chore(integration): combine browse work with iOS crash and CI fixes

.github/workflows/deploy.yml

@@ -54,6 +54,10 @@ jobs:
       # needs Xcode 26+. The macos-15 runner defaults to Xcode 16.4 (Swift 6.1),
       # so the archive fails at SPM resolution. Point the whole toolchain at the
       # Xcode 26.3 that's preinstalled on the image.
+      # Xcode 26.3's Swift 6 compiler also hits an abs() overload-ambiguity bug in
+      # expo-modules-jsi (patched via patches/expo-modules-jsi+*.patch — see Expo
+      # issue #47875). Expo's real fix is Xcode 26.4, not preinstalled on macos-15
+      # yet; bump this to 26.4 and drop the patch once the runner image has it.
       DEVELOPER_DIR: /Applications/Xcode_26.3.app/Contents/Developer
     steps:
       - name: Warn when building a non-main ref

.github/workflows/test.yml

@@ -147,6 +147,47 @@ jobs:
         with:
           run: npm run lint
 
+  # ── i18n ─────────────────────────────────────────────────────────────────────
+  # The locale gates (`i18n-completeness`, `i18n-unused-keys`, `i18n`) live at the
+  # root of `__tests__/`, so neither `test:unit` (`__tests__/unit`) nor
+  # `test:integration` (`__tests__/integration`) matches them and they never ran
+  # on a PR. An English-only key therefore shipped green: Type check only catches
+  # a *missing key* (t() is typed against the locale JSON), never a *missing
+  # translation*, which reaches users as silent English fallback mid-RTL layout.
+  i18n:
+    name: i18n
+    runs-on: ubuntu-latest
+    needs: [gate, setup]
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: i18n
+        if: needs.gate.outputs.skip != 'true'
+        uses: ./.github/actions/run-ci
+        with:
+          run: npm run test:i18n
+
+  # ── Native dependency coherence ──────────────────────────────────────────────
+  # Two versions of one native module resolve fine for npm and compile fine for
+  # Xcode, then abort at launch in dyld with "Symbol not found" — a pod built
+  # against one version of a shared dependency against a different version of it
+  # bundled. TestFlight build 173 shipped that way (expo-camera 57.0.1 against
+  # expo-modules-core 57.0.7) and crashed on every launch with a green CI.
+  # `expo-doctor` catches the duplicate at the npm level in seconds, on Linux;
+  # `scripts/verify-dyld-symbols.sh` is the belt-and-braces check at ship time.
+  native-deps:
+    name: Native deps
+    runs-on: ubuntu-latest
+    needs: [gate, setup]
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Duplicate native modules
+        if: needs.gate.outputs.skip != 'true'
+        uses: ./.github/actions/run-ci
+        with:
+          run: npm run check:native-deps
+
   # ── Maestro E2E (iOS sim, monthly schedule only) ─────────────────────────────
   # Runs once per month (00:00 UTC on the 1st) via the schedule trigger above.
   # The macOS runner + iOS app build is ~15 minutes and ~$3/run, so this acts

AGENTS.md

@@ -67,6 +67,60 @@ The `[skip-ci]` suffix keeps CI from re-triggering on the bump-only commit. Neve
 - Before running any EAS build or submit command, stop and ask the user to confirm — do not proceed automatically
 - Never trigger it as a side-effect of a "ship" or "commit and ship" request
 
