Closer/ClaudeQAPlan.md

Claude QA Playbook — Full-App QA → Fix → Re-QA until flawless

Reusable QA plan for the Closer app. Run report-only first, fix everything, then re-QA until a clean round. Progress/state is tracked in ClaudeReport.md (issues) + ClaudeQACoverage.md (coverage matrix), which are the authoritative source of truth. See the Continuity section before resuming.

Program roadmap: Part 1 = Android QA (this doc) → Part 2 = build the iOS app to Android's current parity → Part 3 = run these same passes on iOS + a cross-platform (Android↔iOS) pass. Parts 2 & 3 live in ClaudeiOSPlan.md (note: iOS build/run/QA requires macOS — not possible from this Linux box).

⛔ This is a LIVING document — improve it whenever you see a gap (do this automatically)

This playbook, the coverage matrix, and the scripts//qa/ scanners are yours to evolve every round — that is part of the job, not a separate task. Whenever a round teaches you something the plan doesn't yet capture, edit it in the same chunk (no need to ask):

• A bug escaped a prior round, was hard to diagnose, or recurred → add the generalized reflex to the right Pass + the durable substance to the Engineering Manual landmine (the MANDATORY-retrospective rule), and if the class is greppable, add/extend a scanner (scripts/theme-scan.sh, scripts/wiring-scan.sh, qa/entrypoint_smoke.sh).
• A step is wrong, contradictory, or stale (e.g. it told you to do something a standing rule forbids) → fix the wording so the next agent isn't misled.
• A new route / feature / notification / collection / gate / asset appeared → fold it into the relevant Pass + ClaudeQACoverage.md (Living discovery ritual).
• The plan is unclear or bloated → tighten it; lead with the answer; keep one canonical home per fact (don't restate a lesson in four places — link by ID). Leave the plan better than you found it each round. When you change a scanner, update its header; when you change a process rule, make sure it doesn't contradict the Guardrails.

✅ Per-round execution checklist (the literal flow — details in the sections below)

1. Resume: read ClaudeReport.md run-state + ClaudeQACoverage.md (the authoritative state); adb devices shows both emulators; installed build == HEAD (rebuild+install if unsure — never QA a stale APK); baseline clean (both free, 0 active sessions, logcat 0 FATAL).
2. Discovery ritual: reconcile routes/notifications/features/assets/backend with coverage; fold new surfaces in.
3. Run the cheap gates FIRST (before live driving): (a) the automated test suites — ./gradlew testDebugUnitTest
   • cd functions && npm test (they cover the fragile logic: encryption format, rate limiter, quiet hours, streak, entitlement math) — a red suite is a P0/P1 regression gate, stop and fix before QA'ing a build; (b) the scanners — qa/entrypoint_smoke.sh (both serials), scripts/theme-scan.sh (Pass C), scripts/wiring-scan.sh (Pass N), scripts/painter-xml-scan.sh (crash guard — painterResource() on a non-<vector> XML drawable throws on render; caught O-ONBOARD-001 class — exit≠0 is a P0 gate); (c) optional monkey fuzz adb shell monkey -p app.closer --throttle 300 --pct-touch 90 -v 5000 (any crash = bug). File 🔴/🟠 to ClaudeReport.md; record counts in coverage.
4. Run the passes report-only, sub-batched to one context window each — recurring set A–N + P (K money-path + O release gates only when a sandbox device / pre-ship is in scope). Checkpoint the MD files after each chunk.
5. Fix phase (after all passes): by severity P0→P1→P2→P3, one at a time, verify each live via the real path + re-run the relevant scanner, flip the row to Fixed, capture the durable substance in the Engineering Manual.
6. Re-QA loop until flawless (see Definition of Done). Prune confirmed-fixed rows.
7. You never git commit/push — the user commits. Your durable state is the MD files (they survive compaction).

📖 Architecture reference (read BEFORE testing the matching area)

For each Pass below, before you start, read the relevant section of docs/Engineering_Reference_Manual.md — it documents the architecture, the wire-format contracts, the security invariants, and the Known landmines (bugs that cost real debugging time and are easy to re-introduce).

This is bidirectional — the manual is a LIVING document, not a read-only reference. Read it before; write back to it after. Whenever a round fixes a bug, changes a contract/flow/gate, or finds the manual stale or missing something, update the manual in the same chunk (see Where every finding goes, the Docs update rule, and the MANDATORY retrospective — all now route durable engineering truth here). Treat it as part of every fix, same as ClaudeReport.md/ClaudeQACoverage.md.

Pass	Manual section to read first
A — Couple-shared premium	Premium-gated features and gate pattern · Billing
B — Games lifecycle	Game session push semantics (idempotent flag-claim) · Foreground game-alert banner · F-RACE-001
C — Visual (light+dark)	Daily question lifecycle · C-NAV-001 · Back-stack gotchas · C-HOME-001
D — Security & encryption	End-to-end encryption model · Firestore security rules · Encryption versions
E — Notifications	Notifications · Notification deep-link routing · E-GAME-001 · E-GAME-002
F — Resilience	End-to-end encryption model · Known limitation: single-device keys
G — Account creation / fake-account	Authentication and pairing flow · Rate limiting on accept
H — Branding & artwork	ClaudeBrandingReview.md (this repo) · docs/brand/visual-identity.md
I — Performance	Engineering conventions · Where to look first
J — Accessibility	CloserTheme · Engineering conventions
K — Billing & subscription lifecycle	Billing · Premium-gated features and gate pattern
L — Messaging & chat (E2E)	End-to-end encryption model · Notifications
M — Settings & account management	Authentication and pairing flow · Notifications
N — Daily question & interactive features	Daily question lifecycle
O — Release build & store readiness	Firestore security rules · Engineering conventions
P — Content, copy & language	docs/brand/visual-identity.md (Store voice) · seed/questions/QUESTION_CONTENT_GUIDE.md (v3)

If you find a bug that LOOKS like it might be a re-introduction of a known landmine (above table or Known landmines), stop and verify the fix is still in place before filing a new ID — it may be a regression on a known issue, not a new bug.

