peachsim 0.1.1620 → 0.1.1635

package/cli/commands/inspect/settling.ts

@@ -1,3 +1,7 @@
+import {
+  accessibilityBounds,
+  sameNativeAccessibilityShape,
+} from '../../maestro-accessibility'
 import {
   inspectReadySnapshot,
   inspectWaitReady,
@@ -16,9 +20,11 @@ export type WaitForSootsimIdleOptions = {
   // in fewer polls instead of paying the count at its slower rate.
   stablePolls?: number
   strict?: boolean
-  // compare consecutive full-screen pixels, as maestro's ios driver does.
+  // compare consecutive full-screen pixels; explicit animation waits use this.
   // frame publications and caret blinks do not require a continuous window.
   paintStill?: boolean
+  // native maestro ios: pixel stability first, then native hierarchy equality.
+  maestro?: boolean
   // settle on motion alone, for the waits around a gesture: image loads and
   // network traffic move nothing on screen, and the hierarchy walk costs
   // seconds per poll on a large tree under load.
@@ -75,6 +81,7 @@ export async function waitForSootsimIdle({
   stablePolls = 3,
   strict = false,
   paintStill = false,
+  maestro = false,
   motionOnly = false,
   transitionStartWindowMs = 80,
   transitionTimeoutMs = 4000,
@@ -108,12 +115,15 @@ export async function waitForSootsimIdle({
       const pollMs = ${Math.max(1, Math.round(pollMs))}
       const strict = ${strict ? 'true' : 'false'}
       const paintStill = ${paintStill ? 'true' : 'false'}
+      const maestro = ${maestro ? 'true' : 'false'}
+      const accessibilityBounds = ${accessibilityBounds.toString()}
+      const sameNativeAccessibilityShape = ${sameNativeAccessibilityShape.toString()}
       const motionOnly = ${motionOnly ? 'true' : 'false'}
       // full mounted-geometry observation needs at least two hierarchy reads
       // plus the stillness window, which a sub-second budget cannot afford:
       // tight budgets (auto-wait) settle on worker stillness plus root
       // membership, generous ones confirm the geometry too.
-      const fullGeometry = ${!motionOnly && maxMs >= 1000 ? 'true' : 'false'}
+      const fullGeometry = ${!maestro && !motionOnly && maxMs >= 1000 ? 'true' : 'false'}
       const requiredStillMs = paintStill ? 0 : ${Math.max(1, Math.round(stablePolls)) * Math.max(1, Math.round(pollMs))}
       const minimumStablePolls = paintStill ? 1 : ${Math.min(2, Math.max(1, Math.round(stablePolls)))}
       const composite = paintStill ? window.SootSim?.bridges?.liveComposite : null
@@ -235,6 +245,9 @@ export async function waitForSootsimIdle({
           // strict also waits out shell motion: a system alert, sheet or menu
           // animates in the shell worker while the app's own stats read still,
           // so its presentation is complete only once the shell's list empties.
+          const accessibilityRead = maestro
+            ? answer(() => window.__sootsimTest.captureAccessibilityTree()).catch(() => null)
+            : null
           const motionRead = strict
             ? answer(() => window.__sootsimTest?.listAnimations?.()).catch(() => null)
             : null
@@ -244,9 +257,12 @@ export async function waitForSootsimIdle({
           // the compositor rejects a stats request it is too busy to answer.
           const compositorRead = answer(() => window.__sootsimCompositor.getStats()).catch(() => BUSY)
           const countsRead = motionOnly ? null : answer(() => window.__sootsimTest?.getRequestCounts?.()).catch(() => null)
-          // mounted-root membership: the live node count moves while the tree
-          // mounts, unmounts or recycles, in either render mode.
-          const nodeCountRead = answer(() => window.__sootsimTest?.getNodeCount?.()).catch(() => null)
+          // mounted-root membership: the count of nodes that can reach the
+          // window moves while the screen mounts, unmounts or recycles, in
+          // either render mode. a screen a pushed one covers or an inactive
+          // tab stays mounted out of the window, so what it mounts is not
+          // screen motion.
+          const nodeCountRead = answer(() => window.__sootsimTest?.getWindowNodeCount?.()).catch(() => null)
           // the mounted hierarchy in render-worker mode comes from the test
           // bridge: window.__sootsimRoot is an empty placeholder there, so a
           // walk over it hashes every screen to the same empty string while
@@ -276,7 +292,7 @@ export async function waitForSootsimIdle({
             const motion = await motionRead
             if (motion === BUSY) workerBusy = true
             else if (Array.isArray(motion)) {
-              const blocking = motion.filter((anim) => anim.realm === 'shell' && anim.priority !== 'low')
+              const blocking = motion.filter((anim) => anim.realm === 'shell' && anim.priority !== 'low' && anim.visible !== false)
               shellAnimating = blocking.length > 0
               // name what holds the settle so a timeout points at its owner.
               shellBlockers = blocking
@@ -308,6 +324,8 @@ export async function waitForSootsimIdle({
               stats.hasActiveAnims === true ||
               stats.hasActiveBlockingNativeAnimations === true ||
               (!paintStill && stats.hasPendingBlockingAnimationFrames === true)
+            // the runtime reports only layout work that can reach the window:
+            // a covered screen or inactive tab committing does not hold it.
             layoutDirty =
               stats.layoutDirty === true && stats.renderRequested === true
             // a freshly-pushed screen showing a skeleton is layout-stable
@@ -340,6 +358,11 @@ export async function waitForSootsimIdle({
           const total = await nodeCountRead
           if (total === BUSY) workerBusy = true
           nodeCount = typeof total === 'number' ? total : null
+          const accessibility = accessibilityRead ? await accessibilityRead : null
+          const nativeHierarchy = Array.isArray(accessibility) && accessibility.length > 0
+            ? accessibility
+            : null
+          if (accessibility === BUSY) workerBusy = true
           let hierarchy = hierarchyRead ? await hierarchyRead : null
           if (hierarchy === BUSY) {
             workerBusy = true
@@ -399,9 +422,15 @@ export async function waitForSootsimIdle({
               webViewStarting = true
               webViewBlocker = 'preview progress element displayed'
             } else {
-              const unreadyWebView = document.querySelector(
+              // a loading document with no area (an app's hidden worker or
+              // executor frame) paints nothing, and a screenshot on device
+              // does not wait for it.
+              const unreadyWebView = [...document.querySelectorAll(
                 '[data-sootsim-webview-loading="true"], [data-sootsim-webview-mode="placeholder"]'
-              )
+              )].find((view) => {
+                const frame = view.querySelector('iframe')
+                return !frame || frame.clientWidth > 0 || frame.clientHeight > 0
+              })
               if (unreadyWebView) {
                 webViewStarting = true
                 webViewBlocker = 'webView document still loading'
@@ -420,7 +449,7 @@ export async function waitForSootsimIdle({
               pixelFrameToken = token
             }
           }
-          return { layout: nodes.join(';'), screenPixels, animating, shellAnimating, shellBlockers, layoutDirty, pendingFetches, workerBusy, webViewStarting, webViewBlocker, renderStatsAvailable, requestTotal, requestInFlight, appCommits, nodeCount }
+          return { layout: nodes.join(';'), nativeHierarchy, screenPixels, animating, shellAnimating, shellBlockers, layoutDirty, pendingFetches, workerBusy, webViewStarting, webViewBlocker, renderStatsAvailable, requestTotal, requestInFlight, appCommits, nodeCount }
         }
 