+---
+
+## PR Titles & Branches
+
+Derived from the dominant pattern in PRs created 2026-07-19 → 2026-07-24
+(e.g. `#348`–`#394`, `#368`, `#376`, `#378`–`#393`). Prefer this over ad-hoc
+formats. Never mention Cursor, Codex, Claude, or other AI tooling in the title,
+branch name, body, or commit message.
+
+### Title format
+
+```
+type(scope): imperative summary
+```
+
+- **type** — one of: `feat`, `fix`, `chore`, `docs`, `test`, `ci`, `perf`, `refactor`
+- **scope** — optional but preferred when clear (`session`, `terminal`, `conversation`,
+  `onboarding`, `servers`, `ios`, `ci`, `deps`, `e2e`, `i18n`, …)
+- **summary** — imperative, lowercase start, no trailing period; describe the
+  user-visible or operational change, not the implementation dump
+
+Examples from recent history:
+
+- `feat(conversation): add in-chat search entry on detail screen`
+- `fix(terminal): fall back to HTTP output when WS replay is blank`
+- `chore(ios): bump build number to 171 [skip-ci]`
+- `docs(session): correct hold_session comments to describe grace-timer behavior`
+- `test(e2e): wire server drag-reorder into the mock Maestro suite`
+- `ci: gate locale parity and dead keys with an i18n job` (scope may be omitted)
+
+Version-bump PRs keep the `[skip-ci]` suffix. Dependabot-style titles stay as
+`chore(deps): …`.
+
+Do **not** use bracket-slug titles like `[feat][task-name]` for new PRs — that
+was a short-lived experiment and does not match the repo’s established style.
+
+### Branch format
+
+```
+type/kebab-case-summary
+```
+
+Match the title’s type. Examples: `feat/in-chat-search`,
+`fix/terminal-empty-replay-fallback`, `chore/bump-ios-version-171`,
+`docs/hold-session-comments`. Never prefix with tool names (`cursor/…`, `cc/…`).
+
+### Base branch (this integration wave)
+
+For work targeting the current integration line, open PRs against
+`integration-merge-354-355-376` (not `main`). Rebase onto the latest tip of that
+branch before merge. Squash title must still follow `type(scope): summary`.
+
+---
+
 ## Merging PRs — Rebase + Squash, Linear History
 
 Keep `main` a straight line — one commit per PR, no merge commits. Every PR follows the same two operations, in this order:

CLAUDE.md

@@ -19,6 +19,20 @@ When adding new Maestro flows:
 
 ---
 
+## Jest — Confirm Suite Failures in Isolation
+
+The `SessionScreen.*` suites are heavy enough that jest's parallel workers oversubscribe some machines, so a batch run can report failures that aren't real. Verify them serially:
+
+```bash
+npx jest --ci --runInBand --testPathPattern "SessionScreen"
+```
+
+**Never classify a batch failure without re-running that suite alone** — a load artifact and a genuine defect are indistinguishable in batch output. Passes alone → artifact. Fails alone → real, fix it. Both mistakes are easy: during the 2026-07-22 integration snapshot four genuinely broken suites were nearly dismissed as flakes, while two others really were artifacts.
+
+The load-sensitive suites, the `.claude/` worktree gotcha (`npx jest` finds **0 tests** there), and their fixes are documented in [`docs/troubleshooting.md`](./docs/troubleshooting.md) → "Jest test suites".
+
+---
+
 ## Comments — Non-Trivial Only
 
 Never add comments that restate what the code already says. Only comment when the code is complex, non-obvious, or would surprise a reader without context.
@@ -173,6 +187,60 @@ To validate a build from one of these tags before merging, run the `Deploy` work
 
 ---
 
+---
+
+## PR Titles & Branches
+
+Derived from the dominant pattern in PRs created 2026-07-19 → 2026-07-24
+(e.g. `#348`–`#394`, `#368`, `#376`, `#378`–`#393`). Prefer this over ad-hoc
+formats. Never mention Cursor, Codex, Claude, or other AI tooling in the title,
+branch name, body, or commit message.
+
+### Title format
+
+```
+type(scope): imperative summary
+```
+
+- **type** — one of: `feat`, `fix`, `chore`, `docs`, `test`, `ci`, `perf`, `refactor`
+- **scope** — optional but preferred when clear (`session`, `terminal`, `conversation`,
+  `onboarding`, `servers`, `ios`, `ci`, `deps`, `e2e`, `i18n`, …)
+- **summary** — imperative, lowercase start, no trailing period; describe the
+  user-visible or operational change, not the implementation dump
+
+Examples from recent history:
+
+- `feat(conversation): add in-chat search entry on detail screen`
+- `fix(terminal): fall back to HTTP output when WS replay is blank`
+- `chore(ios): bump build number to 171 [skip-ci]`
+- `docs(session): correct hold_session comments to describe grace-timer behavior`
+- `test(e2e): wire server drag-reorder into the mock Maestro suite`
+- `ci: gate locale parity and dead keys with an i18n job` (scope may be omitted)
+
+Version-bump PRs keep the `[skip-ci]` suffix. Dependabot-style titles stay as
+`chore(deps): …`.
+
+Do **not** use bracket-slug titles like `[feat][task-name]` for new PRs — that
+was a short-lived experiment and does not match the repo’s established style.
+
+### Branch format
+
+```
+type/kebab-case-summary
+```
+
+Match the title’s type. Examples: `feat/in-chat-search`,
+`fix/terminal-empty-replay-fallback`, `chore/bump-ios-version-171`,
+`docs/hold-session-comments`. Never prefix with tool names (`cursor/…`, `cc/…`).
+
+### Base branch (this integration wave)
+
+For work targeting the current integration line, open PRs against
+`integration-merge-354-355-376` (not `main`). Rebase onto the latest tip of that
+branch before merge. Squash title must still follow `type(scope): summary`.
+
+---
+
 ## Merging PRs — Rebase + Squash, Linear History
 
