release: v0.2.4 里程碑快照(孤儿提交,无父;树 = dev 68362e5)
CI / fmt (rustfmt --check) (push) Waiting to run
CI / clippy (ubuntu-latest) (push) Waiting to run
CI / clippy (windows-latest) (push) Waiting to run
CI / test (ubuntu-latest) (push) Waiting to run
CI / test (windows-latest) (push) Waiting to run
CI / build release (ubuntu-latest) (push) Waiting to run
CI / build release (windows-latest) (push) Waiting to run
CI / runtime smoke (ubuntu x64, xvfb) (push) Waiting to run
CI / runtime smoke (macos arm64, probe) (push) Skipped
CI / update-e2e (windows x64) (push) Waiting to run
GPU evidence / GPU 判据自测(不需 GPU;跳过 ≠ 已验证) (push) Waiting to run
GPU evidence / GPU 证据 — 探测可用产物与硬件适配器 (push) Waiting to run
Guards / core 零 UI 守门 (push) Waiting to run
Guards / presenter 零 UI 守门 (push) Waiting to run
Guards / design 零第三方依赖守门 (push) Waiting to run
Guards / unsafe 落点守门(集合必须恰好等于期望) (push) Waiting to run
Guards / 文件行数 ≤400 守门 (push) Waiting to run
Guards / 依赖许可与来源(cargo-deny) (push) Waiting to run
Release guard / 发布标记语义守门(push tag → 稳定版) (push) Waiting to run
Signing status(播报,不是门禁) / 判据自测(三行/退出码/变异重放/接线) (push) Waiting to run
Signing status(播报,不是门禁) / 签名就绪度(播报) (push) Waiting to run
Release / 选槽位(默认四个:Windows 两档 + Linux 两档) (push) Waiting to run

本批内容见 CHANGELOG.md 的 0.2.4 段。按 D41 用孤儿快照推 main,保留「公开史只有里程碑」的模型。
This commit is contained in:
AskaEth committed 2026-11-11 21:41:38 +08:00
commit a5d1dec495
1580 files changed
+320938

No files matched your search

+313
View File
@@ -0,0 +1,313 @@
#!/usr/bin/env python3
"""GPU 截图证据判定(M249):把「这张图真是 GPU 档出来的」变成机器判据,而不是靠人读日志。
为什么需要它:M246 实测到 `--renderer=gpu` 落到 RX 6600 XT / Vulkan(DiscreteGpu)并出图,但那份
结论是**人读 258 行日志**得出的。日志少一行、`Selected` 换了适配器、PNG 是上一次的残留,人眼复核
一遍就得重来。本脚本把那套判据固定成三分钟能复跑一遍的门禁,且**缺哪条就点名哪条**。
入参 = 一份 stderr 日志 + 一张 PNG(可选 `--expect-refuse` 走负例口径)。正例六条判据(每条抓得住
什么、抓不住什么见 `packaging/docs/gpu-evidence.md` §2):
1. **全部** `[shot] … renderer=` 行都是 `wgpu`——不是「出现过」(把 CPU 那次的 stderr `>>` 拼进来
就会骗过旧写法),也**不是** `[probe] renderer = wgpu`:后者只是启动前的意图,M246 的 A6 负例里
它照样打印 wgpu,却一个像素都没画。
2. `Selected: AdapterInfo { … device_type: X }` 的 `X` 不在 `DISALLOWED_DEVICE_TYPES` 内;
**取不到 `X` 也算失败**(多行 Debug 漂移不许静默给绿)。
3. PNG 能解码且像素数 > 0。
4. PNG 宽高 = 日志里 `[shot] W×H` 声明的宽高(把图和日志钉在一起)。
5. `[backend] final=wgpu result=started`(后端自己的结论行与上面几条不矛盾)。
6. PNG 过 `check_screenshot.py` 的非空白判据(复用它的解码器与阈值,不复制第二份)。
`--expect-refuse` 是负例口径(M246 的 A6):要求 `[backend] final=… via=refuse`、**没有**任何
`[shot]` 出图行、且目标 PNG 不存在。它证明「该拒绝时真拒绝了」,不证明「GPU 能出图」。
零第三方依赖;PNG 只解码一次,用的是同目录 `check_screenshot.py` 的 `decode_png` / `measure` / `verdict`。
用法::
python packaging/check_gpu_shot.py shot.stderr.log shot.png # 正例:判定,0/1 退出
python packaging/check_gpu_shot.py a6.stderr.log --expect-refuse # 负例:判定拒绝是否成立
python packaging/check_gpu_shot.py --self-test # 19 条向量(数据+脚手架都在 tests/)
"""