         // how long generic background network is allowed to keep us waiting AFTER
@@ -434,6 +463,7 @@ export async function waitForSootsimIdle({
         // so the generic request counters are a courtesy, not a gate: honor them
         // only within this grace once the screen is visually still.
         const networkQuietGraceMs = 1200
+        let lastNativeHierarchy = null
         let lastLayout = ''
         let lastScreenPixels = null
         let screenChangedPolls = 0
@@ -461,6 +491,9 @@ export async function waitForSootsimIdle({
           }
           return true
         }
+        const hierarchyUnchanged = (current, previous) =>
+          current !== null && previous !== null && current.length === previous.length &&
+          current.every((node, index) => sameNativeAccessibilityShape(node, previous[index], accessibilityBounds))
         const IDLE_HELD = { idleHeld: true }
         while (true) {
           if (clockNow() >= deadline) {
@@ -506,7 +539,7 @@ export async function waitForSootsimIdle({
           // the loop head drains a transition still in flight before it ends.
           if (snapshot.workerBusy && clockNow() >= deadline && report) continue
           polls++
-          report = { snapshot, layout: lastLayout, screenPixels: lastScreenPixels, requestTotal: lastRequestTotal, appCommits: lastAppCommits, nodeCount: lastNodeCount }
+          report = { snapshot, nativeHierarchy: lastNativeHierarchy, layout: lastLayout, screenPixels: lastScreenPixels, requestTotal: lastRequestTotal, appCommits: lastAppCommits, nodeCount: lastNodeCount }
           if (snapshot.layoutDirty) dirtyPolls++
           if (snapshot.animating || snapshot.shellAnimating) animatingPolls++
           if (snapshot.webViewStarting) webViewStartingPolls++
@@ -527,7 +560,15 @@ export async function waitForSootsimIdle({
           // strictOk and adds shell motion plus the blocking-stats floor. a
           // null count or commit reading means the runtime does not report
           // it, which reads as stable the way a missing stats block does.
-          const visuallyStill =
+          // native screen checks observe output, not animation ownership. after
+          // its 3 s pixel phase, iOS accepts two equal native TreeNodes even
+          // while paint-only playback continues. overlap hierarchy reads so
+          // neither the gesture nor screenshot deadline grows.
+          const maestroStill = snapshot.screenPixels !== null && !snapshot.workerBusy && !snapshot.webViewStarting && (
+            !screenChanged || (clockNow() - start >= 3000 &&
+              hierarchyUnchanged(snapshot.nativeHierarchy, lastNativeHierarchy))
+          )
+          const visuallyStill = maestro ? maestroStill :
             strictOk &&
             !snapshot.animating &&
             !snapshot.layoutDirty &&
@@ -555,7 +596,7 @@ export async function waitForSootsimIdle({
             // grace (continuous background traffic must not block forever).
             if (
               stillProven() &&
-              (motionOnly || networkQuiet || clockNow() - visuallyStillSince >= networkQuietGraceMs)
+              (maestro || motionOnly || networkQuiet || clockNow() - visuallyStillSince >= networkQuietGraceMs)
             ) {
               if (transitionWait) {
                 await transitionWait
@@ -567,6 +608,7 @@ export async function waitForSootsimIdle({
             stillPolls = []
             visuallyStillSince = 0
           }
+          lastNativeHierarchy = snapshot.nativeHierarchy
           lastLayout = snapshot.layout
           lastScreenPixels = snapshot.screenPixels
           lastRequestTotal = snapshot.requestTotal
@@ -589,20 +631,21 @@ export async function waitForSootsimIdle({
         }
         const last = report.snapshot
         const blockers = []
-        if (strict && !last.renderStatsAvailable) blockers.push('render stats unavailable')
-        if (last.animating) blockers.push('animations still running')
-        if (last.shellAnimating) blockers.push('shell animations still running (' + last.shellBlockers + ')')