 Keep `main` a straight line — one commit per PR, no merge commits. Every PR follows the same two operations, in this order:

KICKOFF-landing-runbook.md

@@ -0,0 +1,71 @@
+# Kick-off — build the landing runbook for the mobile PR chain
+
+**Run this from the `merge-prs-v2` worktree** (`integration-dev/v1.0.0-2026-07-22`) so every path below is relative. Paste the section below into a fresh session.
+
+---
+
+Create a landing runbook for this repo's open PR chain — the procedure for merging the 20 open PRs onto `main` one at a time, without re-hitting the documented conflicts or the traps that produce a green signal.
+
+**Format to follow:** `docs/runbooks/_template.md`. Read it first; it opens with the seven rules its sections encode. `docs/runbooks/_example-streamer-land-open-prs.md` is the same template filled in for another repo — use it to see the shape of a finished runbook, but its PR numbers and paths are not ours. `docs/runbooks/README.md` explains why runbooks and merge reports are separate files.
+
+**Source material:** `docs/integration-merge-report-2026-07-22.md`. It is a **run log**: Run 2 is current, **Run 1 is superseded** — do not carry Run 1's resolutions forward. Link to the report by conflict letter (A–I); do not restate it.
+
+**Write to** `docs/runbooks/2026-07-22-land-open-prs.md`.
+
+## Decide these before writing the order
+
+**`#355` — close as superseded by `#354` (investigated, not an open question).** Both PRs are the mobile counterpart to streamer #253 and deliver the **same** live-session capability — the core feature files (`useConversationStream.ts`, `mergeLiveMessages.ts`, `lib/externalSession.ts`, `app/conversation/[id].tsx`) are byte-identical between the two tips. #355 is the *integration-testing twin*: authored on the superseded `bfc800d-2026-07-20` snapshot, so it carries pre-formed cross-PR glue — the #341∪#354 union in `services/api-client.ts` (conflict A) and #339/#341's `cache_alert` frames in `services/ws-client.ts` — that #354 lacks. That glue is only valid on a base that already has #341, and re-forms as conflict A when #341 and #354 both land on `main`; it is **not** live-session functionality. #354 is the one to land: it targets `main` and additionally carries the ar/he/ru take-over translations #355 is missing (`takeOver` keys in `locales/ar/conversation.json`: #354 has 2, #355 has 0). Closing #355 loses no capability — verified by an end-to-end contract trace: every WS event (`conversation_event`, `conversation_events`, `conversation_updated`), type (`ownership`, `processLiveness`) and error code (`CONVERSATION_BUSY`) that streamer #253 emits is consumed by #354. Re-verify before acting: if #355 has been rebased since, re-run the comparison.
+
+**`#291`** (typescript 6→7) is excluded by request.
+
+**Count:** 21 PRs are open and **20** are merged into the snapshot. The report's merge-log table lists only **19** — `#372` (`docs/jest-suite-verification`) is merged and named in the prose but missing from the table. Confirm against `git log --merges origin/main..HEAD` and make the runbook authoritative.
+
+## Stacked pairs
+
+`#341` is based on `#339`, not `main` — identify this from `baseRefName`, never from PR number. The runbook needs both `--onto` rebases and the distinction between them: the one required whenever the parent is force-pushed or gains a commit, and the one required after the parent squash-merges (because the squash produces a commit git cannot match against the originals). Note that a stacked PR may not get full CI until its base is `main`.
+
+## Treat every recorded resolution as perishable
+
+The report demonstrates this itself: `#341`'s run-1 conflict across six files **disappeared** once its branch was rebased onto `#339`. Re-verify preconditions rather than applying resolutions blind — conflict A's resolution in particular depends on a byte-identical precondition the report states explicitly.
+
+The same applies to the fixes pushed upstream during the run — `#346 b84f18c`, `#354 25c83b6` / `5b26bf7`, `#362 ec5260f` / `4c6a275`, `#343 393b9ca`. The runbook's correctness depends on them still being on those branches. Verify, and say in the runbook how to re-check.
+
+## The centrepiece: traps that produce a green signal
+