from __future__ import annotations
import re
import sys
import zlib
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent)) # 同目录 import(-m 方式调用时也要成立)
from check_screenshot import ( # noqa: E402
DEFAULT_MIN_EDGE_RATIO,
DEFAULT_MIN_STDDEV,
DEFAULT_MIN_UNIQUE,
PngError,
decode_png,
measure,
verdict,
)
#: 这些 `device_type` 等于「不是硬件 GPU」:`Cpu` = WARP/软渲染适配器,`Other` = 未分类(本机那条
#: `Other` 是 Gl 兼容层)。用 `not in` 而不是 `in {DiscreteGpu, IntegratedGpu}`,与
#: `crates/app-iced/src/probe.rs:165-167` 的硬件判据同口径:`VirtualGpu` 之类新词条不会因为本脚本
#: 没登记就被当成「不是硬件」。反过来,把 `Cpu` 从这张禁令里拿掉正是最想抓的回归(向量第 3 条)。
DISALLOWED_DEVICE_TYPES = frozenset({"Cpu", "Other"})
#: 截图行里的 renderer。只认 `[shot]` 行 —— `[probe] renderer = …` 是启动前的意图,不是结果。
SHOT_RENDERER_RE = re.compile(r"^\[shot\][^\n]*?\brenderer=(\S+)", re.MULTILINE)
#: `[shot] 1180x760 scale=1 -> …`:出图行声明的像素尺寸。
SHOT_SIZE_RE = re.compile(r"^\[shot\]\s+(\d+)x(\d+)\b", re.MULTILINE)
#: `[gpu] …compositor: Selected: AdapterInfo { … }`:真正建出来的那块适配器。这两处约束不是洁癖:
#: 每份日志里还有一行 `[probe] 取证:出现 …Selected: AdapterInfo…` 的**提示文本**(含同样几个字、
#: 但没有 `{`),锚 `^\[gpu\]` + 要求紧跟 `{` 才不会被那行骗过去。
SELECTED_RE = re.compile(r"^\[gpu\][^\n]*Selected:\s*AdapterInfo\s*\{(?P<body>[^\n]*)", re.MULTILINE)
ADAPTER_NAME_RE = re.compile(r'name:\s*"([^"]*)"')
DEVICE_TYPE_RE = re.compile(r"device_type:\s*(\w+)")
ADAPTER_BACKEND_RE = re.compile(r"backend:\s*(\w+)")
#: `[backend] final=… result=… … via=…`:后端自己的结论行。
BACKEND_FINAL_RE = re.compile(r"^\[backend\] final=(\S+)\s+result=(\S+)(?P<rest>[^\n]*)", re.MULTILINE)
VIA_RE = re.compile(r"\bvia=(\w+)")
#: 判据编号 → 标签。编号是报告里的次序,也在 `--self-test` 的向量里被逐条点名。
LABELS = {
1: "截图行 renderer=wgpu",
2: "Selected: AdapterInfo 的 device_type 不是 Cpu/Other",
3: "PNG 能解码且非空",
4: "PNG 宽高 = 日志 [shot] 的宽高",
5: "后端结论行 final=wgpu result=started",
6: "PNG 不是空白帧",
7: "后端结论行 via=refuse",
8: "没有出图行([shot] 尺寸 + renderer 都不该有)",
9: "目标 PNG 不存在(拒绝出图就该一张都没有)",
}
class Report:
"""逐条判据的结论。失败时由 `print_report` 逐条点名,绝不只印一个 FAILED。"""
def __init__(self) -> None:
self.items: list[tuple[int, bool, str]] = []
def add(self, criterion: int, ok: bool, detail: str = "") -> bool:
self.items.append((criterion, ok, detail))
return ok
def problems(self) -> list[str]:
return [LABELS[criterion] for criterion, ok, _detail in self.items if not ok]
def print_report(self, log_path: Path, png_path: Path | None) -> None:
for criterion, ok, detail in self.items:
print(f"{'[ok]' if ok else '[fail]'} {criterion}. {LABELS[criterion]}" + (f":{detail}" if detail else ""))
target = str(log_path) if png_path is None else f"{log_path} + {png_path}"
problems = self.problems()
if problems:
print(f"[fail] 判据没过 {len(problems)}/{len(self.items)} 条({target})", file=sys.stderr)
else:
print(f"[ok] {len(self.items)} 条判据全过({target})")
def read_log(path: Path) -> str:
"""日志按 UTF-8 读、坏字节替换 —— 判据只看 ASCII 标记,不该因为一行中文乱码就崩。"""