-        if (last.layoutDirty) blockers.push('layout still dirty')
-        if (last.appCommits !== null && report.appCommits !== null && last.appCommits !== report.appCommits) blockers.push('layout commits still landing')
-        if (last.nodeCount !== null && report.nodeCount !== null && last.nodeCount !== report.nodeCount) blockers.push('node count still changing')
-        if (!motionOnly && last.pendingFetches > 0) blockers.push(last.pendingFetches + ' image fetches pending')
+        if (!maestro && strict && !last.renderStatsAvailable) blockers.push('render stats unavailable')
+        if (!maestro && last.animating) blockers.push('animations still running')
+        if (!maestro && last.shellAnimating) blockers.push('shell animations still running (' + last.shellBlockers + ')')
+        if (!maestro && last.layoutDirty) blockers.push('layout still dirty')
+        if (!maestro && last.appCommits !== null && report.appCommits !== null && last.appCommits !== report.appCommits) blockers.push('layout commits still landing')
+        if (!maestro && last.nodeCount !== null && report.nodeCount !== null && last.nodeCount !== report.nodeCount) blockers.push('node count still changing')
+        if (!maestro && !motionOnly && last.pendingFetches > 0) blockers.push(last.pendingFetches + ' image fetches pending')
         if (last.workerBusy) blockers.push('a render worker too busy to report')
         if (last.webViewStarting) blockers.push(last.webViewBlocker || 'webView still starting')
-        if (last.layout !== report.layout) blockers.push('layout still changing')
+        if (!maestro && last.layout !== report.layout) blockers.push('layout still changing')
+        if (maestro && !hierarchyUnchanged(last.nativeHierarchy, report.nativeHierarchy)) blockers.push(last.nativeHierarchy === null ? 'native accessibility hierarchy unavailable' : 'native accessibility hierarchy still changing')
         if (paintStill && !last.screenPixels) blockers.push('screen pixels unavailable')
         else if (paintStill && !screenUnchanged(last.screenPixels, report.screenPixels)) blockers.push('screen pixels still changing')
-        if (!motionOnly && last.requestInFlight > 0) blockers.push(last.requestInFlight + ' requests in flight')
-        else if (!motionOnly && last.requestTotal !== report.requestTotal) blockers.push('new requests issued')
+        if (!maestro && !motionOnly && last.requestInFlight > 0) blockers.push(last.requestInFlight + ' requests in flight')
+        else if (!maestro && !motionOnly && last.requestTotal !== report.requestTotal) blockers.push('new requests issued')
         if (polls < 2) blockers.push('sim too busy to poll twice')
         // the last read is still: motion ended too near the deadline for the
         // stillness window to hold before it.
@@ -654,6 +697,9 @@ export type SettledWaitReady = {
   // the last readiness read: the settled tree once the app went idle,
   // otherwise the read that ended the wait.
   status: WaitReadyStatus
+  // the first read that counted as ready, before the settle: the probe the
+  // wait accepted. null when the app never became ready.
+  accepted: WaitReadyStatus | null
   // the last settle; null when the app never became ready.
   idle: WaitForSootsimIdleResult | null
   elapsedMs: number
@@ -672,8 +718,10 @@ export async function waitReadyAndSettle(
 ): Promise<SettledWaitReady> {
   const start = Date.now()
   const remaining = () => Math.max(0, start + timeoutMs - Date.now())
+  let accepted: WaitReadyStatus | null = null
   const result = (status: WaitReadyStatus, idle: WaitForSootsimIdleResult | null) => ({
     status,
+    accepted,
     idle,
     elapsedMs: Date.now() - start,
   })
@@ -682,6 +730,7 @@ export async function waitReadyAndSettle(
     idle = null
     const ready = await inspectWaitReady(bridge, remaining(), options)
     if (!ready.ready) return result(ready, idle)
+    accepted = ready
     idle = await waitForSootsimIdle({
       bridge,
       simId: options.simId,

