From 13001f2ae42370261da19b4e23dd9fcb2266bda6 Mon Sep 17 00:00:00 2001 From: Wanderingss <1624155937@qq.com> Date: Tue, 27 Oct 2026 11:31:53 +0800 Subject: [PATCH] =?UTF-8?q?feat(=E9=97=A8=E7=A6=81):=20=E6=A0=88=E5=88=A4?= =?UTF-8?q?=E6=8D=AE=E5=AE=9A=E7=A8=BF=E2=80=94=E2=80=94=E9=9D=99=E6=80=81?= =?UTF-8?q?=20margin=20=E5=94=AF=E4=B8=80=E6=8B=A6=E6=8F=90=E4=BA=A4?= =?UTF-8?q?=E3=80=81=E8=BF=90=E8=A1=8C=E6=97=B6=E6=B0=B4=E4=BD=8D=E5=8F=AA?= =?UTF-8?q?=E5=91=8A=E8=AD=A6=E3=80=81=E9=87=87=E9=9B=86=E4=B8=8D=E5=81=A5?= =?UTF-8?q?=E5=BA=B7=E6=95=B4=E8=BD=AE=E4=BD=9C=E5=BA=9F?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit task-32 裁定落成可执行件（道F 交付）。不是二选一而是**分权**：静态最坏链 margin 仍是唯一拦提交项（由 `stack_report.py` 守，它本就把运行时峰值取进 max）；新增 `stack_gate.py` 管台架侧——趋势只出 WARN， **绝对底线**（free_min < 25% cfg）是运行时唯一有 FAIL 权限的一档，**采集健康门不过 = VOID / exit 2**， 不红不绿、不进台账。把「daemon 死了 status 仍报正常、rx_lines=0」那类事故机械化：门不过与设备故障 exit 1 分开记，按码聚合的人不会把端口没开当成又崩了一次栈。 容差是推导的：tol = 上取整到 16 B 的 max(128 B, 2×实测摆动)，实测 80 B ⇒ 统一 160 B，登记进 `scripts/stack_tolerance.tsv` 带来源列；**改数权限写死**——只有台架重测后才允许回填，禁止为让某轮变绿 而调大容差（被这把尺考核的人不得当场决定尺子的松紧），且这句话**紧贴数值写**。 `--selftest` 14 条能红，开发中抓到作者三个真错（身份戳正则把 __DATE__ 月份写成三位数字、一度把跨轮戳 不同当 VOID 会让判据永不可用、文档 §4.0.2 被 doc-links 抓跨文档串号）。**不接 pre-commit**：它的输入 是台架串口采集，提交时既无设备也无当轮日志，挂上去要么永远 SKIP、要么拿上一轮日志判这一轮的码（与 体积门禁那条假凭证同族）。 --- docs/测试说明/启动日志原子性复验流程.md | 35 ++- scripts/stack_gate.py | 310 ++++++++++++++++++++++++ scripts/stack_tolerance.tsv | 21 ++ 3 files changed, 362 insertions(+), 4 deletions(-) create mode 100644 scripts/stack_gate.py create mode 100644 scripts/stack_tolerance.tsv diff --git a/docs/测试说明/启动日志原子性复验流程.md b/docs/测试说明/启动日志原子性复验流程.md index 2be68bf7..01a50469 100644 --- a/docs/测试说明/启动日志原子性复验流程.md +++ b/docs/测试说明/启动日志原子性复验流程.md @@ -100,12 +100,39 @@ 判据写"崩溃遗言一整行、恰好一次、`dropped=0`"，并按 `boot_log_data/README.md` §2「故障注入日志」 运行时放宽 `@test`（`--forbidden` 去掉 `[ERR]`、`--max-boot-count 2`）——**不改仓库配置**。 -## 5. 栈与堆的比对口径 +## 5. 栈判据（task-32 定稿：三档分权，不再是"待定"） -同一份像连跑 4 次仅复位启动，uiTask 峰值实测 4116~4188 B（**自身摆动 72 B**），heap 计数逐字节稳定。 -因此：**不要**用"栈水位不劣于上一轮"这种单次比较当判据。要么取 ≥3 次的最大值对最大值， -要么把 `scripts/stack_report.py` 的静态最坏链当硬线、运行时峰值只作趋势（定稿见 task-32）。 +**一句话裁定：静态最坏链是唯一拦提交项（`scripts/stack_report.py` 每任务 verdict）；运行时 +水位只出告警（`scripts/stack_gate.py`，超容差 = WARN）；采集健康门不过则整轮作废（VOID， +既不红也不绿、也不得写进台账）。** +| 档 | 判据 | 等级与权限 | 由谁跑 | +|---|---|---|---| +| 硬线 | `margin = cfg − max(静态最坏链, 运行时峰值) ≥ max(1024 B ISR, cfg×30%)` | **FAIL = 拦提交** | `scripts/stack_report.py`（看它退出码。