remote/tools/cursor_proof.ps1

281 lines
14 KiB
PowerShell
Raw Normal View History

test(tools): cursor_proof + proof_common 抽取 + rel_mouse 适配 v5/双屏与污染修复 4b 光标覆盖层验收(V-D1/D3/D6/D7)+ 4a 建议 2 的公共库沉淀: - proof_common.ps1:抽 landscape/rel_mouse 已验证的观测面 helper (唤醒只从证明过的隐藏态、正向配对判隐藏、N25 唯一临时文件、按文案取按钮、 静置门控 + SetCursorPos 读回自校验、内存预检、R-7 新鲜守卫)。 老 proof 本轮**不迁移**(改 4 个已验证脚本而不全重跑=假绿风险,登记下轮卫生项) - cursor_proof.ps1:真机判据 V-D1 截图青簇质心 vs CursorMarker 映射 ±6px(含近原点 (48,48),Some(0,0)≠None 由单测+golden 锁) V-D3 静止桌面跳转后 900ms 内标记跟随(覆盖 500ms 保活重发带新鲜 cursor) V-D7 指针进左侧副屏负 X → agent 折叠 None → 无青簇(本机双屏白送的真实判据) reject 阶段 V-D6:旧 v4 APK 连 v5 agent → 屏幕出「协议版本不一致」+ detail client=4,server=5 实测全绿:4 点偏差 ≤0.6px、跟随 0.5px、副屏无簇、拒绝判据命中 - rel_mouse_phone_proof.ps1:适配 v5 + 修两处 harness 缺陷(都是本轮回归实测逼出的) ① HUD v4 断言→v5(PROTO_LABEL 配对);② V-B2 R7 期望值从写死 1920 派生 改为按实时桌面 / 动态算(桌面已漂到 2560x1440,旧写死值假报 575px, 违反 AGENTS.md 坐标期望按当前分辨率重算) ③ 模式守卫改为无条件 Set-Mode relative(旧 HUD.Contains(绝对) 在 HUD 读空时 静默跳过,上一轮 throw 残留的 relative_mode=false 污染了整个 step2—— 绝对模式 DOWN 发绝对移动是正确产品行为,被 2c 误读成 teleport) ④ Set-Mode 唤醒 miss 并入 try 预算(原来单次重读即 throw,烧掉整脚本)+ 失败截图证据 实测:R7 误差 1px、2c maxJump=0、长按右键 0→1、全链路绿