package/cli/commands/inspect/wait-ready.ts

@@ -26,10 +26,12 @@ export async function runWaitReadySubcommand(opts: {
   if (bridge.plane === 'cloud') {
     const status = await inspectWaitReady(bridge, timeoutMs)
     if (!status.ready) timedOut(status, status.elapsedMs)
+    printAcceptedProbe(status)
     printReady(status, status.elapsedMs)
     return
   }
   let status: Awaited<ReturnType<typeof waitReadyAndSettle>>['status']
+  let accepted: Awaited<ReturnType<typeof waitReadyAndSettle>>['accepted']
   let idle: Awaited<ReturnType<typeof waitReadyAndSettle>>['idle']
   let elapsedMs: number
   try {
@@ -41,6 +43,7 @@ export async function runWaitReadySubcommand(opts: {
       },
     })
     status = settledWait.status
+    accepted = settledWait.accepted
     idle = settledWait.idle
     elapsedMs = settledWait.elapsedMs
   } catch (error) {
@@ -51,6 +54,7 @@ export async function runWaitReadySubcommand(opts: {
     throw error
   }
   if (!idle) timedOut(status, elapsedMs)
+  if (accepted) printAcceptedProbe(accepted)
   if (!idle.settled) {
     console.error(
       `  ⚠ wait ready timed out after ${elapsedMs}ms — app mounted but did not settle: ${idle.blockedBy} (nodes: ${status.nodes}, targets: ${status.targets})`,
@@ -66,6 +70,14 @@ export async function runWaitReadySubcommand(opts: {
   printReady(status, elapsedMs)
 }
 
