feat(mobile): PR review discussion UX + toast entry feedback

.kilo_workflow/learnings/mobile-android-a11y-label-leaves-geometric-grouping.md

@@ -0,0 +1,19 @@
+# mobile/android: RN a11y-label container nodes are LEAVES — group card content by geometric containment
+
+Symptom: an Appium page-source parse shows a PR-review thread card's label node
+(`Discussion thread ...`) with its full card bounds but ZERO children; the snippet,
+comments, and pills look "outside" the card when checked by XML parentage.
+
+Cause: on Android (uiautomator2), React Native hoists a labeled container's accessible
+children to siblings in the flattened a11y tree; the label node keeps the card's rect
+but stays a leaf.
+
+Fix: assign content to cards GEOMETRICALLY — a node belongs to the card whose
+label-node rect contains it (`inside(cardRect, nodeRect)` with a 2px tolerance). Works
+with the viewport-clamp learning (clamped bands still nest). Two companion traps:
+`scrollUntilVisible` stops the instant the label peeks past the clamp edge — the card's
+header controls are not necessarily rendered yet, so swipe in a bounded loop until the
+card's own control is INSIDE its rect; and after a settle swipe, re-query rects (the
+card moved) — a stale rect in a swipe loop reads as missing content. Expansion state is
+component state and survives Appium sessions: remount the tab (tap sibling tab, tap
+back) at flow start when a flow depends on default expansion.

.kilo_workflow/learnings/mobile-android-pr-review-e2e-toast-a11y-deeplink-dedupe.md

@@ -0,0 +1,49 @@
+# mobile/android PR-review E2E: toast a11y-invisibility, deep-link dedupe, disabled-Pressable attribute (pr-review-ux-7f22 r2, 2026-07-30)
+
+Non-machine-specific techniques confirmed on the pixel9 API35 emulator:
+
+1. **sonner-native toasts are NOT in the uiautomator tree at all.** A
+   `toast.error('Clipboard is empty')` rendered visually for ~4-8s while 10
+   consecutive UiSelector polls (text + content-desc, full-string) returned
+   zero matches. Detection must be pixel-based: screenshot ~1.2s after the
+   triggering tap, crop the top band (~y17-237 on 1080x2424), diff against a
+   baseline frame; auto-dismiss shows up as the band returning to baseline.
+   Extends `android-picker-tap-races-and-toast-capture` (which assumed the
+   toast is catchable by timing; on this build it is never in the tree).
+
+2. **RN disabled Pressable on Android exports `enabled="false"` but keeps
+   `clickable="true"`** (handler artifact). Assert `enabled === 'false'`, and
+   for a read-only control also do a functional no-op probe (elementClick,
+   then assert state + backend request log unchanged).
+
+3. **Same-route `mobile: deepLink` dedupes**: the router keeps the mounted
+   screen, so scroll offset, expansion state, and uncontrolled-field text all
+   survive across verifier runs. Scroll-to-top loops must not stop at the
+   first a11y sliver of a card (clamped band) — scroll until the needed child
+   (e.g. the first hunk gutter line) intersects the card rect. For a truly
+   fresh mount: `stopApp` + `launchApp`, then wait for a shell element
+   (`Home, tab, 1 of 4`) BEFORE the deep link — a link fired during boot is
+   dropped.
+
+4. **CSS `uppercase` transforms the a11y text**: the Resolved badge matches
+   `RESOLVED`, not `Resolved`. Check the source for text-transform classes
+   before pinning selectors.
+
+5. **DiffLine a11y labels (`Added line 9: ...`) are not exported to
+   uiautomator**; the visible per-line signal is the gutter glyph text
+   (`+ 9`, `- 8`, `· 15`) plus the merged code text node. Scope glyph/text
+   assertions to the card rect (sibling cards render the same strings).
+
+6. **Android uncontrolled-field cleanup**: the entry screen's clear button
+   calls `TextInput.clear()`, which on Android unmounts the button (state
+   says empty) but leaves the native text. Use driver-level
+   `eraseText()` (elementClear) — a real edit that fires onChangeText and
+   clears dependent helpers.
+
+7. **adb wedge recovery variant**: guest wedged under host load (3 slots,
+   load ~30-60) presented as adbd timeouts -> `adb devices` shows the device
+   OFFLINE with qemu at 0% CPU. `adb root` (then `unroot`) restarted adbd and
+   the guest came back; bounded polls of `getprop sys.boot_completed`,
+   `pidof system_server`, and `pm list packages` confirm health before
+   rerunning login.sh. No emulator relaunch needed. Same family as
+   `android-emulator-systemserver-restart-systemui-anr-under-load`.

.kilo_workflow/learnings/mobile-e2e-fullstring-regex-selector-escaping.md

@@ -0,0 +1,22 @@
+# mobile e2e: helper patterns are full-string REGEXES — escape parens/dots/plus in literal labels
+
+Symptom: a flow's `waitVisible('Discussion thread src/alpha.ts L10 (RIGHT)')` times out
+for 20s while the exact card is on screen; `scrollUntilVisible('Conversation comment at T+2')`
+overscrolls to the list bottom; an invariant that tolerates empty `findRects` results then
+passes VACUOUSLY.
+
+Cause: `e2e/wdio/helpers.js` `findAll` feeds the pattern to Android UiSelector
+`textMatches`/`descriptionMatches` (whole-string Java regex) or the iOS `MATCHES` predicate.
+Literal labels containing regex specials silently never match: `(RIGHT)` reads as a group
+(missing the literal parens), `.` matches anything, and `T+2` reads as "one or more T then 2".
+Worse, `(await findRects(bad))` returns `[]`, so "no X intersects Y" invariants pass with the
+target never located.
+
+Fix: in flow files, wrap every literal label in an escaper and assert presence before geometry:
+
+```js
+const rx = s => new RegExp('^' + s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&') + '$');
+```
+
+A `^...$` anchor is harmless (the driver matches whole strings anyway). When a wait/scroll
+times out on something visibly on screen, suspect the pattern before suspecting the app.