**Lead 裁定 2026-10-27：提交时守这条的一直是 stack_report 本身；`stack_gate.py` 定为台架侧门禁、不接 pre-commit**——提交时既无设备也无当轮日志，挂上去只会「永远 SKIP」或「拿上一轮日志判这一轮的码」，后者是假凭证同族） | +| 趋势 | 同镜像 ≥3 圈取最大值 vs 基线轮 ≥3 圈最大值，`Δ > tol(task)` | **WARN = 记录 + 归因，永不单独拦提交** | `scripts/stack_gate.py` | +| 绝对底线 | 任一任务 `free_min < 25% × cfg` | **FAIL**（运行时唯一有 FAIL 权限的一条，不需要跨轮比较） | `scripts/stack_gate.py` | +| 采集健康门 | 锚点 `reset cause:` 恰 1 次 / 身份戳存在且**同轮唯一** / 每块 ≥20 行 / 每块都有 `[STACK]` 行 / ≥3 圈（基线轮同样要过门，否则本轮只跳趋势比较） | **VOID，exit 2** | `scripts/stack_gate.py` | + +**容差是推导值，不是拍脑袋**：`tol(task) = 向上取整到 16 B 格的 max(128 B, 2 × 同镜像实测最大摆动)`。 +实测来源：同一份像 4 圈仅复位里 uiTask 摆 72 B、adcTask 摆 80 B（task-30），2026-10-27 那份像 4 圈零摆动 +（task-42）→ 最大摆动 80 B ⇒ **全任务统一 160 B**，登记在 `scripts/stack_tolerance.tsv`（含来源列）。 +**谁能改这个数**：只有台架在「改栈配置 / 抬栈 / 新增 `DBG_*` 调用点」之后重测同镜像摆动才允许回填； +**禁止为了让某一轮变绿而调大容差**——被这把尺考核的人不得当场决定尺子的松紧（与 §7 里 +"Forbidden 令牌不得落在英文单词内部"同源：判据的松紧与写法都要由测量侧一次性定死）。 + +**身份戳与归因**：跨轮比较本来就是两份镜像，正常；但 Δ 只有在**附上两次构建之间的源码 diff** +时才可归因到某笔改动。台账行必须写**两次的 `fw build` 时间戳（含 `suite=`/`dstr=`）**， +SHA 只当坐标（2026-10-26 那天 SHA 被 amend 挪了三轮）。同镜像比较里出现"零摆动"不代表判据失效， +只代表这次没噪声。 + +**为什么必须留 VOID 这一档**（设计输入=当天两次自纠）：常驻 daemon 死掉后 `status` 会自动拉起 +一个新 daemon 并"正常"应答，`rx_lines=0` 也照样返回成功 JSON；而 Forbidden 令牌 `Fault` 会命中 +`de`+`fault` 制造假红。所以判据**不许依赖采集健康的自我感觉**：门不过 → exit 2 与 exit 1 分开， +按码聚合的人不会把"端口没开"记成"又崩了一次栈"。 + +**判据自己必须能红**：`& '.venv\Scripts\python.exe' scripts/stack_gate.py --selftest`（14 条，含 +"72 B 实测摆动不算红"、"超容差只 WARN"、"单圈基线拒比"、"死采集器 VOID"、"同轮混两份身份戳 VOID"、 +"绝对底线 FAIL"、"峰值取最坏一圈而非最后一圈"）。**改完判据不跑这条，就等于没改。** ## 6. 锁与脚手架 - bench 锁是**会话级**，必须手工 acquire/release，且**用同一个名字**释放。 diff --git a/scripts/stack_gate.py b/scripts/stack_gate.py new file mode 100644 index 00000000..4989c17f --- /dev/null +++ b/scripts/stack_gate.py @@ -0,0 +1,310 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- +"""stack_gate.py - 栈水位趋势判据（task-32 定稿的可执行形态，纯标准库）。 + +模块名称：台架栈水位判据器（趋势 + 采集健康门） +模块功能：把 N 段启动日志里的 [STACK] 水位汇成"本轮最大值"，与基线轮最大值按每任务 + 容差比较，输出四档结论；**先过采集健康门，再谈比较**。 +创建日期：2026-10-27（裁定见 docs/内存管理/RAM使用分析与优化方案.md §4.0.6 第③条） + +为什么要有它：旧判据「栈水位不劣于上一轮」拿单次运行时峰值跨轮比较，而同一份镜像自己 +就会摆（实测 uiTask 4116~4188 = 72 B、adcTask 920~1000 = 80 B），于是既假红又漏报。 +本脚本把那条判据拆成三件事，等级不同、互不越权： + 1) 硬线（唯一拦提交项）= scripts/stack_report.py 的每任务 verdict： + margin = cfg - max(静态最坏链, 运行时峰值) >= max(isr_margin, cfg*30%) + 本脚本不重算也不替代它，只在末尾提示其退出码（接不接 pre-commit 由 Lead 定）。 + 2) 趋势（告警级，永不单独拦提交）= 同一身份戳、>=3 圈取最大值，与基线轮最大值比， + 超每任务容差 -> WARN。容差按字节，来源见 scripts/stack_tolerance.tsv。 + 3) 作废（VOID，优先于前两者）= 采集健康门不过：本轮既不判过也不判不过，也不得写进台账 + 结论。判据不许依赖"采集看起来很健康"这种自我感觉——2026-10-27 实证：常驻 daemon + 死后 status 会自动拉起一个新 daemon，仍报"正常"而 rx_lines=0。 + +退出码（与 verify_log.py 同族口径：测量问题绝不伪装成设备故障）： + 0 = PASS（含通过但有趋势 WARN） + 1 = FAIL（运行时绝对底线破线：free_min < 25% of cfg —— 设备侧真消息） + 2 = VOID_MEASUREMENT（采集健康门未过：本轮不可用，也不算设备的账） + 3 = 用法错误 +""" + +import argparse +import os +import re +import sys + +RUNTIME_FLOOR_PCT = 25 # 运行时绝对底线（与静态 30% 硬线错开一档：峰值只是下界证据） +MIN_CIRCLES = 3 # 少于 3 圈 = 噪声样本，不做跨轮比较 +DEFAULT_TOL = 160 # 未登记任务的兜底容差（字节），推导见 scripts/stack_tolerance.tsv 头注 + +STACK_RE = re.compile(r"\[STACK\]\s+(\w+)\s+peak_used=(\d+)/(\d+)\s+free_min=(\d+)\s*\((\d+)%\)") +# __DATE__ 的月份是**英文名**（"Oct 26 2026 22:52:09"），不是数字——早先写成 \d{3} 会让 +# 每一份真实日志都判成"没有身份戳"而整轮作废（判据全红比判据漏更响，但也别拿假红当纪律）。 +STAMP_RE = re.compile(r"fw build:\s*([A-Za-z]{3}\s+\d{1,2}\s+\d{4}\s+\d{2}:\d{2}:\d{2})\s+suite=(\S+)\s+dstr=(\S+)") +ANCHOR = "reset cause:" + + +def parse_block(text): + """取一段日志的 (每任务峰值, 身份戳, 锚点数, 有效行数)。 + + 同一任务在一块里出现多次时取最大值——栈峰值的语义就是"最坏那一次"，不做平均。 + """ + payload = [l for l in text.splitlines() if l.strip() and not l.lstrip().startswith("#")] + peaks = {} + for ln in payload: + m = STACK_RE.search(ln) + if m: + name = m.group(1) + used, cfg, free = int(m.group(2)), int(m.group(3)), int(m.group(4)) + prev = peaks.get(name) + if prev is None or used > prev["used"]: + peaks[name] = {"used": used, "cfg": cfg, "free": free} + stamps = STAMP_RE.findall("\n".join(payload)) + uniq = sorted(set(stamps)) + stamp = None + if uniq: + stamp = uniq[0] if len(uniq) == 1 else ("MIXED:" + str(len(uniq))) + return {"peaks": peaks, "anchors": sum(1 for l in payload if ANCHOR in l), + "stamp": stamp, "lines": len(payload)} + + +def measurement_gate(blocks, label): + """(ok, reasons)。任一条不满足就整轮作废：宁可不出结论，也不出错结论。""" + reasons = [] + if len(blocks) < MIN_CIRCLES: + reasons.append("%s: only %d circle(s), need >= %d for a peak-vs-peak comparison" + % (label, len(blocks), MIN_CIRCLES)) + stamps = set() + for i, b in enumerate(blocks): + n = i + 1 + if b["anchors"] != 1: + reasons.append("%s block#%d: boot anchor count=%d (want exactly 1) -> capture did " + "not start at a reset, or the windows overlap" % (label, n, b["anchors"])) + if b["lines"] < 20: + reasons.append("%s block#%d: only %d payload line(s) -> the collector was not " + "carrying this boot (a respawned daemon still answers 'ok')" + % (label, n, b["lines"])) + if not b["peaks"]: + reasons.append("%s block#%d: zero [STACK] lines (a missing line is not zero used)" + % (label, n)) + if not b["stamp"]: + reasons.append("%s block#%d: no 'fw build:' identity stamp -> cannot prove which " + "image these numbers belong to" % (label, n)) + elif str(b["stamp"]).startswith("MIXED"): + reasons.append("%s block#%d: %s identity stamps inside one block -> the capture " + "crossed two images" % (label, n, b["stamp"])) + else: + stamps.add(b["stamp"]) + # 同一轮内部出现多个身份戳 = 采集窗跨了一次重烧，这轮的数字不属于任何一份镜像 -> 作废。 + # 注意：**跨轮**戳不同是正常的（那正是要比的两份镜像），不能按作废处理，见 evaluate()。 + if len(stamps) > 1: + reasons.append("%s: this round mixes %d identity stamps %s -> one circle was reflashed " + "mid-round, these peaks belong to no single image" + % (label, len(stamps), sorted(stamps))) + return (not reasons), reasons + + +def load_tolerances(path): + """容差表：每行 taskbytes[来源]，# 开头为注释。