+// the probe the wait accepted, kept so a ready that later proves wrong can be
+// read from the command's own output.
+function printAcceptedProbe(status: WaitReadyStatus): void {
+  console.error(
+    `  accepted probe at ${Math.round(status.elapsedMs)}ms: basis ${status.readyBasis}, flag ${String(status.flag)}, ${status.nodes} nodes (quiet ${Math.round(status.treeQuietMs)}ms), ${status.targets} targets, painted ${status.paintedAppContent}, live frame ${status.liveFrameActive} (${status.liveFrameChannels} channels, ${status.liveFramePublishes} publishes), browser surfaces ${status.browserSurfaceCount}`,
+  )
+}
+
 function printReady(status: WaitReadyStatus, elapsedMs: number): void {
   console.log(
     `  ready in ${elapsedMs}ms: ${status.nodes} nodes, ${status.targets} targets${status.liveFrameActive ? `, ${status.liveFrameChannels} live surfaces` : ''}${status.browserSurfaceCount ? `, ${status.browserSurfaceCount} browser surfaces` : ''}`,

package/cli/commands/inspect.ts

@@ -34,6 +34,7 @@ import {
   parseBridgeCliArgs,
   printBridgeWorldNotice,
   type WsBridge,
+  type BridgeSimInfo,
 } from '../ws-bridge'
 import {
   isTapSuccess,
@@ -596,13 +597,13 @@ export function printShellPerfReport(result: Record<string, any>) {
     console.log(``)
     console.log(`    cadence (frame delivery, ${cadence.vsyncTicks} vsync ticks):`)
     console.log(
-      `      display clock:  ${cadence.displayHz}hz · host rAF interval p50 ${cadence.hostIntervalP50} / p95 ${cadence.hostIntervalP95} / max ${cadence.hostIntervalMax} ms${cadence.hostGaps ? ` · ${cadence.hostGaps} gaps >1.5x (host rAF starved)` : ''}`,
+      `      display clock:  ${cadence.displayHz}hz · host rAF interval p50 ${cadence.hostIntervalP50} / p95 ${cadence.hostIntervalP95} / max ${cadence.hostIntervalMax} ms${cadence.hostGaps ? ` · ${cadence.hostGaps} gaps >1.5x (host rAF starved)` : ''}${cadence.idleBreaks ? ` · ${cadence.idleBreaks} host idle breaks excluded` : ''}`,
     )
     console.log(
       `      delivery lag:   p50 ${cadence.deliveryLagP50} / p95 ${cadence.deliveryLagP95} / max ${cadence.deliveryLagMax} ms (vsync postMessage → shell receipt)`,
     )
     console.log(
-      `      paint interval: p50 ${cadence.paintIntervalP50} / p95 ${cadence.paintIntervalP95} / max ${cadence.paintIntervalMax} ms${cadence.idleBreaks ? ` · ${cadence.idleBreaks} idle breaks excluded` : ''}`,
+      `      present interval: p50 ${cadence.paintIntervalP50} / p95 ${cadence.paintIntervalP95} / max ${cadence.paintIntervalMax} ms (during motion, ${cadence.paintIntervals} intervals, including native frames)`,
     )
   }
   const compositorCadence = result.compositorCadence
@@ -1247,7 +1248,7 @@ export async function runInspect(args: string[], opts: InspectOptions) {
     return
   }
 
-  // `peach list --drivers` — pure local registry view, no bridge needed.
+  // driver discovery works offline; a live bridge also reports each running renderer.
   // short-circuit here so users can discover drivers even when no peach
   // session is running.
   if (
@@ -1256,6 +1257,19 @@ export async function runInspect(args: string[], opts: InspectOptions) {
   ) {
     const { buildDriverListRows } = await import('../drivers')
     const rows = buildDriverListRows()
+    const bridge = (opts.createBridge ?? createBridgeFromParsed)(parsed)
+    let sims: BridgeSimInfo[] = []
+    try {
+      sims = await bridge.listSims()
+    } catch {
+      // local availability remains useful when no bridge is running.
+    } finally {
+      bridge.close()
+    }
+    if (wantsJson(effectiveArgs)) {
+      printJson({ drivers: rows, sims })
+      return
+    }
     console.log(`  available drivers (${rows.length}):\n`)
     const idWidth = Math.max(...rows.map((r) => r.id.length), 6)
     const kindWidth = Math.max(...rows.map((r) => r.kind.length), 4)
@@ -1270,6 +1284,15 @@ export async function runInspect(args: string[], opts: InspectOptions) {
         console.log(`    unavailable: ${row.reason}`)
       }
     }
+    for (const sim of sims) {
+      const renderer = sim.meta?.peachRenderer
+      const executable = sim.meta?.peachHostExecutable
+      if (sim.readyState === 'open' && typeof renderer === 'string') {
+        console.log(
+          `  in use: ${sim.id} · ${String(sim.meta?.sootsimHostDriver)} · ${renderer}${typeof executable === 'string' ? ` @ ${executable}` : ''}`,
+        )
+      }
+    }
     return
   }
 