Where every finding goes (route it here — exactly one home each)

What you found	Where it goes	Form
A bug — broken / incorrect / crashing / insecure, premium bypass, wrong-or-missing notification, dead-end nav	ClaudeReport.md	Table row: stable ID (A-001, E-003…) + severity (P0–P3) + repro + status
An idea / improvement — works but could be better, confusing copy, missing affordance, rough-but-not-broken flow, "it'd be great if…", feature idea	Future.md ## QA	Short title + what prompted it + suggested improvement
New artwork to create — illustrations, glyphs, image-gen prompts	ClaudeBrandingReview.md	House-style prompt + placement
What got tested + its status (pass / fail / todo / deferred)	ClaudeQACoverage.md	Coverage cell (the resume anchor)
Automated scanner findings	ClaudeReport.md (CRITICAL/MAJOR that break themes/functionality) + ClaudeQACoverage.md (execution counts + filing status)	ID + file:line + pattern + fix suggestion
Durable engineering knowledge — a fixed bug's root cause + how it's easy to re-introduce, a new architecture fact / data path / wire-format contract / security invariant / gate pattern, or anything the manual is now stale/missing about	docs/Engineering_Reference_Manual.md (esp. Known landmines and recent fixes)	New landmine entry (ID + cause + the guard) and/or an updated architecture/gate/flow section

• A branding defect (mis-colored, clipped, off-brand, low-contrast art) is a bug → ClaudeReport.md, not a brand idea — only new art to create goes to ClaudeBrandingReview.md.
• WRONG LANGUAGE IS A BUG (not a Future.md idea). A typo, grammar/punctuation error, off-brand or cold/salesy voice, non-inclusive/assumptive wording, leaked placeholder/dev/raw-error text, copy that doesn't match behavior, or a broken/duplicate/off-guide question → ClaudeReport.md (see Pass P). Only genuinely-working copy that could be warmer/clearer (a rewording for delight) goes to Future.md. "Confusing copy" that actually misleads the user is a bug.
• ONE canonical home per fact; everywhere else is a pointer (ID/anchor), never a paraphrase. This is the rule that keeps the five docs from duplicating each other (and wasting tokens re-stating the same lesson). Route by purpose: the defect (repro/severity/status) → ClaudeReport.md (transient — prunes to an ID after one confirm); the substance (root cause / why it's fragile / how to not re-introduce it) → the Engineering Reference Manual (permanent, engineer-facing); the reflex (how to FIND the class next round) → this ClaudeQAPlan.md Pass (generalized, citing the ID); coverage status → ClaudeQACoverage.md; cross-session ops not in the repo (accounts, tooling, auth) → memory/. State a fact in its home once; elsewhere cite the ID. Don't restate a fix in four docs.
• The Engineering Reference Manual is a LIVING document — read it before a pass, write back to it after. When a round teaches the codebase something durable (a fixed bug's re-introduction risk, a new/changed architecture fact, data path, contract, gate, flow, collection/Function/route, or the manual disagreeing with reality), update the manual in the same chunk. A fix is not complete until its durable substance is in the manual (see the MANDATORY-retrospective rule). The report row and the Pass reflex just reference the manual's landmine ID — they don't re-tell it.
• Logging an idea in Future.md is never a substitute for filing a real defect: if it's broken, it gets an ID in ClaudeReport.md too.
• Bug lifecycle: filed in ClaudeReport.md → fixed → kept one confirmation round → pruned to the archived-ID line (detail lives in git). Future.md ideas sit in the backlog until built. (See Report hygiene under Reporting.)

Context

Drive the real app on both emulators, verify each thing live, report, fix, re-verify. Core QA dimensions (cornerstones):

1. Couple-shared premium — if EITHER partner is premium, all premium features unlock for both.
2. Games — each starts, plays, joins, resumes, finishes, and reopens results correctly on both devices.
3. Full visual pass, light + dark — every screen, text readable, nothing clipped/invisible.
4. Security & encryption (cornerstone) — every private field is ciphertext at rest, rules hold against non-members, keys/recovery are sound. Findings here default to P0.
5. Notifications — the full suite: every type delivers to the right partner (foreground/background/killed), deep-links correctly, opens the right destination on both clients, covers all game/join-game flows, handles stale notifications, and leaks no private content.

These five are the original cornerstones; the playbook has since grown to cover the rest of the app as first-class passes (see QA passes below): K Billing & subscription lifecycle (the real purchase/restore/cancel/expiry money path, not just the admin entitlement toggle), L Messaging & chat (E2E send/receive/react/media, both clients), M Settings & account management (every toggle persists + takes effect; biometric lock, quiet hours, unpair/delete), N Daily-question/reveal/check-ins + Bucket List/Date Builder/Activity (the interactive non-game features), and O Release build & store readiness (the minified release build, signing/AAB, App Check, i18n, deep/App-Links, Play Data-Safety — everything else runs on the debug APK), and P Content, copy & language (typos/grammar, brand voice, inclusive language, and the question-bank content — wrong language is a bug, not a Future.md idea). Plus the existing F resilience, G account/abuse, H branding, I performance, J accessibility. Pass letters are stable IDs — never renumber (issue IDs and coverage rows reference them; note D/E/G are not in strict alphabetical position for that reason).

Scope decisions: exhaustive visual pass (all ~50 screens, both modes); full scope incl. pre-pairing flows (fresh throwaway account); couple-shared everywhere — per-user gates are bugs, fixed by routing through core/billing/CouplePremiumChecker.kt; full notification suite — every type, game + join-game pushes, deep-links, stale-notification handling, and all in-app paths into joining/resuming/results, verified on both clients.

Early known signal: only chat uses CouplePremiumChecker; games/packs/dates/wheel gate on the user's own EntitlementChecker.isPremium() — so premium almost certainly does NOT unlock for the free partner there. Pass A confirms + enumerates this; the fix phase applies couple-shared everywhere.