缺文件则全用默认值。""" + tol = {} + if not path or not os.path.isfile(path): + return tol, "no tolerance file -> default %d B for every task" % DEFAULT_TOL + for raw in open(path, "r", encoding="utf-8").read().splitlines(): + line = raw.strip() + if not line or line.startswith("#"): + continue + parts = line.split("\t") + if len(parts) >= 2 and parts[1].isdigit(): + tol[parts[0]] = int(parts[1]) + return tol, "loaded %d task tolerance(s) from %s" % (len(tol), os.path.basename(path)) + + +def worst_of(blocks): + """把本轮各圈压成"每任务最大值"（峰值就该取最坏一次）。""" + agg = {} + for b in blocks: + for name, info in b["peaks"].items(): + prev = agg.get(name) + if prev is None or info["used"] > prev["used"]: + agg[name] = dict(info) + return agg + + +def decide(cur, base, tol_map): + """跨轮趋势：只出 WARN。FAIL 的权限只给绝对底线，硬线归 stack_report.py。""" + rows, warns = [], [] + for name in sorted(cur): + info = cur[name] + bl = base.get(name) + tol = tol_map.get(name, DEFAULT_TOL) + pct = (100 * info["free"] // info["cfg"]) if info["cfg"] else 0 + row = {"task": name, "cur_max": info["used"], "cfg": info["cfg"], "free_min": info["free"], + "floor_pct": pct, "tol": tol, "base_max": (bl or {}).get("used"), + "delta": (info["used"] - bl["used"]) if bl else None, "level": "PASS"} + if pct < RUNTIME_FLOOR_PCT: + row["level"] = "FAIL" + warns.append("%s free_min=%d/%d (%d%%) below the %d%% runtime floor" + % (name, info["free"], info["cfg"], pct, RUNTIME_FLOOR_PCT)) + elif bl is not None and (info["used"] - bl["used"]) > tol: + row["level"] = "WARN" + warns.append("%s peak rose %d B over baseline max (%d -> %d, tol=%d B): trend only, " + "the blocking decision is stack_report.py's margin verdict" + % (name, info["used"] - bl["used"], bl["used"], info["used"], tol)) + rows.append(row) + return rows, warns + + +def evaluate(cur_blocks, base_blocks, tol_map): + """(exit_code, verdict, 报告行)。顺序：先门，再绝对底线，最后才是趋势告警。""" + out = [] + ok_cur, why_cur = measurement_gate(cur_blocks, "current") + if not ok_cur: + out.append("[VOID_MEASUREMENT] this round is not usable:") + out.extend(" - " + r for r in why_cur) + out.append(" -> fix the measurement first (open the port and watch rx_lines grow,") + out.append(" capture from a reset, prove the identity stamp). A dead collector must") + out.append(" not become a stack FAIL, and this round must NOT enter the ledger.") + return 2, "VOID_MEASUREMENT", out + if base_blocks: + ok_base, why_base = measurement_gate(base_blocks, "baseline") + if not ok_base: + out.append("[WARN] baseline round failed its own gate -> trend skipped this run:") + out.extend(" - " + r for r in why_base) + out.append(" (we