package/cli/commands/install-desktop.ts

@@ -126,6 +126,7 @@ async function installLinuxDeb(packagePath: string): Promise<string> {
 async function installWindowsExe(exePath: string): Promise<string> {
   // nsis installer prompts and self-elevates; run it and let it take over.
   await new Promise<void>((resolvePromise, rejectPromise) => {
+    // runner-toolchain-platform: win32
     const child = spawn('cmd', ['/c', 'start', '""', '/wait', exePath], {
       stdio: 'inherit',
     })

package/cli/commands/maestro-compare.ts

@@ -8,21 +8,23 @@ import * as fs from 'fs'
 import * as os from 'os'
 import * as path from 'path'
 import { devices } from 'peach-engine/settings/devices'
+import { PNG } from 'pngjs'
 import { rnxHomeDir } from '../../src/home-paths'
 import { parseFlowFile } from '../flow-file'
 import {
   buildCompareHtml,
+  buildCompareSide,
   compareScreens,
   comparePassed,
   instrumentFlowSource,
   nativeStepsFromCommandsJson,
-  nativeFlowFailureFromLog,
+  nativeFlowVerdictsFromJunit,
+  junitFlowKey,
   peachStepsFromCommandsJson,
   screensOverThreshold,
   type CompareCapture,
   type CompareFlow,
   type CompareReport,
-  type CompareSide,
 } from '../maestro-compare'
 import { MaestroJsContext } from '../maestro-js'
 import { collectFlowFiles, loadMaestroWorkspaceConfig } from '../maestro-workspace'
@@ -161,12 +163,44 @@ function statusBarPx(png: Buffer): number {
   const height = png.readUInt32BE(20)
   const spec = Object.values(devices).find(
     (d) =>
-      Math.round(d.width * d.scale) === width &&
-      Math.round(d.height * d.scale) === height,
+      // maestro can truncate fractional point bounds one pixel below simctl.
+      Math.abs(Math.round(d.width * d.scale) - width) <= 1 &&
+      Math.abs(Math.round(d.height * d.scale) - height) <= 1,
   )
   return spec ? spec.statusBarHeight * spec.scale : 0
 }
 
+// maestro's xctest shot on an ios 27 simulator drops the rightmost column of
+// an odd framebuffer (1179 becomes 1178). diffing at that short width
+// resamples peach off the pixel grid, so restore the column first.
+function alignNativeCapture(png: Buffer): Buffer {
+  const width = png.readUInt32BE(16)
+  const height = png.readUInt32BE(20)
+  const spec = Object.values(devices).find(
+    (d) =>
+      Math.abs(Math.round(d.width * d.scale) - width) <= 1 &&
+      Math.abs(Math.round(d.height * d.scale) - height) <= 1,
+  )
+  if (!spec) return png
+  const expectedW = Math.round(spec.width * spec.scale)
+  const expectedH = Math.round(spec.height * spec.scale)
+  if (expectedH !== height || expectedW !== width + 1) return png
+  const src = PNG.sync.read(png)
+  const dst = new PNG({ width: expectedW, height })
+  for (let y = 0; y < height; y++) {
+    const srow = y * src.width * 4
+    const drow = y * expectedW * 4
+    src.data.copy(dst.data, drow, srow, srow + src.width * 4)
+    const last = srow + (src.width - 1) * 4
+    const o = drow + src.width * 4
+    dst.data[o] = src.data[last] ?? 0
+    dst.data[o + 1] = src.data[last + 1] ?? 0
+    dst.data[o + 2] = src.data[last + 2] ?? 0
+    dst.data[o + 3] = 255
+  }
+  return PNG.sync.write(dst)
+}
+
 // the flow as its author would name it: relative inside the working
 // directory, home-relative or absolute outside it.
 function displayPath(cwd: string, file: string): string {
@@ -201,45 +235,6 @@ function formatFailedRequest(entry: unknown): string {
   return `request failed: ${e.method ?? 'GET'} ${e.url ?? '?'} ${e.status ?? e.error ?? ''}`.trim()
 }
 