+**`#356` × `#360`** — a locale-key deletion git applies with **no conflict marker**. The report proves with `git merge-tree` in both directions that reordering does not help, and that the merge hands you a conflict *in the very file silently losing keys*, so it looks handled. What works: rebase, then merge only on a **fresh post-rebase green**, because Type check catches it (`lib/i18n.types.ts` makes a missing key a `TS2345`). This needs its own "Known traps" subsection with the exact post-merge check.
+
+**`#354` × `#346`** — `#354`'s suite omits `useNavigation`; the screen only calls it once `#346` is present. Neither PR is wrong alone.
+
+## Mechanics that must be stated
+
+Worktree **outside `.claude/`** or jest silently finds 0 tests. Each worktree needs its own `npm ci` (~3 min). Verify heavy `SessionScreen` suites with `--runInBand`, and **re-run any single failure in isolation before calling it a flake** — this run had four genuine failures nearly dismissed as flakes and two real load artifacts, and batch output cannot distinguish them.
+
+One post-merge action that is not a merge: **after `#368` lands, add `i18n` to the required status checks**, or the new job runs without being able to block anything.
+
+## Also derive
+
+- **Who should run it** — model *and* reasoning effort per phase, justified by what failure looks like rather than command difficulty. Raise effort where the failure mode is *absence*: the locale cluster (`#356`, `#357`, `#360`, `#362`) is where a clean merge silently deletes keys.
+- **Stop points** — three lists: always-stop (irreversible/outward-facing), stop-because-this-isn't-the-described-situation, and explicitly **do-not-stop**. The third is what keeps the first two credible.
+- **Pre-flight** — sweep all 21 PRs for `CONFLICTING` and red checks, and record what it catches. `gh pr view` must run **twice**: GitHub computes mergeability lazily and the first call only triggers it.
+
+## Done when
+
+Someone who has not read the merge report can land the chain from the runbook alone, and every trap that produces a green signal has an explicit "after merging X, check Y" step.
+
+---
+
+## Follow-up PRs opened after this kick-off (session-name display)
+
+These are **not** part of the 20-PR chain this kick-off builds a runbook for — they are new work opened after, targeting `main` directly.
+When you write `docs/runbooks/2026-07-22-land-open-prs.md`, carry this cross-repo chain into it as a separate "Follow-up PRs" section (do not fold them into the main order — different base, different set).
+
+Fix: interactive Claude Code conversations carried no session name (the scanner only read the `slug` field, which the human REPL never writes), so mobile showed the project name in the list, the conversation view, and the live-session view instead of a real title.
+
+| Order | Repo | PR | Branch | What it does |
+|---|---|---|---|---|
+| 1 | tb-scanner | [#53](https://github.com/RonenMars/threadbase-scanner/pull/53) | `fix/session-name-from-first-message` | Derive session name from the first user message when no `slug`. The data source. |
+| 2 | tb-streamer | [#267](https://github.com/RonenMars/threadbase-streamer/pull/267) | `fix/emit-session-name` | Emit `session_name` in the conversation detail `meta` block. |
+| 3 | tb-mobile | [#376](https://github.com/RonenMars/threadbase-mobile/pull/376) | `fix/session-name-display` | Read `session_name` in the list, conversation, and live-session views (user rename → session name → project name). |
+
+**The mobile PR (#376) is safe to land alone** — it is additive and shows nothing new until the server pipeline (scanner #53 → scanner release → streamer dep bump → streamer #267) lands.
+So the *visible* result is gated on the server side, but the merge order for correctness is only: land #376 whenever; the name simply stays blank until the upstream chain completes.

__tests__/e2e/onboarding-flow.test.tsx

@@ -190,9 +190,9 @@ describe('Onboarding – error handling', () => {
 // ── First-launch flow ─────────────────────────────────────────────────────────
 
 describe('Onboarding – first-launch flow', () => {
-  it('TOTAL_STEPS is 3', () => {
+  it('TOTAL_STEPS is 4', () => {
     // eslint-disable-next-line @typescript-eslint/no-var-requires
     const { TOTAL_STEPS } = require('@/components/onboarding/OnboardingNavigator')
-    expect(TOTAL_STEPS).toBe(3)
+    expect(TOTAL_STEPS).toBe(4)
   })
 })