refuse to compare against numbers we cannot trust; the current") + out.append(" round still gets its absolute-floor check)") + base_blocks = [] + cur = worst_of(cur_blocks) + base = worst_of(base_blocks) if base_blocks else {} + rows, warns = decide(cur, base, tol_map) + cur_stamp = cur_blocks[0]["stamp"] + base_stamp = base_blocks[0]["stamp"] if base_blocks else None + cross_image = bool(base_blocks) and cur_stamp != base_stamp + out.append("%-9s %8s %9s %6s %7s %6s %5s %s" + % ("task", "cur_max", "base_max", "delta", "cfg", "free%", "tol", "level")) + for r in rows: + out.append("%-9s %8d %9s %6s %7d %5d%% %5d %s" % ( + r["task"], r["cur_max"], + "-" if r["base_max"] is None else r["base_max"], + "-" if r["delta"] is None else ("%+d" % r["delta"]), + r["cfg"], r["floor_pct"], r["tol"], r["level"])) + fails = [w for w in warns if "floor" in w] + trend = [w for w in warns if "floor" not in w] + for w in fails: + out.append("[FAIL] " + w) + for w in trend: + out.append("[WARN] " + w) + if not base_blocks: + out.append("note: no usable baseline -> trend comparison off this run (levels come from " + "the absolute runtime floor only)") + elif cross_image: + out.append("note: CROSS-IMAGE comparison (current=%s vs baseline=%s). Different stamps " + "are normal here -- but a delta is only attributable to a given change if the " + "source diff between those two builds is attached; write BOTH stamps into the " + "ledger row, never the commit SHA alone (SHAs moved three times on 2026-10-26)." + % (cur_stamp[0], base_stamp[0])) + else: + out.append("note: same-image comparison (stamps match: %s) -> a delta over tolerance is " + "measurement noise or an unexplained regression; re-measure before writing it up." + % (cur_stamp[0],)) + out.append("hard line unchanged: run scripts/stack_report.py; its per-task verdict " + "(margin >= max(isr_margin, 30% of cfg), computed from the larger of the static " + "chain and the runtime peak) is the only stack criterion that blocks a commit.") + if fails: + return 1, "FAIL", out + if trend: + return 0, "PASS_WITH_WARN", out + return 0, "PASS", out + + +def synth(stamp, anchors, entries, filler=24): + """造一段日志文本（selftest 用）。entries = [(task, used, cfg, free)]。""" + L = ["[RX] 2026-10-27T10:00:%02d.000+08:00 [PROF] filler=%d" % (i % 60, i) for i in range(filler)] + if stamp: + L.append("[RX] 2026-10-27T10:00:00.000+08:00 [2026-10-27 02:00:00][APP] fw build: " + + stamp + " suite=NONE dstr=off") + for _ in range(anchors): + L.append("[RX] 2026-10-27T10:00:00.100+08:00 [2026-10-27 02:00:00][APP] reset cause: " + "software (NVIC_SystemReset) (CSR=0x14000000)") + for task, used, cfg, free in entries: + L.append("[RX] 2026-10-27T10:00:01.000+08:00 [2026-10-27 02:00:01][STACK] %s " + "peak_used=%d/%d free_min=%d (%d%%)" % (task, used, cfg, free, 100 * free // cfg)) + return "\n".join(L) + "\n" + + +def selftest(out=print): + """证明四档都能真的亮——不能失败的判据等于没有判据。""" + cases = [] + + def chk(name, got, want): + cases.append((name, got == want, "got=%r want=%r" % (got, want))) + + tasks = ("default", "net", "ftp", "ui", "timesync", "adc", "iec104", "protect", "test") + clean = [(t, 1000, 4096, 3096) for t in tasks] + S = "Oct 26 2026 22:52:09" + b1 = parse_block(synth(S, 1, clean)) + b2 = parse_block(synth(S, 1, [(t, 1040, 4096, 3056) for t in tasks])) + b3 = parse_block(synth(S, 1, [(t, 990, 4096, 3106) for t in tasks])) + chk("three clean circles -> PASS", evaluate([b1, b2, b3], [], {})[1], "PASS") + base72 = [parse_block(synth("Oct 26 2026 19:43:23", 1, [("ui", 4116, 16384, 12268)]))] + cur72 = [parse_block(synth("Oct 26 2026 19:43:23", 1, [("ui", 4188, 16384, 12196)]))] * 3 + chk("the measured 72 B swing is not a red flag", evaluate(cur72, base72, {"ui": 160})[1], "PASS") + cur_big = [parse_block(synth("Oct 26 2026 19:43:23", 1, [("ui", 4388, 16384, 11996)]))] * 3 + base72r = base72 * 3 # 基线轮也要 >=3 圈：单圈基线是噪声样本，不许拿它比 + chk("delta over tolerance = WARN not FAIL", evaluate(cur_big, base72r, {"ui": 160})[1], "PASS_WITH_WARN") + chk("single-circle baseline is refused (trend skipped, no false red)", + evaluate(cur_big, base72, {"ui": 160})[1], "PASS") + chk("cross-round different stamps still compares (with attribution note)", + 0, evaluate([parse_block(synth("Oct 26 2026 22:52:09", 1, [("ui", 4400, 16384, 11984)]))] * 3, + base72r, {"ui": 160})[0]) + low = [parse_block(synth("Oct 26 2026 19:43:23", 1, [("ui", 13000, 16384, 3384)]))] * 3 + chk("runtime floor breach = FAIL", evaluate(low, [], {})[1], "FAIL") + dead = parse_block(synth(S, 0, [])) + chk("dead collector -> VOID (not FAIL)", evaluate([dead] * 3, [], {})[1], "VOID_MEASUREMENT") + two = parse_block(synth(S, 2, clean)) + chk("two boot anchors -> VOID", evaluate([two, two, two], [], {})[1], "VOID_MEASUREMENT") + nostamp = parse_block(synth("", 1, clean)) + chk("missing identity stamp -> VOID", evaluate([nostamp] * 3, [], {})[1], "VOID_MEASUREMENT") + mixed = parse_block(synth(S, 1, clean) + synth("Oct 25 2026 09:00:00", 1, clean)) + chk("mixed stamps in one block -> VOID", evaluate([mixed] * 3, [], {})[1], "VOID_MEASUREMENT") + cross = parse_block(synth("Oct 25 2026 09:00:00", 1, clean)) + chk("round that mixes two stamps -> VOID", evaluate([b1, b2, cross], [], {})[1], "VOID_MEASUREMENT") + chk("only two circles -> VOID", evaluate([b1, b2], [], {})[1], "VOID_MEASUREMENT") + chk("unusable baseline -> trend skipped, floor still enforced", + evaluate([b1, b2, b3], [dead], {})[1], "PASS") + chk("peak