.kilo_workflow/learnings/mobile-e2e-scrolluntilvisible-lands-but-throws.md

@@ -0,0 +1,14 @@
+# mobile: scrollUntilVisible throws "scrolled N times without finding" but lands on the target
+
+Symptom: `scrollUntilVisible('<full label>')` throws after exhausting maxScrolls, yet a
+hierarchy dump taken immediately after shows the target fully visible and correctly
+positioned (observed twice on iOS PR-review Discussion tab, pr-review-ux-7f22).
+
+Cause: the helper checks visibility between flicks; the final flick moves the list
+past the last check point (momentum + FlashList re-layout), so the loop exhausts while
+the end state is fine. It is the same flick-overshoot family as
+`mobile-e2e-top-clip-positioning-flashlist-clamp` — not a new defect.
+
+Fix: after a failed scrollUntilVisible, dump the hierarchy before retrying or
+classifying — the target is often already on screen. For flows that must not throw,
+wrap in `.catch(() => {})` and follow with an explicit `assertVisible` probe.

.kilo_workflow/learnings/mobile-ios-pr-link-open-button-clearx-shift.md

@@ -0,0 +1,18 @@
+# mobile iOS: PR-link entry screen — Open button shifts +7pt when the field gains content (clear-X)
+
+Symptom: verifying "invalid link renders INLINE in the reserved slot, Open button bounds
+unchanged" by recording Open bounds on the empty entry screen, pasting `not-a-url`, and
+re-measuring FAILS: Open moves y 236→243 and grows h 39→40 even though the helper slot is
+documented as layout-stable.
+
+Cause: the in-field clear-X (`Clear pull request link`, h-13/w-13 ≈ 46pt) appears whenever the
+field has content; it is taller than the 39pt input, so the input ROW grows 39→46 and pushes the
+helper slot and Open button down ~7pt. This is field-CONTENT state, not helper-message state —
+the clear-X predates the pr-review-ux-7f22 changes (present at r1 head `dd659d4a9`).
+
+Fix (decisive experiment for the slot invariant): hold field state constant across the helper
+transition — paste `not-a-url` (helper visible, clear-X present), tap the clear-X: the shipped
+clear handler does NOT clear `helperMessage`, so the invalid helper stays mounted with an empty
+field, and Open returns EXACTLY to the initial rect ({x:21,y:236,w:360,h:39} on iPhone 17).
+That byte-identical comparison is the assertion the AC intends; the +7pt with a filled field is
+pre-existing approved behavior, not a regression.

.kilo_workflow/learnings/mobile-ios-xcuitest-a11y-tree-geometry-containment.md

@@ -0,0 +1,29 @@
+# mobile iOS: XCUI tree — a11y-labeled parent Views are leaves; assert by geometry, not XML ancestry
+
+Symptom: a flow parses `driver.getPageSource()` to assert "snippet inside thread card" via XML
+parent/child containment and finds the card node has ZERO descendants — while pills, snippets and
+comment texts all render on screen. Plain-text labels also appear exactly TWICE in the tree, and
+`visible="true"` matches nothing, so visibility filtering empties the probe.
+
+Cause (iOS 26.5 sim, RN 0.86, PR-review Discussion tab, verified 2026-07-30):
+- A RN View with `accessibilityLabel` (thread cards, header Pressables) becomes ONE accessibility
+  element; its children are NOT XML descendants — they appear as geometric SIBLINGS elsewhere in
+  the tree. Containment must be rect-in-rect on the page-source coordinates (`x/y/width/height`).
+- Plain RN Texts are mirrored across two windows at identical rects — dedupe by
+  (label,x,y,width,height) before counting, and never assert `count === 1` on a Text label
+  (transiently or consistently 2). Interactive elements (Pressables with labels) appear once.
+- Badge Texts using NativeWind `uppercase` expose the TRANSFORMED string (`RESOLVED`, `OUTDATED`,
+  `FILE`) — match the uppercase form.
+- Off-screen FlashList rows stay mounted in the tree; the `visible` attribute is useless (never
+  `true`). Before coordinate-tapping a node, require `0 <= y && y+h <= screenHeight`; swipe
+  (clamped, harmless) until the node is in the viewport.
+- Diff-line gutters (line numbers, +/- markers) are `accessibilityElementsHidden`; the code
+  container's `buildDiffLineAccessibilityLabel` (`Deleted line 8: ...`) is NOT what reaches the
+  tree — the inner selectable RNText's raw code string is. Assert snippet content via code texts
+  (`// stub change`, `return 1;`), and rely on screenshots for line-number evidence.
+- Emoji labels (`👍 reaction, 2 reactions`) match fine through the helpers' predicate path.
+
+Fix: parse page source into a rect tree (fast-xml-parser from the repo's `.pnpm` store requires
+fine from scratch flows), keep `flat()` (dedupe) + `insideRect()` + `inViewport()` helpers, and
+tap by node-rect centre via W3C actions instead of global `tapOn` when the same label exists in
+several cards (e.g. `Resolve thread`).