Execution mode — run to completion (autonomous; do NOT stop)

• Do not stop to check in or ask for approval. Run all passes (A–P — recurring set A–N + P each round; K's real-money path and O's release/store gates run when a sandbox device / pre-ship is in scope) → the fix phase → re-QA rounds continuously until a flawless round (zero open P0–P2, Passes D + E clean, every game fully played through, all notification routes verified, navigation/back-stack verified). Don't hand control back early.
• Unblock yourself: if anything blocks progress (a stale/blocking session, a crash, a build break, a missing prerequisite state, a broken nav path that prevents reaching a screen), fix it immediately and continue — even though passes are otherwise report-only. Blocking issues are fixed inline so the run can proceed; non-blocking findings are still logged and fixed in the fix phase.
• "Once executed, complete it": never declare done before the Definition of Done is met — keep cycling fix → re-QA until flawless, then stop.
• Context limits ≠ stopping — do NOT hand back to the user when context fills. The harness auto-summarizes a long conversation and continues in the next window; you continue without the user. (You cannot self-invoke /compact — and you don't need to; auto-compaction handles it.) The committed ClaudeReport.md run-state + ClaudeQACoverage.md are the authoritative state and survive any compaction — after a summary, re-read them and continue at the next chunk. Never pause a run merely because context is getting long; only stop for a true blocker (a denied gated action even with standing auth, or the macOS requirement for iOS).
• Checkpoint (save the working-tree MD files + run-state) before anything interruptible so a mid-chunk compaction never loses progress (the user commits — never run git yourself). Keep chunks atomic; if a chunk is cut off mid-way (e.g., a game session left active), the session-start ritual recovers it (clear the stuck session via in-app "End their game", then redo that chunk). Right-sized chunks (see Batch sizing) make this rare.
• Don't pause for "by-design vs bug": log the ambiguous finding and keep going (don't unilaterally rewrite deliberate design — the log captures it). Never halt the run to ask.
• Only true stop = a gated action you cannot perform. Production deploys, admin Firestore writes/seeds, and entitlement toggles still need per-occurrence authorization (the classifier enforces this regardless of this doc). If one is genuinely required to proceed and is denied, do all other work first, then surface only that single blocker — don't halt the whole run for it.

Methodology (every pass)