2026-11-01 23:39:23 -05:00
# 4b cursor-overlay proof: marker position, off-screen fold, refresh, old-client reject.
#
# ASCII-only (PS 5.1 reads BOM-less .ps1 as GBK; repo lessons). Chinese expected
# texts are built from code points (R-8: never trust an encoding round trip).
#
# Judgements (loop-4b-plan.md section 9):
# V-D6 OLD v4 APK + new v5 agent -> form shows the version-mismatch text and
# the session never goes live. RUN THIS PHASE BEFORE INSTALLING THE NEW
# APK: -Phase reject
# V-D1 screencap cyan cluster centroid == CursorMarker expectation +/- 6 px
# for 4 desktop points incl. (48,48) ~ origin (Some(0,0) vs None: the
# marker MUST be painted at the corner, not skipped)
# V-D3 after moving the PC cursor on a quiet desktop the marker reaches the
# new spot within ~1.2s (covers the 500ms idle-resend path; the frame
# header rebuild with fresh cursor was the whole point of R-A)
# V-D7 cursor into the LEFT SECOND MONITOR (negative X, machine is dual-screen
# since 2026-11-02) -> agent folds to None -> cyan cluster GONE
#
# Prereqs: v5 release agent.exe built; new (v5) APK installed for -Phase main;
# second monitor actually present (V-D7 skips loudly if not - see probe output);
# nobody touching the real mouse during a step (gates retry, loud fail otherwise).
#
# powershell -ExecutionPolicy Bypass -File tools\cursor_proof.ps1 -Host_ 192.168.1.181 -Phase reject
# powershell -ExecutionPolicy Bypass -File tools\cursor_proof.ps1 -Host_ 192.168.1.181 -Phase main
param(
[Parameter(Mandatory)][string]$Host_,
[int]$Port = 47800,
[string]$Code = "t",
[ValidateSet("reject", "main")][string]$Phase = "main",
[string]$AgentExe = ""
)
$ErrorActionPreference = "Stop"
$repo = (Resolve-Path "$PSScriptRoot\..").Path
. "$PSScriptRoot\proof_common.ps1"
# ---- expected Chinese fragments as code points (R-8) ----
# "connect" button label
$F_CONNECT = -join ([char]0x8FDE, [char]0x63A5)
# "reconnect"
$F_RECONNECT = -join ([char]0x91CD, [char]0x65B0, [char]0x8FDE, [char]0x63A5)
# "protocol version mismatch" error summary (ConnectErrorText KIND_VERSION)
$F_VERSION = -join ([char]0x534F, [char]0x8BAE, [char]0x7248, [char]0x672C,
[char]0x4E0D, [char]0x4E00, [char]0x81F4)
# "live" HUD prefix (yi-lian-jie = connected)
$F_LIVE = -join ([char]0x5DF2, [char]0x8FDE, [char]0x63A5)
$ev = Join-Path (Split-Path $repo -Parent) "evidence\4b"
New-Item -ItemType Directory -Force -Path $ev | Out-Null
# ---- cyan marker pixel detector (LockBits scan, full centroid) ----
Add-Type -AssemblyName System.Windows.Forms # VirtualScreen for V-D7
if (-not ('PccCyan' -as [type])) {
Add-Type -ReferencedAssemblies System.Drawing -TypeDefinition @"
using System; using System.Drawing; using System.Drawing.Imaging;
using System.Runtime.InteropServices; using System.Collections.Generic;
public class PccCyan {
public static string Find(string path) {
using (var bmp = new Bitmap(path)) {
var rect = new Rectangle(0,0,bmp.Width,bmp.Height);
var data = bmp.LockBits(rect, ImageLockMode.ReadOnly, PixelFormat.Format32bppArgb);
int n = data.Stride * data.Height; byte[] buf = new byte[n];
Marshal.Copy(data.Scan0, buf, 0, n);
bmp.UnlockBits(data);
// dot color = RGB(0x00,0xE5,0xFF); BGRA in memory. Loose tolerance:
// the ring is white/black; a desktop with a big cyan region would poison the
// centroid, so the total-count guard below rejects it instead of averaging.
int total = 0;
for (int y=0; y<data.Height; y++) {
int row = y*data.Stride;
for (int x=0; x<data.Width; x++) {
int i = row + x*4;
byte b=buf[i], g=buf[i+1], r=buf[i+2];
if (r < 40 && (g>=0xC0 && g<=0xFF) && b>=0xD0) total++;
}
}
if (total == 0) return "none";
if (total > 4000) return "patch:" + total; // suspicious giant color field
// Centroid over ALL cyan pixels would be corrupted by a few stray ones far
// away (measured 2026-11-02: n=181 at the right spot but spread=955 - the
// desktop itself had a handful of cyan pixels elsewhere). Cluster properly:
// bin the hits on an 8px grid, take the DENSEST bin as the marker seed,
// then centroid/spread over hits within 40px of that seed only.