return path.read_text(encoding="utf-8", errors="replace")
def shot_renderers(text: str) -> list[str]:
return SHOT_RENDERER_RE.findall(text)
def shot_size(text: str) -> tuple[int, int] | None:
match = SHOT_SIZE_RE.search(text)
return (int(match.group(1)), int(match.group(2))) if match else None
def _pick(pattern: re.Pattern[str], text: str, default: str | None = "?") -> str | None:
match = pattern.search(text)
return match.group(1) if match else default
def selected_adapter(text: str) -> dict[str, str | None] | None:
"""只取 `Selected:` 那一行:`Available adapters: [...]` 里有 12 条 `AdapterInfo`,取错就全错。
`device_type` 取不到时返回 `None`(**不是** `"?"`):一旦 iced 改成 pretty-print Debug,
字段就落不到 `{` 同一行,此时按「取不到 = 判红」处理,不许静默给绿(M249 r1 P2-1)。
"""
match = SELECTED_RE.search(text)
if match is None:
return None
body = match.group("body")
return {
"name": _pick(ADAPTER_NAME_RE, body),
"device_type": _pick(DEVICE_TYPE_RE, body, None),
"backend": _pick(ADAPTER_BACKEND_RE, body),
}
def backend_final(text: str) -> dict[str, str] | None:
match = BACKEND_FINAL_RE.search(text)
if match is None:
return None
via = VIA_RE.search(match.group("rest"))
return {"final": match.group(1), "result": match.group(2), "via": via.group(1) if via else ""}
def decode_shot(png_path: Path, report: Report) -> tuple[int, int, dict | None]:
"""解码 PNG 并核对像素数,返回 (宽, 高, 统计)。读不出就记一条失败并返回 0 尺寸。"""
try:
width, height, rows, channels = decode_png(png_path)
except (PngError, OSError, zlib.error) as error:
report.add(3, False, f"读不出 {png_path}:{error}")
return 0, 0, None
pixels = width * height
report.add(3, pixels > 0, f"{width}x{height} = {pixels} 像素")
return (width, height, measure(width, height, rows, channels)) if pixels > 0 else (width, height, None)
def judge_evidence(log_path: Path, png_path: Path) -> Report:
"""正例口径:这份日志 + 这张 PNG 能不能当「GPU 档真的出了这一帧」的证据。"""
report = Report()
try:
text = read_log(log_path)
except OSError as error:
report.add(1, False, f"日志读不出 {log_path}:{error}")
return report
renderers = list(dict.fromkeys(shot_renderers(text)))
# 「这张图是 wgpu 出的」= 日志里**每一行** `[shot] … renderer=` 都说 wgpu。用 `in` 就退化成
# 「曾经出现过」:把 CPU 那次的 stderr `>>` 追加进来、配 CPU 那张图,会静默判绿(M249 r1 P2-2)。
report.add(
1,
set(renderers) == {"wgpu"},
"实际 renderer=" + "/".join(renderers) if renderers else "日志里没有 `[shot] … renderer=…` 行(没走到出图那一步?)",
)
adapter = selected_adapter(text)
if adapter is None:
report.add(2, False, "日志里没有 `Selected: AdapterInfo { … }`(wgpu 没建成设备,或只打印了 Available adapters)")
else:
device_type = adapter["device_type"]
# `None` = 字段不在 `{` 同一行(多行 Debug 漂移)或字段缺失:按取不到即失败处理,
# 宁可假红也不许「格式一变就静默给绿」。
report.add(
2,
device_type not in (None, "?") and device_type not in DISALLOWED_DEVICE_TYPES,
f'device_type={device_type}({adapter["name"]} / {adapter["backend"]})'
if device_type not in (None, "?")