• EVIDENCE OVER ASSUMPTION — read the logs, never assume, always verify (the #1 rule). Every conclusion — pass, fail, fixed, "it works", "the notification didn't open" — must be backed by observed evidence, never by what the UI appears to do or by reasoning about the code. Concretely:
  • Read logcat on EVERY action, not only when something looks wrong. logcat -c before a tap/flow, then after, scan for FATAL EXCEPTION/ANR/PERMISSION_DENIED/exceptions. Absence of a visible symptom ≠ success — a screen that "looks fine" can be masking a swallowed exception, a denied read, or a crash on another device.
  • Verify with ground truth, not appearance: confirm persisted state via admin reads (Firestore), confirm delivery via notification_queue/dumpsys notification, confirm routing via the landed screen + back stack, confirm encryption via the raw stored bytes. "Looked right" is not verified.
  • Don't theorize a root cause — reproduce it and read the stack. If behavior is "didn't work / closed / flashed", pull the crash log FIRST (this session's bug was misdiagnosed by reasoning until the live stack named the splash NPE).
  • Don't trust a synthetic pass (am start, admin write, direct call) for launch/notification/permission paths — verify through the real channel (see Reproduction fidelity). A green that didn't exercise the user's path is not green.
• Devices: 5554 (QA), 5556 (Sam), paired; one fresh throwaway account for pre-pairing flows.
• Drive via adb tap/swipe; resolve coords from uiautomator dump bounds; downscale screenshots to read; scan logcat for FATAL EXCEPTION/ANR on each screen.
• Premium toggled via scratchpad/set_premium.js (admin, user-authorized each time).
• Theme toggled via Settings → Appearance (Light/Dark) (MainActivity ThemeMode).
• REPORT-ONLY during passes — never fix mid-pass.
• THINK AS A CONSUMER — approach everything from different angles. Beyond "does it work", constantly ask "is this what a real person would expect / want here? is this delightful, confusing, or annoying?" Come at each flow from multiple angles (first-time user, returning user, the partner who didn't start it, someone tapping fast, someone reading carefully, the skeptic, the impatient). Vary inputs, depths, orders, and entry points (don't repeat one happy path). A thing can be bug-free yet still worse than it should be — notice that too.
• CAPTURE IMPROVEMENT / FEATURE IDEAS → Future.md (section ## QA). Bugs (broken/incorrect behavior) go to ClaudeReport.md as always. But anything that works yet could be better — confusing copy, a missing affordance, a rough-but-not-broken flow, a "it'd be great if…" feature idea — append it to Future.md under ## QA with a short title, what prompted it, and the suggested improvement. This is an idea backlog, not the bug log; logging here is never a substitute for filing an actual defect in ClaudeReport.md.
• Environment (senior-QA rec): prefer the Firebase Local Emulator Suite or a dedicated staging project over production — isolates test data, makes seeding / entitlement toggles / D3 negative tests free (no gated prod writes), and avoids polluting real users. Caveat: App Check, RevenueCat IAP, and real FCM/APNs push need real services — run those against staging/prod with test accounts. (We've been on prod with test accounts — works, but every seed/toggle/deploy hits the gate.)
• Device/OS matrix (pre-ship gate — currently NOT met; track it honestly): per-round QA runs on our two identical emulators (5554/5556, same API + screen size) — that's the realistic recurring setup, not full coverage. Before any store push, certify across minSdk + targetSdk, a small and a large screen, and at least one physical device (App Check / Play Integrity behave differently on emulators). Because this is unmet today, keep a blocked→needs-device row for it in ClaudeQACoverage.md (alongside Pass K money-path + Pass O) so the gap stays visible rather than silently assumed-covered — don't claim "device matrix ✓" off two same-size emulators.
• ⛔ First-run / cold-path lane (fixture blind-spot — learned from O-ONBOARD-001, a P0 we shipped past). The two recurring emulators (5554/5556) are paired + signed-in + onboarding-complete — a stable fixture that is great for the A–N passes but structurally cannot reach the entire first-run surface: onboarding (every slide + Skip), sign-up, login, the auth-screen logo, pairing/invite-accept, recovery-on-new-device, and day-1 empty states. A bug anywhere in that region is invisible no matter how thorough the passes are — O-ONBOARD-001 (every fresh install crashed on the last onboarding slide) sat undetected precisely because the fixtures skip it. So: run a fresh-install lane on a THROWAWAY device (e.g. emulator-5558, or a fresh AVD — never pm clear 5554/5556, it breaks the App Check debug token): install the build, walk onboarding all the way through (+ Skip) → sign-up → login → pairing → first daily Q, asserting 0 FATAL and each screen renders. Trigger it every time you touch onboarding / auth / pairing / branding / launcher or any res/drawable asset, and always before a store push. Keep a first-run row in ClaudeQACoverage.md so this blind-spot stays visible instead of assumed-covered.
• Render-level coverage gap (the other half of why O-ONBOARD-001 slipped). Our cheap gates are all static or logic-level — unit/functions tests, theme-scan, wiring-scan, painter-xml-scan — none of them actually render a composable. painterResource on a <bitmap> compiled fine, passed all 205 unit tests, and only threw on first paint. There is a whole class of "composes fine, crashes on render" bugs (resource resolution, LocalContext casts, bad painterResource) with zero automated coverage. Until a screenshot/instrumented render test exists (see Future.md — Roborazzi/Paparazzi JVM screenshot test or a Compose/Espresso smoke), the fresh-install lane above is the only net for render crashes on first-run screens — treat it as mandatory, not optional.
• Automate the regression smoke: capture the smoke checklist as a runnable script (adb/Maestro) so every round re-checks it cheaply instead of by hand. Built: qa/entrypoint_smoke.sh <serial> <recipient_uid> (+ helper qa/qa_push.js) — the cold-start / entry-point launch-integrity smoke. It launches via the launcher AND sends a real push to a killed (am kill) app and taps the actual OS notification for each type, asserting the app opens and STAYS (process alive, 0 FATAL, off the launcher). This is the smoke that catches the "opens-and-closes" splash-crash class that am start can't. Run it every round and after any change touching MainActivity / splash / theme / manifest / nav / notifications. FAIL = an app crash (real bug); BLOCK = push not delivered (flaky emulator FCM — rerun, not a bug).
• Run the project's OWN test suites every round (they are the cheapest, most deterministic regression net). Before the scanners and live driving, run ./gradlew testDebugUnitTest (19 unit tests — FieldEncryptorTest, SealedAnswerEncryptorTest, NotificationRateLimiterTest, QuietHoursManagerTest, StreakCalculatorTest, ChallengeStateMachineTest, PartnerNotificationManagerTest, HomePriorityEngineTest, DateMatchRepositoryImplTest, CloserBrandCopyTest, …) and cd functions && npm test (entitlementLogic.test.ts). A failing test is a regression bug (P0/P1) — file it and do not QA a build with a red suite. A fix that breaks a test isn't "Fixed" (see Fix phase). These guard the exact invariants this QA chases (ciphertext format, rate limiting, quiet-hours suppression, entitlement math), so a green run is a precondition, not a bonus. (Coverage gap: there are 0 instrumented tests — no on-device Compose/Espresso smoke; logged in Future.md. Until that exists, the live passes + scanners are the only UI-behavior net.)
• Stress / monkey fuzz (cheap random-crash net the manual nav-fuzz misses): once per build run adb shell monkey -p app.closer --throttle 300 --pct-touch 90 -v 5000 on each emulator with logcat capturing — any FATAL EXCEPTION/ANR it triggers is a bug (file it with the monkey seed). This complements Pass C's targeted nav fuzzing with broad random input.
• Run associated automated scanners BEFORE the manual pass. Every pass with a supporting script must start with it:
  • Pass C: run scripts/theme-scan.sh and review /tmp/claude-theme-scan-<date>.md before looking at any screen.
  • Pass N (+ discovery ritual): run scripts/wiring-scan.sh and review /tmp/claude-wiring-scan-<date>.md before driving the interactive features — it catches the silent dead-feature class (N-001 Bucket List, N-002 Date Builder): 🔴 a setX() ViewModel setter with no caller, 🟠 a repository read method with no ui/ caller (data written but never displayed), 🟡 if (x.isEmpty()) return bail-guards to confirm the state is actually provided. Every 🔴 is a likely dead feature — prove the feature works by persisting real data and reading it back from Firestore (admin), not by trusting the empty-state render.
  • If a scanner does not yet exist for a pass but the pass is highly automatable (e.g. touch-target sizing for Pass J, enc:v1: leak grep for Pass L, redundant-read count for Pass I), consider building it and adding it here.
  • Scanner findings narrow the manual sweep: every 🔴 CRITICAL must be verified (both themes for C; live persist→read for N); 🟠 MAJOR must be reviewed for theme/art breakage or orphan data; 🟡 REVIEW is checked during the sweep.
  • If a manual finding is something the scanner should have caught, improve the scanner (see Living discovery ritual).
• Test-data hygiene: keep known test accounts; clean up artifacts (stray messages/reactions/sessions) between rounds so they don't masquerade as bugs.
• Evidence standard: every filed bug must be reproducible from text alone: build/commit, device, account, theme, app/process state, screen/route, exact tap/input sequence, expected result, actual result, and whether logcat showed a crash/ANR/permission denial. Screenshots/videos are helpful but never the only evidence because session artifacts may not survive compaction.
• Flake policy: if something fails once and then passes, do not dismiss it. Repeat from a clean state, vary timing (rapid tap / slow network / background-resume), inspect logs, and file it as intermittent if it cannot be made fully deterministic. Intermittent routing, notification, encryption, duplicate-write, or crash behavior is still a bug.
• Reproduction fidelity (how we catch DEEP bugs) — the test harness must exercise the SAME path as the user. A synthetic shortcut (am start extras, admin writes, calling a function directly, am force-stop) can pass while the real path crashes — the splash-handover NPE only fires on a real notification cold-start, and am force-stop can't even receive FCM. So for launch / notification / permission / IPC / deep-link behavior, reproduce through the real OS mechanism (real push tapped from the shade, real launcher cold-start, real permission dialog). Record which angle proved it in ClaudeQACoverage.md; "synthetic/UI-shortcut only" is not a pass for these paths.
• Symptom→inspection reflexes (apply before theorizing a root cause): (1) "opens-and-closes / flashes / silently fails" ⇒ it's a crash until the stack says otherwise — logcat -c then capture FATAL EXCEPTION from the live repro before proposing a cause (don't fix by reasoning, like the routing red-herring on this very bug). (2) Many features break at once ⇒ inspect the SHARED code path (launch/onCreate/splash/auth/key-load), not each feature. (3) "worked before, broken now" ⇒ diff & history-check before you fix: git blame/git log -L/git diff the failing line to the introducing commit (incl. other agents' commits — Codex/kimi/Ripley co-edit this repo), and search the Engineering Manual landmines + the report's archived-ID line for a prior fix of the same symptom — a match means regression, not a new bug (full procedure: the Fix-phase Regression triage step). (4) Treat cosmetic/branding/theme/manifest/splash commits as capable of deep crashes — re-run the cold-start + notification smoke after them.