var bins = new Dictionary<long,int>();
for (int y=0; y<data.Height; y++) {
int row = y*data.Stride;
for (int x=0; x<data.Width; x++) {
int i = row + x*4;
byte b=buf[i], g=buf[i+1], r=buf[i+2];
if (r < 40 && (g>=0xC0 && g<=0xFF) && b>=0xD0) {
long key = (long)(x/8)*1000000 + (y/8);
bins[key] = bins.ContainsKey(key) ? bins[key]+1 : 1;
}
}
}
long bestK = -1; int bestN = 0; int secondN = 0;
foreach (var kv in bins) {
if (kv.Value > bestN) { secondN = bestN; bestN = kv.Value; bestK = kv.Key; }
else if (kv.Value > secondN) { secondN = kv.Value; }
}
int bx = (int)(bestK/1000000)*8 + 4, by = (int)(bestK%1000000)*8 + 4;
long sx=0, sy=0; int cnt=0; int minx=99999999,miny=99999999,maxx=-1,maxy=-1;
for (int y=0; y<data.Height; y++) {
int row = y*data.Stride;
for (int x=0; x<data.Width; x++) {
int i = row + x*4;
byte b=buf[i], g=buf[i+1], r=buf[i+2];
if (r < 40 && (g>=0xC0 && g<=0xFF) && b>=0xD0) {
if (Math.Abs(x-bx)<=40 && Math.Abs(y-by)<=40) {
sx+=x; sy+=y; cnt++;
if(x<minx)minx=x; if(x>maxx)maxx=x; if(y<miny)miny=y; if(y>maxy)maxy=y;
}
}
}
}
if (cnt == 0) return "none";
int spread = Math.Max(maxx-minx, maxy-miny);
// report seed too: harness prints it when a check fails (diagnosis, not judgement)
return string.Format("{0}|{1}|{2}|{3}|{4}|{5}|{6}", sx/(double)cnt, sy/(double)cnt, cnt, spread, total, bx, secondN);
}
}
}
"@
}
function Find-Marker([string]$pngPath) {
$r = [PccCyan]::Find($pngPath)
if ($r -eq "none") { return $null }
if ($r -like "patch:*") { throw "screenshot contains a huge cyan-ish color field ($($r.Substring(6)) px) - clear the desktop (wallpaper/browser) and re-run; refusing to average a fake marker (R-8)" }
$p = $r.Split('|')
@{ x = [double]$p[0]; y = [double]$p[1]; n = [int]$p[2]; spread = [int]$p[3]; total = [int]$p[4]; seed = "$($p[5]),$($p[6])" }
}
function Take-Shot([string]$tag) {
$dev = "/sdcard/pcc_$PID.png"
$local = Join-Path $ev "$tag.png"
$old = $ErrorActionPreference; $ErrorActionPreference = "Continue"
try {
[void](AdbShell @("screencap", "-p", $dev))
$r = Adb -AdbArgs @("pull", $dev, $local)
} finally { $ErrorActionPreference = $old }
if ($r.code -ne 0 -or -not (Test-Path $local)) { throw "screencap/pull failed for $tag" }
return $local
}
Write-Host "=== 0. preflight (memory / fresh v5 agent) ==="
Memory-Precheck
if (-not $AgentExe) { $AgentExe = Join-Path $repo "target\release\agent.exe" }
Ensure-FreshAgent $AgentExe $Port $Code
$installed = ((AdbShell @("dumpsys", "package", "com.verify.remote"))) -join "`n"
$vn = [regex]::Match($installed, 'versionName=(\S+)').Groups[1].Value
Write-Host " installed APK: $vn"
# ======================================================================
if ($Phase -eq "reject") {
Write-Host "=== V-D6: OLD (v4) APK against the NEW v5 agent must be rejected ==="
if ($vn -like "*phase4a*") {
Write-Host " (running against the v4 client - this phase MUST come before installing 4b)"
} else {
Write-Host " WARNING: installed APK is $vn - the v4-rejection test is only meaningful against a real v4 client"
