docs(mobile): Hermes V1 pin A/B measured no RAM reduction

.kilo_workflow/learnings/e2e-slot-release-stopped-another-sections-stack.md

@@ -35,3 +35,15 @@ a `slot-acq` window under a proxy `<section>-e2e-seed` session and parked a
 `sleep` to hold it. That both violates the E2E-loop rule and creates a
 self-deadlock: the verifier polls for a free slot that its own section's
 sleeper holds, and neither side can progress.
+
+Victim-side recovery (worked twice in hermes-mem-c716 r1, the section named in
+the symptom): the kill is silent for the victim — `pnpm dev:status` just prints
+"No dev session running" mid-round and the dev client loses Metro. While the
+slot is still held, restart the stack in place: `pnpm dev:start --no-attach
+mobile cloud-agent-next kiloclaw event-service` (the same ports return), restart
+the CLI relay (`remote-cli.sh exec remote` → `Remote connection enabled.`),
+re-prove `session list --pure`, re-anchor the dev client with the deep link, and
+re-run only the iteration that was in flight — device-local measurements (e.g.
+meminfo) from completed iterations stay valid. The companion `stack=none` shown
+by `e2e-slot.sh status` even for live stacks was the #4836 name-mangling bug
+(the same mangled session name made `has-session` always fail), already fixed.

.kilo_workflow/learnings/mobile-android-maestro-flow-pitfalls.md

@@ -0,0 +1,20 @@
+# mobile-android-maestro-flow-pitfalls
+
+Symptom: a Maestro iteration flow on Android fails or hangs in ways the exit
+code does not reveal.
+
+Cause/fix (three independent traps, all hit in the hermes-mem baseline round):
+1. `maestro test` exits 0 even when a flow assertion FAILED (observed on
+   maestro 2.7.0, Android). Never trust the exit code; grep the log for
+   `FAILED` after every run.
+2. Maestro text matching is full-string regex: the Agents list search field is
+   `Search sessions...` (with ellipsis), so `visible: 'Search sessions'` never
+   matches. Use `Search sessions.*` (and generally re-dump and copy exact
+   strings after any UI change).
+3. Under parallel-workflow host load (load avg >100) a cold dev-client launch
+   can sit on a blank Compose splash for 3-15 min and the emulator throws
+   `Process system isn't responding` ANR dialogs. `open-app.yaml`'s 30 s wait
+   is too short. A verifier temp flow needs: a 420 s launch wait, a
+   `.*isn.t responding.*` → `tapOn: 'Wait'` handler after the launch wait and
+   after settle, and 60-120 s waits for list/transcript asserts. `dumpsys
+   meminfo` itself can hit its 10 s service timeout under load — retry it.

.kilo_workflow/learnings/mobile-android-stale-claims-explicit-port.md

@@ -0,0 +1,20 @@
+# mobile-android-stale-claims-explicit-port
+
+Symptom: `pnpm dev:mobile:android claim emulator-5554` fails with "claimed by
+/Users/igor/Projects/.worktrees/<other-worktree>" even though the owning
+worktree has no live tmux session and the claiming PID is dead.
+
+Cause: Android claims are per-serial JSON files in
+`$TMPDIR/kilo-mobile-android-claims/`. Before 96154f115 ownership was checked
+by worktreeRoot equality only, so a claim was never stale while the owning
+worktree directory existed, and dead-emulator claims wedged the default
+serials (emulator-5554, emulator-5556) permanently. Claims now record the
+guest kernel boot id, so a foreign claim self-clears once the emulator that
+wrote it no longer answers on the serial — but a *live* foreign emulator on a
+default serial still blocks your boot, which lands on the first free serial.
+
+Fix: boot the AVD on an explicit free even console port so the serial is
+unclaimed: `pnpm dev:mobile:android emulator -avd <avd> -port 5558
+-no-snapshot-save -no-boot-anim -gpu host` → serial `emulator-5558`, then
+`claim emulator-5558`. Never delete another worktree's claim file. Record the
+port in the round handoff so the next round reuses the same serial.