Living discovery ritual (before each round, and whenever reality disagrees with the docs)

The app is allowed to grow; the QA plan must keep up. Before a pass or chunk, quickly inventory the current code/app surface and reconcile it with ClaudeQACoverage.md:

• Routes/screens: inspect core/navigation/AppRoute.kt, navigation graph call sites, Settings sub-pages, dialogs, bottom tabs, deep links, and any new composables reachable by buttons/cards.
• Notifications: inspect notification type enums/classes, Cloud Function triggers, Android intent/deep-link handling, notification channels/actions, FCM token registration, and Android runtime notification permission paths.
• Features/gates: grep for premium checks, permission requests, media pickers, billing/paywall entry points, destructive actions, account/couple lifecycle actions, and admin/server-only writes.
• Assets/content: inventory new drawables, drawable-night* variants, pack art, empty states, strings, feature flags, remote config, and any debug-only screens that should not ship.
• Backend/rules: inspect Firestore rules, indexes/queries, Functions triggers/callables, Storage paths, scheduled jobs, and migrations for new data shapes or access paths.
• Docs update rule: if the inventory finds a page, feature, notification, asset, state, backend path, or edge case missing from the playbook/coverage, update ClaudeQAPlan.md and ClaudeQACoverage.md before marking the chunk done.
• Scanner update rule: if a manual finding is a pattern an existing scanner should have caught (e.g. a hardcoded surface color the theme scanner missed, a route the smoke should have exercised), improve that script and document the change in its header. If no scanner exists for a repeated failure mode, consider writing one and adding it to Methodology. If it is product polish, also add it to Future.md; if it needs new artwork, add it to ClaudeBrandingReview.md. And if the discovery is a durable engineering fact (new route/collection/Function/flag/contract, a changed wire format, a renamed file, a gate/flow that the manual describes wrongly or omits), update docs/Engineering_Reference_Manual.md in the same chunk — the discovery ritual is exactly when the manual drifts out of date, so reconcile it then, not "later".

Multi-angle attack mandate (go DEEPER than "does the happy path work")

A capability can pass via the UI yet fail when hit directly. Probe each meaningful capability (read/write a private field, gate a premium feature, deliver/route a notification, start/finish a game, pair/unpair, create an account) from as many independent angles as apply — not just the in-app happy path:

• Real UI (play-as-user) — the baseline angle.
• Crafted intent / deep-link — fire the exact intent a notification/link carries (bypasses UI nav) to test routing in isolation; also send malformed/missing extras → must route gracefully or no-op, never crash.
• Raw API against the DEPLOYED backend — hit Firestore/Storage/Functions REST directly with a real token, as a member AND a non-member, to exercise rules + App Check from OUTSIDE the app. A non-member (or no-App-Check) request must be DENIED — App Check 403 or rules PERMISSION_DENIED. The member request characterizes which layer enforces. Any unauthorized 200 returning couple data = P0.
• Admin inspection (ground truth) — read the RAW stored docs/objects (admin bypasses rules) to assert what is actually persisted: ciphertext only, no plaintext, no raw keys/invite-seeds, no private content in pushes.
• Concurrency / race — two partners (or two rapid taps) hit the same thing at once.
• Killed / cold state — kill with am kill <pkg>, NOT am force-stop: a force-stopped app is in Android's stopped state and is excluded from FCM broadcasts (GCM broadcast …result=CANCELLED), so the push never arrives and you get a false "no notification". Then deliver a real push and tap the actual OS notification (one at a time — clear the shade first; tapping a grouped summary launches with no extras and falsely lands on Home). am start … --es type … is not equivalent to a real notification tap (different launch path — see the crash-triage note in Pass E). Also cold-start straight onto a deep link.
• Malformed / abusive input — oversized, empty, rapid-fire, injection-ish, forged FCM payloads, replayed/expired tokens & invite codes.
• Offline / flaky — drop network mid-action → graceful failure, recover on reconnect.

Record which angles were tried per area in ClaudeQACoverage.md. For security- or data-sensitive capabilities, "UI happy path only" is not a pass. D3/Pass G negative access MUST be executed live via the raw-API angle each round — never deferred to "only 2 emulators." (Mint a token for a non-member UID via admin → exchange for an ID token via the Identity Toolkit REST signInWithCustomToken → use it as Bearer against the Firestore REST API.)