-function sideFromSteps(
-  steps: CompareSide['steps'],
-  ran: boolean,
-  extraErrors: string[],
-  // a run that failed outside its steps (the sim or app never came up)
-  runFailure?: string,
-): CompareSide {
-  const failed = steps.find((s) => s.status === 'failed')
-  return {
-    status: !ran ? 'not run' : failed || runFailure ? 'failed' : 'passed',
-    durationMs: steps.reduce((sum, s) => sum + s.durationMs, 0),
-    steps,
-    errors: [
-      ...(failed
-        ? [`step "${failed.label}" failed: ${failed.error ?? 'no message'}`]
-        : runFailure
-          ? [runFailure]
-          : []),
-      ...extraErrors,
-    ],
-  }
-}
-
-// peach names each flow's output dir after the file stem, suffixing repeats
-// in run order (commands/maestro.ts uniqueFlowDir).
-function peachFlowDirs(root: string, files: string[]): string[] {
-  const seen = new Map<string, number>()
-  return files.map((file) => {
-    const slug =
-      path
-        .basename(file)
-        .replace(/\.(ya?ml)$/i, '')
-        .replace(/[^A-Za-z0-9._-]+/g, '_') || 'flow'
-    const n = (seen.get(slug) ?? 0) + 1
-    seen.set(slug, n)
-    return path.join(root, n === 1 ? slug : `${slug}-${n}`)
-  })
-}
-
 // the flow's appId as maestro resolves it: its `${...}` templates evaluate
 // against --env values (in process.env) and the flow's own env block.
 function resolveFlowAppId(frontmatter: Record<string, unknown>): string | undefined {
@@ -447,6 +442,10 @@ export async function runMaestroCompare(
         '--test-output-dir',
         nativeOut,
         '--no-ansi',
+        '--format',
+        'junit',
+        '--output',
+        path.join(out, 'ios-report.xml'),
         ...shared,
         ...envPairs,
       ],
@@ -456,18 +455,23 @@ export async function runMaestroCompare(
     spawnSync('xcrun', ['simctl', 'status_bar', udid, 'clear'], { stdio: 'ignore' })
   }
   console.log(`\n  ios │ maestro exited ${nativeExit}\n`)
-  const nativeLog = fs.readFileSync(path.join(out, 'ios.log'), 'utf8')
+  const nativeVerdicts = nativeFlowVerdictsFromJunit(
+    fs.readFileSync(path.join(out, 'ios-report.xml'), 'utf8'),
+  )
 