}
# NOTE (first-run lesson 2026-11-02): launching via `am start --es host ...`
# does NOT dial: the intent only prefill-prefers stale SharedPreferences and
# the form just sits there ("未连接"). Rejection must be observed through the
# REAL dial path - type the form, tap connect, read the failure screen.
$x = Restart-AppForm
FillAndConnect $x $Host_ "$Port" $Code $F_CONNECT $F_RECONNECT
$deadline = (Get-Date).AddSeconds(25); $shown = $null
while ((Get-Date) -lt $deadline) {
Start-Sleep -Milliseconds 800
$ui = Get-UiRetry
$texts = (VisibleTexts $ui) -join " / "
if ($texts -match "version mismatch" -or $texts.Contains($F_VERSION)) { $shown = $texts; break }
if ($texts.Contains($F_LIVE)) { throw "V-D6 FAIL: v4 client CONNECTED to the v5 agent ('$F_LIVE' on screen) - version gate broken" }
}
if (-not $shown) { throw "V-D6 FAIL: no version-mismatch text within 25s; last screen: [$((VisibleTexts (Get-Ui)) -join ' / ')]" }
Write-Host " screen: $shown"
Write-Host " PASS: v5 agent refused the v4 client with the actionable Chinese hint"
& powershell -NoProfile -ExecutionPolicy Bypass -File "$repo\tools\shot.ps1" -Out "$ev\v-d6-reject.png" | Out-Null
# pollution discipline (4a): leave no live session for the next proof
[void](AdbShell @("am", "force-stop", "com.verify.remote"))
Write-Host "=== CURSOR PROOF (reject phase) COMPLETE ==="
exit 0
}
# ======================================================================
# ---- main phase: needs the NEW apk; go live once, then geometry checks ----
Write-Host "=== 1. go live (new v5 APK), capture HUD (v5 pin + desktop size) ==="
if ($vn -notlike "*phase4b*") { throw "main phase needs the 4b APK installed (found $vn)" }
$x = Restart-AppForm
FillAndConnect $x $Host_ "$Port" $Code $F_CONNECT $F_RECONNECT
$hud = Wait-LiveHud 20
if (-not $hud) { throw "never reached live: [$((VisibleTexts (Get-Ui)) -join ' / ')]" }
Write-Host " HUD: $($hud -replace "`n", ' | ')"
if (-not $hud.Contains("v5")) { throw "V-D5x FAIL: HUD does not report v5 (PROTO_LABEL out of sync?): [$hud]" }
$m = [regex]::Match($hud, '(\d+)x(\d+)')
if (-not $m.Success) { throw "cannot parse desktop size from HUD: [$hud]" }
$dW = [int]$m.Groups[1].Value; $dH = [int]$m.Groups[2].Value
Write-Host " desktop from HUD: ${dW}x${dH}"
# Marker geometry works in EITHER orientation (letterbox mapping is proportional),
# but landscape is the 4a headline case. AUTO rotates for a wide desktop - give the
# rotation a moment to land, then proceed on whatever orientation is actually up.
$rot = (Get-Date).AddSeconds(3); $d = LocalDisplay
while ((Get-Date) -lt $rot) {
$d = LocalDisplay
if ($d.w -gt $d.h) { break }
Start-Sleep -Milliseconds 250
}
Write-Host (" orientation: {0}x{1} ({2})" -f $d.w, $d.h, $(if ($d.w -gt $d.h) { "landscape" } else { "portrait" }))
Start-Sleep -Seconds 2 # let the hide timer run so the video fills its final rect
$ui = Get-UiRetry
$vb = VideoBounds $ui
if (-not $vb) { throw "video (16:9 View) not found in the tree - every coordinate expectation below would be meaningless (N21)" }
Write-Host (" video bounds: l={0} t={1} w={2} h={3} ratio err {4:N4}" -f $vb.l, $vb.t, $vb.w, $vb.h, $vb.err)
if ($vb.err -gt 0.02) { throw "video not 16:9 - geometry changed since 4a, re-derive the mapping here" }
