%；尾 gap (40,80) MTE2 全覆 → DATA 100%。
+# 错误形态：坏 JSON → per-core error 留痕不崩、scalar 为空。
+TB_TMP="$SMOKE_TMP/tb"; mkdir -p "$TB_TMP/bad"
+cat > "$TB_TMP/trace_core_smoke.json" <<'TRACE'
+{
+  "traceEvents": [
+    {"ph": "M", "name": "process_name", "pid": 1, "args": {"name": "RVECLD"}},
+    {"ph": "M", "name": "process_name", "pid": 2, "args": {"name": "MTE2"}},
+    {"ph": "X", "name": "RVECLD", "pid": 1, "ts": 0, "dur": 40},
+    {"ph": "X", "name": "MTE2", "pid": 2, "ts": 20, "dur": 60}
+  ]
+}
+TRACE
+echo "not-json{" > "$TB_TMP/bad/trace_core_0.json"
+if python3 "$REPO_ROOT/ops-verdict/scripts/trace_bubble.py" \
+    --trace "$TB_TMP/trace_core_smoke.json" \
+    --scalar-out "$TB_TMP/scalars.json" >/dev/null 2>&1 \
+    && python3 -c "
+import json
+d = json.load(open('$TB_TMP/scalars.json'))
+assert d['scalars'] == {'overlap_pct': 25.0, 'top_stall_pct': 100.0, 'bubble_pct': 50.0}, d['scalars']
+assert d['per_core']['trace_core_smoke.json']['main_pipe'] == 'VECTOR'
+assert d['per_core']['trace_core_smoke.json']['top_stall_category'] == 'DATA'
+assert 'verdict.py' in d['next']
+" 2>/dev/null; then
+  pass "trace_bubble 合成 trace（overlap/bubble/top_stall 手算值断言）"
+else
+  fail "trace_bubble 合成 trace 断言"
+fi
+if python3 "$REPO_ROOT/ops-verdict/scripts/trace_bubble.py" \
+    --trace-dir "$TB_TMP/bad" >/dev/null 2>&1 \
+    && ! python3 "$REPO_ROOT/ops-verdict/scripts/trace_bubble.py" \
+    --trace-dir "$TB_TMP/bad" 2>/dev/null | grep -q "scalar: overlap"; then
+  pass "trace_bubble 坏 JSON 降级（error 留痕不崩，scalar 为空）"
+else
+  fail "trace_bubble 坏 JSON 降级"
+fi
+
+echo ""
+if [[ $FAILURES -eq 0 ]]; then
+  echo "smoke: 全部通过"
+  exit 0
+else
+  echo "smoke: $FAILURES 项失败 —— 运行时 bug 防线失效，禁止合入"
+  exit 1
+fi

package/skills/asc-tools/tool-ops-tune/ops-verdict/fixtures/core_tail_a5.csv

@@ -0,0 +1,25 @@
+block_id,sub_block_id,aic_time(us),aiv_time(us),aiv_vec_ratio,aic_cube_ratio,aic_mte2_ratio,aic_mte1_ratio,aic_fixpipe_ratio,ai_scalar_ratio,aic_mte3_ratio,ai_icache_miss_rate
+0,Vec0,0,96.3,0.32,0.00,0.28,0.02,0.01,0.22,0.03,0.9
+1,Vec0,0,184.7,0.61,0.00,0.55,0.04,0.02,0.15,0.05,1.1
+2,Vec0,0,91.8,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9
+3,Vec0,0,182.4,0.60,0.00,0.54,0.04,0.02,0.14,0.05,1.0
+4,Vec0,0,95.1,0.31,0.00,0.27,0.02,0.01,0.22,0.03,0.9
+5,Vec0,0,186.2,0.62,0.00,0.56,0.04,0.02,0.15,0.05,1.1
+6,Vec0,0,93.6,0.31,0.00,0.27,0.02,0.01,0.22,0.03,0.9
+7,Vec0,0,185.5,0.61,0.00,0.55,0.04,0.02,0.15,0.05,1.0
+8,Vec0,0,92.7,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9
+9,Vec0,0,184.9,0.61,0.00,0.55,0.04,0.02,0.15,0.05,1.1
+10,Vec0,0,94.2,0.31,0.00,0.27,0.02,0.01,0.22,0.03,0.9
+11,Vec0,0,186.8,0.62,0.00,0.56,0.04,0.02,0.15,0.05,1.1
+12,Vec0,0,93.1,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9
+13,Vec0,0,185.8,0.61,0.00,0.55,0.04,0.02,0.15,0.05,1.0
+14,Vec0,0,94.9,0.31,0.00,0.27,0.02,0.01,0.22,0.03,0.9
+15,Vec0,0,187.3,0.63,0.00,0.56,0.04,0.02,0.15,0.05,1.1
+16,Vec0,0,92.4,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9
+17,Vec0,0,183.6,0.60,0.00,0.54,0.04,0.02,0.14,0.05,1.0
+18,Vec0,0,94.6,0.31,0.00,0.27,0.02,0.01,0.22,0.03,0.9
+19,Vec0,0,185.2,0.61,0.00,0.55,0.04,0.02,0.15,0.05,1.1
+20,Vec0,0,93.4,0.30,0.00,0.27,0.02,0.01,0.21,0.03,0.9
+21,Vec0,0,186.5,0.62,0.00,0.56,0.04,0.02,0.15,0.05,1.1
+22,Vec0,0,92.9,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9
+23,Vec0,0,184.1,0.61,0.00,0.55,0.04,0.02,0.15,0.05,1.0