Continuity & resumability (this effort WILL span many context windows — don't lose state)

State lives in files, not memory:

• ClaudeReport.md = the issue log (committed). Each issue row is self-contained in text (repro + expected
  • actual) — screenshots are session-only and won't survive a compaction; never rely on a screenshot path alone.
• ClaudeQACoverage.md = the coverage matrix: every screen×mode, feature×premium-state, game×lifecycle, notification×{foreground,background,killed}, each todo | pass | fail→id | not implemented→Future.md | blocked→id. The resume anchor.
• Future.md (## QA) = the non-bug improvement/idea backlog; ClaudeBrandingReview.md = the branding/artwork review + image-prompt backlog. Both committed alongside the report/coverage.
• Persistent memory (memory/): QA methodology + exact commands; emulator↔account↔coupleId mapping; scratchpad/set_premium.js + admin tooling; the couple-shared-premium-everywhere goal + the per-user-gate gap.
• Run-state header pinned at the TOP of ClaudeReport.md, always current: Round N | Pass X | Chunk Y | NEXT ACTION: … — first thing to read, last thing to update before stopping.
• Stable issue IDs: A-001 / B-002 / C-… / D-… / E-… (pass-letter + number); coverage references the ID for every fail. Never renumber or reuse.
• Source of truth: the two MD files are authoritative; the TodoWrite list is scratch for the current chunk only. Update the MD files + run-state header before ending a session.
• Living playbook rule: when QA discovers any new app surface or recurring lesson — a new page/route, feature, setting, game state, notification type/action/channel, entry point, background/killed-state behavior, asset/art placement, repeatable bug class, missed edge case, fragile route, confusing state, image/layout failure mode, security angle, or anything else that should be checked every future round — update this ClaudeQAPlan.md in the relevant pass before ending the chunk. Also add the matching row/cell to ClaudeQACoverage.md if it needs recurring verification. And update docs/Engineering_Reference_Manual.md when the discovery is durable engineering truth (a new architecture fact, data path, contract, gate, flow, or a fixed bug's re-introduction risk) — the QA plan captures what to re-test, the manual captures what the system is and why it's fragile; both are living and both get updated. Do this even after the immediate bug is filed/fixed so the lesson or newly discovered surface is not lost to memory or git history.
• Learn from every ESCAPED or DEEP bug — MANDATORY retrospective (do this automatically, not only when asked). Any bug that (a) escaped a prior round, (b) needed non-obvious diagnosis (a crash, an "opens-and-closes", a "didn't work", an intermittent, a wrong-root-cause first guess), or (c) recurred triggers a short retrospective the moment it's fixed — the fix is not complete until all four are done:
  1. Add the guard that would have caught it — a new qa/ smoke check, a coverage row, or a concrete pass step (e.g. the cold-start bug → qa/entrypoint_smoke.sh). If an existing smoke missed it, extend the smoke.
  2. Capture the lesson in its ONE canonical home, then link by ID elsewhere — never paraphrase it twice. Split by purpose: the reflex (how to find this class next round) goes in the relevant Pass of this doc, written generalized and citing the bug ID as an example (do NOT re-narrate the bug here); the substance (root cause + where it lives now + re-introduction risk + the guard) goes in docs/Engineering_Reference_Manual.md → Known landmines and recent fixes (and update the matching architecture/gate/flow section if the fix changed it). The manual is the next engineer's first read; a landmine that isn't in it will be re-introduced. Do NOT copy the fix into memory/ — per the memory rules, memory holds only cross-session facts NOT in the repo (emulator↔account map, admin tooling/commands, standing auth, never-commit); past fixes belong to the manual, so memory just points to the landmine ID if needed.
  3. Name the missing state/angle/entry-point that let it hide and add it to the multi-angle / state matrices so it's exercised every round (e.g. "real notification tap on an am kill'd app", not just am start).
  4. Note any wrong turn in diagnosis so the misstep isn't repeated (e.g. "synthetic test passed while the real path crashed → don't fix by reasoning; reproduce via the real channel + read the stack"). This is how the plan self-improves between rounds — treat the human pointing out a missed bug as a signal the plan had a gap, and close the gap here, not just the bug.
• Checkpoint cadence: save ClaudeReport.md + ClaudeQACoverage.md + run-state after each pass and each chunk (the user commits — never run git yourself; see Guardrails).
• Chunking: run small chunks (Pass C one screen-group; Pass A one feature), checkpoint after each.
• Session-start ritual: (1) read run-state header + both MD files; (2) adb devices shows both emulators online; (3) installed build == current HEAD (rebuild+reinstall if unsure — never QA a stale APK); (4) continue at the first todo / unverified-fix; (5) if a prior chunk left an active/stuck game session, recover it via in-app "End their game" (log if needed), then redo that chunk.

Batch sizing — sub-batch each pass to ONE context window (Round-1 calibration)

A pass is a category, not a unit of work. Execute each pass as sub-batches (chunks), where a chunk = the largest coherent unit that reliably finishes AND checkpoints within one context window, with margin. End every chunk by saving the MD files + run-state (the user commits — never run git yourself). If a chunk starts overflowing, split it; if chunks feel trivial, merge them. Why: in Round 1, A & D fit as single batches, but B/C/E were too large → got cut off → deferred. Sub-batching prevents half-done/lost work and gives cleaner per-chunk verification + revertable history.

Default small: if a chunk requires two-device live driving, screenshots/montage review, logcat checks, or admin/API verification, keep it to one small route family, one game phase, or one notification type. A chunk is too large if it cannot produce a precise coverage update, issue log, and file-checkpoint before context gets tight. Split before starting rather than leaving a half-tested matrix behind. Prefer Claude-friendly micro-batches: smaller chunks let the agent fully inspect screenshots, tap every CTA, vary app states, update files accurately, and avoid shallow "covered" rows.