   // 2. peach, same instrumented flows (its --env already applied above)
   const peachOut = path.join(out, 'peach')
-  const peachFailures = new Map<string, string>()
+  const peachResults = new Map<
+    string,
+    { exit: number; playbackFailure: string; outputDir: string | null }
+  >()
   await runMaestro(
     ['test', runTarget, '--test-output-dir', peachOut, ...shared, ...peachArgs],
     {
       ...opts,
       captureConsole: true,
-      onFlowResult: ({ file, exit, playbackFailure }) => {
-        if (exit !== 0) peachFailures.set(path.resolve(file), playbackFailure)
+      onFlowResult: ({ file, exit, playbackFailure, outputDir }) => {
+        peachResults.set(path.resolve(file), { exit, playbackFailure, outputDir })
       },
     },
   )
@@ -481,24 +485,56 @@ export async function runMaestroCompare(
         .sort()
         .pop()
     : undefined
-  const peachDirs = peachFlowDirs(
-    peachOut,
-    instrumented.map((f) => f.copy),
-  )
+  const nativeOutputs = new Map<string, Array<{ dir: string; commands: unknown }>>()
+  for (const entry of nativeRun
+    ? fs.readdirSync(nativeRun, { withFileTypes: true })
+    : []) {
+    if (!entry.isDirectory()) continue
+    const dir = path.join(nativeRun!, entry.name)
+    const commands = readJson(path.join(dir, 'commands.json'))
+    if (!Array.isArray(commands)) continue
+    // display names are not paths: maestro replaces slashes and suffixes collisions.
+    // the top-level bookkeeping retains the original flow identity.
+    let name: string | undefined
+    let fileStem: string | undefined
+    for (const command of commands) {
+      if (command?.metadata?.depth !== 0) continue
+      const config = command?.command?.applyConfigurationCommand?.config
+      if (typeof config?.name === 'string') name ??= config.name
+      const env = command?.command?.defineVariablesCommand?.env
+      if (typeof env?.MAESTRO_FILENAME === 'string') fileStem ??= env.MAESTRO_FILENAME
+    }
+    if (fileStem === undefined) continue
+    const key = JSON.stringify([name ?? fileStem, fileStem])
+    const outputs = nativeOutputs.get(key) ?? []
+    outputs.push({ dir, commands })
+    nativeOutputs.set(key, outputs)
+  }
   const imagesDir = path.join(out, 'images')
   const flows: CompareFlow[] = []
   for (const [i, flow] of instrumented.entries()) {
-    const nativeDir = nativeRun ? path.join(nativeRun, flow.name) : null
-    const peachDir = peachDirs[i]!
-    const nativeRan = Boolean(nativeDir && fs.existsSync(nativeDir))
-    const peachRan = fs.existsSync(path.join(peachDir, 'commands.json'))
+    const fileStem = path.basename(flow.copy).replace(/\.(ya?ml)$/i, '')
+    const nativeOutput = nativeOutputs.get(JSON.stringify([flow.name, fileStem]))?.shift()
+    const nativeDir = nativeOutput?.dir
+    const nativeRan = Boolean(nativeOutput)
+    const peachResult = peachResults.get(path.resolve(flow.copy))
+    const peachDir = peachResult?.outputDir
+    const verdicts = nativeVerdicts.get(junitFlowKey(flow.name, fileStem))
+    const nativeFailure = verdicts?.length
+      ? verdicts.shift()
+      : 'no native verdict in ios-report.xml'
+    const peachRan = Boolean(
+      peachDir && fs.existsSync(path.join(peachDir, 'commands.json')),
+    )
     const nativeSteps = nativeRan
-      ? nativeStepsFromCommandsJson(readJson(path.join(nativeDir!, 'commands.json')))
+      ? nativeStepsFromCommandsJson(nativeOutput?.commands)
       : []
     const peachSteps = peachRan
-      ? peachStepsFromCommandsJson(readJson(path.join(peachDir, 'commands.json')))
+      ? peachStepsFromCommandsJson(readJson(path.join(peachDir!, 'commands.json')))
       : []
-    const consoleJson = readJson(path.join(peachDir, 'logs', 'console.json')) as {
+    const consoleJson = (
+      peachDir ? readJson(path.join(peachDir, 'logs', 'console.json')) : null
+    ) as {
       errors?: unknown[]
       requests?: unknown[]
     } | null
@@ -513,15 +549,17 @@ export async function runMaestroCompare(
       const nativePng = nativeDir
         ? path.join(nativeDir, 'takeScreenshot', `${capture.name}.png`)
         : ''
-      const peachPng = path.join(peachDir, 'screenshots', `${capture.name}.png`)
+      const peachPng = peachDir
+        ? path.join(peachDir, 'screenshots', `${capture.name}.png`)
+        : ''
       const hasNative = Boolean(nativePng) && fs.existsSync(nativePng)
-      const hasPeach = fs.existsSync(peachPng)
+      const hasPeach = Boolean(peachPng) && fs.existsSync(peachPng)
       const imageRel = path.join(
         'images',
         `${i + 1}-${capture.name.split('/').pop()}.png`,
       )
       if (hasNative && hasPeach) {
-        const nativeBytes = fs.readFileSync(nativePng)
+        const nativeBytes = alignNativeCapture(fs.readFileSync(nativePng))
         const { diff, triptych } = compareScreens(
           nativeBytes,
           fs.readFileSync(peachPng),
@@ -545,21 +583,21 @@ export async function runMaestroCompare(
     flows.push({
       name: flow.name,
       file: displayPath(cwd, flow.file),
-      native: sideFromSteps(
+      native: buildCompareSide(
         nativeSteps,
-        nativeRan,
+        nativeFailure ? 'failed' : nativeRan ? 'passed' : 'not run',
         nativeRan
           ? []
           : [
               `maestro produced no output for this flow (exit ${nativeExit}); see ios.log`,
             ],
-        nativeFlowFailureFromLog(nativeLog, flow.name),
+        nativeFailure,
       ),
-      peach: sideFromSteps(
+      peach: buildCompareSide(
         peachSteps,
-        peachRan,
+        peachResult ? (peachResult.exit === 0 ? 'passed' : 'failed') : 'not run',
         peachConsole,
-        peachFailures.get(path.resolve(flow.copy)),
+        peachResult?.exit ? peachResult.playbackFailure : undefined,
       ),
       screens,
     })