takes the worst circle, not the last", worst_of([b1, b3, b2])["ui"]["used"], 1040) + + bad = 0 + for name, ok, detail in cases: + out("%-5s %s%s" % ("[OK]" if ok else "[FAIL]", name, "" if ok else " " + detail)) + if not ok: + bad += 1 + out("SELFTEST: %d case(s), %d failure(s) -> %s" % (len(cases), bad, "PASS" if bad == 0 else "FAIL")) + return 0 if bad == 0 else 1 + + +def main(argv=None): + ap = argparse.ArgumentParser(description="bench stack water-level trend gate (task-32 ruling)") + ap.add_argument("--block", action="append", default=[], + help="current-round boot log block, repeatable (need >= %d)" % MIN_CIRCLES) + ap.add_argument("--baseline-block", action="append", default=[], + help="previous-round boot log block, repeatable; omit = trend comparison off") + ap.add_argument("--tolerance-tsv", default="scripts/stack_tolerance.tsv", + help="per-task tolerance table (task TAB bytes); missing task -> %d B" % DEFAULT_TOL) + ap.add_argument("--selftest", action="store_true", help="run built-in classification cases and exit") + a = ap.parse_args(argv) + if a.selftest: + return selftest() + if not a.block: + print("usage error: give at least one --block (or --selftest)") + return 3 + tol_map, note = load_tolerances(a.tolerance_tsv) + blocks = [parse_block(open(p, "r", encoding="utf-8", errors="replace").read()) for p in a.block] + base = [parse_block(open(p, "r", encoding="utf-8", errors="replace").read()) for p in a.baseline_block] + print("stack trend gate -- %s (floor=%d%% of cfg, min circles=%d)" % (note, RUNTIME_FLOOR_PCT, MIN_CIRCLES)) + code, verdict, lines = evaluate(blocks, base, tol_map) + for ln in lines: + print(ln) + print("VERDICT: %s" % verdict) + return code + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/scripts/stack_tolerance.tsv b/scripts/stack_tolerance.tsv new file mode 100644 index 00000000..dbb1195f --- /dev/null +++ b/scripts/stack_tolerance.tsv @@ -0,0 +1,21 @@ +# scripts/stack_tolerance.tsv -- 栈水位跨轮容差（字节）。判据裁定见 +# docs/内存管理/RAM使用分析与优化方案.md §4.0.6 第③条；用法：scripts/stack_gate.py。 +# +# 容差是**推导值不是拍的**：tol(task) = 向上取整到 16 B 格的 max(128 B, 2 x 同镜像实测最大摆动)。 +# 实测来源（台架原始日志在 docs/测试说明/boot_log_data/）： +# 2026-10-26 task-30（同一份 1d4b6a17 像、4 圈仅复位）：uiTask 4116~4188 = 72 B、adcTask 80 B +# 2026-10-27 task-42（HEAD 像、4 圈仅复位）：全部任务零摆动 +# -> 最大同镜像摆动 80 B => tol = max(128, 160) = **160 B**，全任务统一。 +# 谁改这个数：**只允许台架**在「改栈配置 / 抬栈 / 新增 DBG 日志调用点」后重测同镜像摆动并回填， +# 并在下面那列写上来源；**禁止为了让某一轮变绿而调大容差**（与 Forbidden 令牌那条同源： +# 判据的松紧不能由被它考核的人当场决定）。 +# 格式：taskbytes来源 +default 160 task-30/42 实测同镜像最大摆动 80 B x2（2026-10-27） +net 160 同上 +ftp 160 同上 +ui 160 同上 +timesync 160 同上 +adc 160 同上 +iec104 160 同上 +protect 160 同上 +test 160 同上（仅测试构建创建）