Pass	Chunk granularity	~chunks
A Premium	one gated-feature family per chunk if live toggles are needed; otherwise free-state sweep → couple-shared verify	2–4
B Games	one game per chunk max; split complex games into lifecycle/playthrough chunk + join/resume/results/notification-entry chunk	7–14
C Visual	one small route family per chunk (both themes, ~2–3 screens/states, screenshots reviewed + nav/back + image-fit + all CTAs for that family) — never "all screens" or a broad tab at once	16–25
D Security	one security assertion group per chunk: D1 at-rest · D2 rules static · D3 live negative raw API · D4 keys/recovery · D5/D6 leaks · D7 migration	~6
E Notifications	one notification type per chunk with the full contract below; split a type into direction/state subchunks if needed, but do not mark the type pass until both clients + source screens + fg/bg/killed + stale/malformed + payload/back-stack are covered	16–30
F Resilience	one dimension per chunk (concurrency · lifecycle/process-death · network · time · account-lifecycle)	~5
G Account creation	one creation/abuse dimension per chunk (happy/validation · duplicate/conflict · fake-account abuse · lifecycle)	~4
H Branding	one small route family per chunk (~2–3 screens/states) consumer brand walk + ready-to-paste art prompts + existing-image integration verdict	8–14
I Performance	one route-group per chunk — gfxinfo/jank + read-count instrumentation (build the route smoke checklist)	~3
J Accessibility	one a11y setting per chunk (font scale · TalkBack · contrast · targets · keyboard · reduce-motion)	~5
K Billing	one money-path per chunk (purchase · restore · plan-switch · cancel→expiry-relock · refund · webhook auth) — needs a real device/sandbox	~6
L Messaging	one chat dimension per chunk (send-types both dirs · reactions/receipts/typing · failed-send/offline · media perms · inbox/entry-points · delete/moderation)	~6
M Settings	one settings group per chunk (appearance · notif toggles · quiet hours · biometric lock · edit profile · unpair/delete · security/recovery)	~6
N Interactive features	one feature per chunk (daily-question loop · outcomes/check-ins · Bucket List · Date Builder · Activity feed)	~5
O Release/store	one gate per chunk (minified release smoke · signing/AAB · App Check (staging) · deep/App-Links · permissions/manifest · i18n · Data-Safety/store) — pre-ship, not per-round	~6
P Content/language	one surface per chunk (UI microcopy of a route family · voice/tone sweep · inclusive-language sweep · question-bank by category/depth · legal/store copy)	~5

Context-cost tips: prefer code/admin-read audits (cheap) before live UI sweeps; montage screenshots (dark|light pairs) to review many at once; keep one chunk = one TodoWrite focus.

Guardrails & efficiency

• ⛔ NEVER git commit / git push — the USER does ALL commits. This overrides every "commit" verb elsewhere in this doc: wherever a step says "commit," read it as "checkpoint = save the working-tree files (ClaudeReport.md + ClaudeQACoverage.md + run-state, plus any code/docs)" and leave the actual git commit to the user. Your durable state lives in those files (they survive compaction), not in a commit you make. Never stage, commit, push, branch, or amend.
• Never pm clear / wipe app data — breaks the App Check debug token. Pre-pairing QA: sign-out → fresh sign-up.
• Never run seed/build_db.py. Admin seeds/writes, entitlement toggles, and any deploys are user-authorized per occurrence.
• By-design vs bug: if a finding may be intended behavior, log it and keep going (don't stop to ask; don't unilaterally rewrite deliberate design — the log captures it).
• Pass C parallelism: set 5554 = Dark, 5556 = Light to capture both themes at once.
• Never log decrypted message/answer content.

Severity scale (label every issue)

• P0 Critical — crash/ANR, data loss, encryption/security leak, feature fully broken, premium bypass.
• P1 Major — feature partly broken, premium not unlocking for partner, wrong/missing notification, dead-end nav.
• P2 Minor — readability/contrast, clipping/overflow/truncation, theme not adapting, inconsistent styling, wrong/double-back navigation.
• P3 Polish — spacing/alignment/copy nits.

QA passes (Round 1 = baseline)

Pass A — Couple-shared premium (target: either partner premium → both unlock)

Test each gated feature in 3 states: neither premium → locked + paywall; partner-only premium → BOTH unlock; self premium → unlock. Toggle Sam premium, confirm QA (free) unlocks; toggle off. Features: Play-hub games (Desire Sync + any premium-badged), Connection Challenges, Memory Lane; Question Packs; Spin the Wheel / Category Picker / Wheel History (+ any premium wheel categories); Date Match / Plan Date / Date Builder; chat media + reactions + any premium chat tools (regression — already couple-shared); Subscription/Settings reflects entitlement. Gated files (for the fix): ui/play/PlayHubViewModel, ui/desiresync/DesireSyncScreen, ui/wheel/{CategoryPicker,SpinWheel,WheelHistory}*, ui/questions/QuestionPackLibrary*, ui/dates/{DateMatch,DateMatches}Screen, ui/memorylane/MemoryLaneScreen, ui/challenges/ConnectionChallengesScreen. Also: any VM/screen calling EntitlementChecker.isPremium() directly (grep for it) is a candidate gate.

• ENFORCEMENT, not just a checker-usage grep (mandatory — RETROSPECTIVE from A-201, R12). A feature can carry an isPremium content flag + a cosmetic PremiumBadge with NO gate at all — that's exactly how Date Match shipped a premium bypass (free users could view/like/match ★Premium date ideas; getDateIdeas() returned DateIdeaSeed.all, no CouplePremiumChecker, badge only). Prior rounds missed it because the audit grepped for CouplePremiumChecker usages and found the gated features, never noticing the feature that had no checker. So every round: (1) grep for isPremium / PremiumBadge / premium content flags (DateIdea.isPremium, category.access=="premium", challenge.isPremium, …) and for each confirm a real enforcement path exists — a CouplePremiumChecker filter OR a paywall-on-interaction — not just a badge; (2) actually TRY TO USE the premium content as a free user (like/open/play it), don't just confirm the lock renders — "badge shows" ≠ "gated". A badge with no enforcement = premium bypass (P1+). Inspection lesson: "shows a Premium badge" is a display fact, not a gate; prove the gate by using the content while free.

Pass B — Games lifecycle (MANDATORY: play each game ONE complete time through ALL different play stayles of the game)

Games: This or That, How Well Do You Know Me, Desire Sync, Connection Challenges, Memory Lane, Spin the Wheel, + Date Match.