else "取不到 device_type:`Selected: AdapterInfo {` 与字段不在同一行(多行 Debug),或字段缺失",
)
width, height, stats = decode_shot(png_path, report)
size = shot_size(text)
if size is None:
report.add(4, False, "日志里没有 `[shot] WxH …` 出图行,尺寸无从对账")
elif width == 0:
report.add(4, False, f"PNG 没解出来,无法核对日志声明的 {size[0]}x{size[1]}")
else:
report.add(4, (width, height) == size, f"日志 {size[0]}x{size[1]} vs PNG {width}x{height}")
final = backend_final(text)
if final is None:
report.add(5, False, "日志里没有 `[backend] final=…` 行(ULL_ACCEPTANCE=1 才有)")
else:
report.add(
5,
final["final"] == "wgpu" and final["result"] == "started",
f"final={final['final']} result={final['result']}" + (f" via={final['via']}" if final["via"] else ""),
)
if stats is None:
report.add(6, False, "上一步没解出像素,无从统计")
else:
blank = verdict(stats, DEFAULT_MIN_UNIQUE, DEFAULT_MIN_STDDEV, DEFAULT_MIN_EDGE_RATIO)
report.add(
6,
not blank,
f"唯一色 {stats['unique_colors']} 亮度σ {stats['luma_stddev']} 强边缘 {stats['edge_ratio']}",
)
return report
def judge_refuse(log_path: Path, png_path: Path | None) -> Report:
"""负例口径:这次启动是不是**确实拒绝**降级出图(而不是悄悄换后端出了一张)。"""
report = Report()
try:
text = read_log(log_path)
except OSError as error:
report.add(7, False, f"日志读不出 {log_path}:{error}")
return report
final = backend_final(text)
if final is None:
report.add(7, False, "日志里没有 `[backend] final=…` 行")
else:
report.add(7, final["via"] == "refuse", f"final={final['final']} result={final['result']} via={final['via'] or '<缺>'}")
renderers = list(dict.fromkeys(shot_renderers(text)))
size = shot_size(text)
surplus = (["renderer=" + "/".join(renderers)] if renderers else []) + (
[f"[shot] {size[0]}x{size[1]}"] if size is not None else []
)
report.add(8, not surplus, "存在出图行:" + "、".join(surplus) if surplus else "")
if png_path is None:
return report
if png_path.exists():
report.add(9, False, f"{png_path} 存在({png_path.stat().st_size} 字节)")
else:
report.add(9, True, f"{png_path} 不存在")
return report
def main(argv: list[str]) -> int:
args = list(argv)
positional: list[str] = []
expect_refuse = False
while args:
arg = args.pop(0)
if arg == "--expect-refuse":
expect_refuse = True
elif arg == "--self-test":
sys.path.insert(0, str(Path(__file__).resolve().parent / "tests"))
from gpu_shot_selftest import run_selftest # 延迟 import:脚手架反过来要拿本模块的判据函数
return run_selftest(judge_evidence, judge_refuse)
elif arg in ("-h", "--help"):
print(__doc__)
return 0
elif arg.startswith("-"):
print(f"::error::不认识的参数:{arg}", file=sys.stderr)
return 2
else:
positional.append(arg)
if not positional:
print("::error::至少要给一份 stderr 日志(--help 看用法)", file=sys.stderr)
return 2
if len(positional) > 2:
print(f"::error::最多两个位置参数(日志 + PNG),实际 {len(positional)} 个", file=sys.stderr)
return 2
log_path = Path(positional[0])
png_path = Path(positional[1]) if len(positional) == 2 else None
if png_path is None and not expect_refuse:
print("::error::正例必须给出那张 PNG(只有负例可用 --expect-refuse 省略)", file=sys.stderr)
return 2
report = judge_refuse(log_path, png_path) if expect_refuse else judge_evidence(log_path, png_path)
report.print_report(log_path, png_path)
return 1 if report.problems() else 0
# ────────────────────────── 自测 ──────────────────────────
#
# 向量数据在 `packaging/tests/gpu_shot_vectors.json`,造样本与逐条比对的脚手架在
# `packaging/tests/gpu_shot_selftest.py`(判据脚本自己守 ≤400 行,脚手架不占这里的行数)。
# 每条的 `expect_fail` 是**期望的失败判据清单**,逐条对:关掉一条判据、或让某条判据被别的
# 东西掩盖,清单就对不上。真图不在仓库里(`DevDocs/` 被 ignore),所以这颗钉子守判据,真实
# 产物由报告里的实跑退出码守。
if __name__ == "__main__":
for stream in (sys.stdout, sys.stderr):
if hasattr(stream, "reconfigure"):
stream.reconfigure(errors="backslashreplace")
sys.exit(main(sys.argv[1:]))