package/skills/asc-tools/tool-ops-tune/ops-verdict/fixtures/few_cores_a5.csv

@@ -0,0 +1,5 @@
+block_id,sub_block_id,aic_time(us),aiv_time(us),aiv_vec_ratio,aic_cube_ratio,aic_mte2_ratio,aic_mte1_ratio,aic_fixpipe_ratio,ai_scalar_ratio,aic_mte3_ratio,ai_icache_miss_rate
+0,Vec0,0,100.0,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9
+1,Vec0,0,100.0,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9
+2,Vec0,0,100.0,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9
+3,Vec0,0,100.0,0.30,0.00,0.26,0.02,0.01,0.21,0.03,0.9

package/skills/asc-tools/tool-ops-tune/ops-verdict/fixtures/mte2_bound_a2.csv

@@ -0,0 +1,17 @@
+block_id,sub_block_id,aic_time(us),aiv_time(us),aiv_vec_ratio,aic_cube_ratio,aic_mte2_ratio,aic_mte1_ratio,aic_fixpipe_ratio,ai_scalar_ratio,aic_mte3_ratio,ai_icache_miss_rate
+0,Vec0,0,482.1,0.52,0.00,0.78,0.05,0.03,0.12,0.02,1.2
+1,Vec0,0,479.8,0.53,0.00,0.77,0.05,0.03,0.11,0.02,1.1
+2,Vec0,0,481.0,0.51,0.00,0.79,0.05,0.03,0.12,0.02,1.3
+3,Vec0,0,480.5,0.52,0.00,0.78,0.05,0.03,0.11,0.02,1.2
+4,Vec0,0,483.2,0.53,0.00,0.77,0.05,0.03,0.12,0.02,1.2
+5,Vec0,0,478.9,0.51,0.00,0.78,0.05,0.03,0.11,0.02,1.1
+6,Vec0,0,481.7,0.52,0.00,0.78,0.05,0.03,0.12,0.02,1.2
+7,Vec0,0,480.2,0.52,0.00,0.77,0.05,0.03,0.11,0.02,1.2
+8,Vec0,0,479.4,0.51,0.00,0.78,0.05,0.03,0.12,0.02,1.1
+9,Vec0,0,482.8,0.52,0.00,0.78,0.05,0.03,0.11,0.02,1.2
+10,Vec0,0,480.9,0.53,0.00,0.77,0.05,0.03,0.12,0.02,1.2
+11,Vec0,0,481.3,0.51,0.00,0.79,0.05,0.03,0.11,0.02,1.1
+12,Vec0,0,478.6,0.52,0.00,0.78,0.05,0.03,0.12,0.02,1.2
+13,Vec0,0,480.7,0.52,0.00,0.77,0.05,0.03,0.11,0.02,1.2
+14,Vec0,0,481.5,0.51,0.00,0.78,0.05,0.03,0.12,0.02,1.1
+15,Vec0,0,479.9,0.52,0.00,0.78,0.05,0.03,0.11,0.02,1.2