.kilo_workflow/learnings/mobile-e2e-cli-relay-window-wedge.md

@@ -0,0 +1,17 @@
+# mobile-e2e-cli-relay-window-wedge
+
+Symptom: the `relay` window of an orchestrator-prepared
+`kilo-e2e-cli-<worktree>` tmux session shows an idle shell (`pane_current_command`
+= zsh), echoes typed input, but never executes it (no output, no new prompt);
+sending C-c then collapses the window.
+
+Cause: the pane's shell was wedged (likely a stopped/zombie job state) after
+the original `kilo remote` relay process died; the tmux window only looked
+alive in `tmux ls`.
+
+Fix: do not keep probing the wedged pane. Let the window die (or kill it),
+recreate it with `tmux new-window -t <cli-session> -n relay`, probe with
+`echo OK`, then restart the relay with
+`apps/mobile/e2e/remote-cli.sh exec remote` (absolute path) and confirm
+`Remote connection enabled.`. The mobile Agents list only discovers the
+session while this relay is connected.

.kilo_workflow/learnings/mobile-ios-cold-build-shareextension-pods-root.md

@@ -0,0 +1,19 @@
+# mobile-ios-cold-build-shareextension-pods-root
+
+Symptom: a genuinely cold `pnpm dev:mobile:ios build <udid>` (cache-busted) fails at
+link with `(3 failures)` and one root error:
+`error: unable to spawn process '/../../../../node_modules/.pnpm/react-native@0.86.0[_patch_hash=...]/node_modules/react-native/scripts/xcode/ccache-clang.sh' (No such file or directory) (in target 'ShareExtension' from project 'Kilo')`.
+
+Cause: the generated `apps/mobile/ios/Kilo.xcodeproj` sets project-level
+`REACT_NATIVE_PATH = "${PODS_ROOT}/../../../../node_modules/...react-native"` and
+`CC/CXX/LD/LDPLUSPLUS = "$(REACT_NATIVE_PATH)/scripts/xcode/ccache-clang[++] .sh"`.
+`PODS_ROOT` is defined only by Pods target xcconfigs, so for the ShareExtension
+target (no Pods xcconfig) `PODS_ROOT` expands empty and the spawn path resolves to
+`/node_modules/...`. Warm-cache iOS runs never hit this; the template is
+tree-independent (present with or without the Hermes patch).
+
+Workaround used in the round (temporary, in the gitignored generated project):
+replace the project-level Debug `REACT_NATIVE_PATH` line with the absolute correct
+path (delete duplicates — a later duplicate key wins). Build then succeeds.
+Proper fix candidate for the repo: have the ShareExtension config plugin set an
+absolute REACT_NATIVE_PATH (or stop overriding CC/LD) for that target.

.kilo_workflow/learnings/mobile-ios-maestro-session-row-and-back.md

@@ -0,0 +1,18 @@
+# mobile-ios-maestro-session-row-and-back
+
+Symptom: on iOS, a Maestro flow that matches the Agents session row and navigates
+back fails two ways: (1) `visible: '<session title>'` never matches even though the
+row renders; (2) the `back` command does not leave a pushed session-detail screen.
+
+Cause: iOS exposes the session row as ONE accessibility element whose label is the
+comma-joined summary (`Hermes-mem-c716 baseline session setup, CLOUD, cost 5 cents,
+6 hours ago, from CLI`), not the bare title; and Maestro's `back` is a no-op on this
+pushed screen — the app renders an explicit `Go back` accessibility element instead.
+
+Fix: match with a prefix regex (`'<title>,.*'`) and tap `Go back` to return. Also
+make reruns self-normalizing: start with an optional `Go back` tap and a
+Home-tab gate before driving tabs.
+
+Bonus zsh trap: `"$CLI:relay"` expands as `$CLI` with zsh's `:r` modifier plus
+literal `elay` (target becomes `<session>elay`, tmux "can't find pane"). Always
+brace: `"${CLI}:relay"`.