function Expect-OnScreen([int]$cx, [int]$cy) {
# CursorMarker on-screen anchor = video view origin + mapped point.
# PS parses `a / [double]b, c` inside @() as an array construction (the type
# literal + comma is a known pitfall - measured 2026-11-02: this threw
# op_Division on Object[] / silently produced 3 elements). Parenthesize each
# element so each division is evaluated as a scalar first.
@( ($vb.l + $cx * $vb.w / [double]$dW), ($vb.t + $cy * $vb.h / [double]$dH) )
}
function Check-MarkerAt([int]$cx, [int]$cy, [string]$tag, [double]$tol = 6.0) {
Set-CursorQuiet $cx $cy
Start-Sleep -Milliseconds 700 # >= one 500ms idle frame + render
$png = Take-Shot $tag
$mk = Find-Marker $png
if (-not $mk) { throw "V-D1 FAIL [$tag]: no marker at desktop ($cx,$cy) - cyan cluster absent" }
$e = Expect-OnScreen $cx $cy
$dx = [Math]::Abs($mk.x - $e[0]); $dy = [Math]::Abs($mk.y - $e[1])
Write-Host (" [{0}] marker ({1:N0},{2:N0}) n={3} spread={4} expect ({5:N0},{6:N0}) dev=({7:N1},{8:N1})" -f `
$tag, $mk.x, $mk.y, $mk.n, $mk.spread, $e[0], $e[1], $dx, $dy)
if ($dx -gt $tol -or $dy -gt $tol) { throw "V-D1 FAIL [$tag]: deviation ($dx,$dy) > $tol px (shot: $png)" }
if ($mk.spread -gt 90) { throw "V-D1 FAIL [$tag]: cyan spread $($mk.spread)px - too wide to be the marker dot (false positive path)" }
return $mk
}
Write-Host "=== 2. V-D1: four desktop points (incl. near-origin) ==="
# (48,48) instead of (0,0): the dot radius would clip at the exact corner and
# bias the centroid; Some(0,0)-vs-None semantics are already locked by unit +
# golden tests. Corner-proximity error is bounded well below tol here.
Check-MarkerAt 48 48 "v-d1-origin" | Out-Null
Check-MarkerAt 1280 720 "v-d1-center" | Out-Null
Check-MarkerAt 2400 1300 "v-d1-lowright" | Out-Null
Check-MarkerAt 2000 300 "v-d1-upperright" | Out-Null
Write-Host " PASS: marker centroid within 6 px for all 4 points"
Write-Host "=== 3. V-D3: idle-desktop move refreshes the marker <= ~1.2s ==="
# Quiet gate, then one decisive jump; 700ms above already covers 500ms resend.
$from = Check-MarkerAt 300 1200 "v-d3-from"
$e2 = Expect-OnScreen 1600 400
Set-CursorQuiet 1600 400
Start-Sleep -Milliseconds 900
$mk3 = Find-Marker (Take-Shot "v-d3-to")
if (-not $mk3) { throw "V-D3 FAIL: marker vanished after moving the cursor on a quiet desktop" }
$dx3 = [Math]::Abs($mk3.x - $e2[0]); $dy3 = [Math]::Abs($mk3.y - $e2[1])
Write-Host (" after jump: marker ({0:N0},{1:N0}) expect ({2:N0},{3:N0}) dev=({4:N1},{5:N1})" -f `
$mk3.x, $mk3.y, $e2[0], $e2[1], $dx3, $dy3)
if ($dx3 -gt 6 -or $dy3 -gt 6) { throw "V-D3 FAIL: marker did not follow within 900ms (stale cursor field?)" }
Write-Host " PASS: marker refreshes within one idle-resend window"
Write-Host "=== 4. V-D7: cursor into the LEFT second monitor folds to None ==="
$vd = [System.Windows.Forms.SystemInformation]::VirtualScreen
if ($vd.X -ge 0) {
Write-Host " SKIP: no negative-X monitor on this machine - V-D7 (off-screen fold) cannot be tested here; the fold rule is locked by fold_cursor unit tests"
} else {
$ox = [int]($vd.X + 100) # deep into the second display (negative X)
$oy = 500
Set-CursorQuiet $ox $oy
Start-Sleep -Milliseconds 900
$mk7 = Find-Marker (Take-Shot "v-d7-second-monitor")
if ($mk7) { throw "V-D7 FAIL: marker still drawn with cursor at ($ox,$oy) outside the captured display (n=$($mk7.n) at {x=$($mk7.x);y=$($mk7.y)})" }
Write-Host " PASS: no cyan cluster with the pointer at ($ox,$oy) - folded to None end to end"
}
Write-Host "=== 5. tidy: park the pointer inside the primary screen ==="
Set-CursorQuiet ([int]($dW/2)) ([int]($dH/2))
Write-Host "=== CURSOR PROOF COMPLETE ==="