.kilo_workflow/learnings/mobile-metro-file-map-stale-after-node-modules-relayout.md

@@ -0,0 +1,20 @@
+# mobile-metro-file-map-stale-after-node-modules-relayout
+
+Symptom: after a dependency-layout change (e.g. a root patch changes pnpm peer-hash
+dirs), the dev client deterministically white-screens and logcat shows
+`ReactNativeJS: [runtime not ready]: RangeError: Maximum call stack size exceeded`,
+stack `metroRequire` → repeated `get NativeModules@<line>:56`. Metro serves a bundle
+in which a proxy module (here `react-native-css/dist/commonjs/components/index.cjs`)
+has `require("react-native")` resolved back to itself.
+
+Cause: Metro's jest-haste-map cache in `$TMPDIR/metro-file-map-expo-*` persists per
+project root across `pnpm install`s. Stale entries pointing into old `.pnpm` peer-hash
+directories still validate while those dirs exist on disk (pnpm keeps them), so the
+resolver serves modules from the stale instance and NativeWind's
+"am I inside react-native-css?" exclusion misfires.
+
+Fix: identify the project file-map (`grep -ac '<worktree-name>'` the
+`$TMPDIR/metro-file-map-expo-*` files — only one matches), delete it, restart Metro
+(`pnpm dev:restart mobile`), re-fetch the bundle with a cache-bust query param, and
+verify the proxy's dep map points at the real `react-native/index.js` module id.
+Never diagnose "the patch broke the app" from a bundle served before this check.

.kilo_workflow/learnings/mobile-native-build-cache-ignores-root-patches.md

@@ -0,0 +1,19 @@
+# mobile-native-build-cache-ignores-root-patches
+
+Symptom: after applying a root-level `patches/` or `pnpm-workspace.yaml` change
+(e.g. a `react-native` Hermes pin), `pnpm dev:mobile:android build` /
+`pnpm dev:mobile:ios build` still installs a binary that behaves as if the
+patch were absent. Easy to conclude "the patch had no effect."
+
+Cause: the Android and iOS native build wrappers key on an `@expo/fingerprint`
+hash computed over `apps/mobile` (plus toolchain) only. Root `patches/`,
+`pnpm-workspace.yaml`, and `pnpm-lock.yaml` are outside that root, so the
+fingerprint is unchanged and a warm
+`~/Library/Caches/Kilo/mobile-android-builds` /
+`…/mobile-ios-builds` entry is reused.
+
+Fix: force a rebuild by deleting `entries/<key>` for the `key` printed by
+`pnpm dev:mobile:android fingerprint` (or `pnpm dev:mobile:ios fingerprint`).
+Parse with `pnpm --silent … fingerprint` and slice JSON from the first `{`.
+Never conclude a root patch was a no-op from a cached binary. EAS and CI build
+from a clean checkout and are unaffected.

.kilo_workflow/learnings/pnpm-install-noop-leaves-virtual-store-drift.md

@@ -0,0 +1,19 @@
+# pnpm-install-noop-leaves-virtual-store-drift
+
+Symptom: `pnpm install` prints "Already up to date" in <1s and does NOT re-apply
+patches or recreate deleted `.pnpm` package dirs — even `pnpm install --force` and
+removing `node_modules/.modules.yaml` no-op. Hand-edited files under
+`node_modules/.pnpm/<pkg>/` stay edited; deleted package dirs stay deleted (symlinks
+dangle), and stale duplicate peer-hash instances (react-native-css, nativewind,
+react-native) accumulate and poison Metro bundling (see
+mobile-metro-file-map-stale-after-node-modules-relayout.md).
+
+Cause: pnpm v11 short-circuits headless install when
+`node_modules/.pnpm-workspace-state-v1.json` matches the lockfile; it never
+re-verifies virtual-store contents. Patch re-application only happens during a real
+(re)link pass.
+
+Fix: `rm -f node_modules/.pnpm-workspace-state-v1.json && pnpm install` forces the
+full relink (~1 min, re-applies patches, prunes stale peer-hash instances). Verify
+with `xxd`/`cat` on the patched files afterwards. The install does not disturb
+running dev services; restart Metro afterwards as cheap insurance.