• PLAY AS THE USER (mandatory mindset for this pass): drive every game the way a real user would — reach it through the actual in-app navigation a person would tap (Play hub → the game's card → its buttons), not via deep-links, admin pokes, forced state, or any shortcut a user doesn't have. Expect what the user expects: if a tap/button/flow doesn't do the obvious thing, or a screen doesn't behave the way a normal user would assume, that itself is a finding — log it.
• When something doesn't work: REPORT FIRST, then a minimal workaround (in that order). Do not silently engineer around breakage by taking extra steps the user wouldn't take. The moment the natural user path fails: (1) log the issue in ClaudeReport.md with severity + the exact user action that failed and what was expected; (2) only then apply the smallest workaround needed to keep the pass moving. The workaround never replaces the report — a flow that needs a workaround to proceed is, by definition, broken and must be filed to fix. If a workaround is impossible, mark the game fail→<id> (blocked) and continue with the next.
• A launch/crash check is NOT sufficient. Each game MUST be played one full way through, end-to-end, on BOTH devices — start → answer/interact through every step/round/question on each device → reach the finish/reveal/results screen → confirm the result renders correctly for both partners. Verify each intermediate screen and interaction works (selections register, progress advances, both-answered gating, reveal/scoring/summary correct). Premium games (Desire Sync, Memory Lane) need a premium toggle to play.
• The session lifecycle is exercised by the real playthrough: status active→completed; reveal/results correct on both.
• GAME JOIN PATHS (mandatory — the second partner must JOIN, not just co-play): the starter begins from real in-app nav; the joiner then enters from every user-facing entry point — notification tap, Play-hub active state, Home active-game card, Today prompt, waiting-room/resume screen, in-app foreground banner, game history/replay, and (after the natural paths) deep-link/crafted intent + cold-start from a push. A game isn't complete unless both partners can start, join, resume, finish, reopen results, and recover from a stale/ended session — with no duplicate sessions, wrong routes, stuck waiting screens, broken back nav, or premium-gate mistakes.
• FIRST-FINISHER → WAITING-PARTNER NOTIFICATION (mandatory state — async games): explicitly exercise the asymmetric state where one partner finishes their part and the OTHER is idle/away. The waiting partner MUST get a "your turn to play" nudge (partner_completed_part via onGamePartFinished) the moment the first finishes — async games (this_or_that / wheel / how_well / desire_sync) only flip to completed (→ partner_finished_game) once BOTH answer, so without the first-finish nudge the waiting partner is told nothing. Verify the idle partner (on Home, or backgrounded/killed) actually receives + can tap into the game. (This state was missed for a long time precisely because QA always played both sides through; "one finishes, the other never played" is its own required angle.)
• VARY THE STYLE OF PLAY (don't just repeat the happy path): across runs, deliberately exercise different ways a real couple would play each game, because different inputs hit different code paths:
  • Different DEPTHS and QUESTION COUNTS — cover the matrix, don't settle for one combo: play each game across every depth/mood (Light, Everyday, Deep, All-topics/shuffle) AND every round length / number of questions (5 / 10 / 15), in different pairings across runs (e.g. Light×5, Deep×15, Everyday×10, All×5) — short and long sessions, shallow and deep content. Different depths surface different question sets, tones, and edge content (e.g. Deep/Desire-Sync sensitive prompts); different counts stress pacing, progress, and the both-answered gate. Also exercise each distinct answer type (A/B, Yes/No, True/False, 1–5 scale, multi-select, free-text).
  • Different answer patterns that change the result — all-match vs all-mismatch vs partial; both-yes vs both-no vs split (so reveals show "shared", "all private", "0 matches", "perfect/zero score" — verify each renders right).
  • Different turn orders / who-starts — partner A starts vs partner B starts; the guesser opens before vs after the subject finishes; both open simultaneously (race); one device much slower than the other.
  • Different exit/resume styles — finish normally; quit mid-game; background mid-game then resume; cold-kill mid-game then reopen; "End their game"; re-open a completed session for the replay/results; play two games back-to-back, and a different game type immediately after.
  • Edge inputs — submit with nothing selected (should be blocked), rapid double-taps on answer/confirm/next, spamming the start button, tapping during the reveal animation, switching tabs mid-game, receiving/tapping a notification mid-game. None should crash, duplicate, or desync.
• Edges: re-open a completed session, leave mid-game (resume), no stuck session, no crash, logcat clean.
• Game start/finish pushes (onGameSessionUpdate) exercised here; full delivery/deep-link audit in Pass E.
• Media permissions (CAMERA, RECORD_AUDIO): granted works, denied degrades gracefully.
• Done = every game has one verified complete playthrough (a launch-only "opens, no crash" row is partial, not pass). Coverage row format: game × starter × join-entry × premium-state × depth/count × lifecycle-edge × result; only pass when start/join/play/finish/reopen/recover are all verified.

⛔ Pass C — Visual pass, light + dark, ALL screens (MANDATORY: run scan BEFORE sweep)

⛔ CLAUDE: Run the automated theme scan (below, Automated Tier 1) before starting the visual sweep. Read the output at /tmp/claude-theme-scan-<date>.md and file findings to ClaudeReport.md first. The sweep must verify every flagged screen in BOTH themes.

Every route in core/navigation/AppRoute.kt (~50), in both modes: text contrast/readability (no invisible/ low-contrast), no clipping/overflow/ellipsis breakage, icons visible, backgrounds adapt, controls legible. Groups: auth/onboarding/pairing (fresh acct); Home (solo + paired); Play + every game; Today + reveal/history; Messages (inbox + conversation); Packs; Dates (Match/Builder/Matches/Bucket List); Wheel (picker/session/complete/history); Settings + all sub-pages (Account, Notifications, Appearance, Privacy, Subscription, Relationship, Security, Delete Account); Paywall; Your Progress/Activity; Recovery.

• Images must belong to the screen: during the UI sweep, visually inspect every illustration, glyph, banner, empty-state image, pack art, celebration asset, and dark/light variant in context. It should feel intentionally integrated with the page hierarchy, copy, spacing, and action area — not like a forgotten placeholder dropped into an empty slot. Check crop, scale, padding, alignment, corner radius, background/tile treatment, theme variant, edge treatment, loading/fallback state, and whether the image competes with or clarifies the primary task. If it is broken, clipped, low-contrast, off-brand, stale, or placeholder-looking, file a bug in ClaudeReport.md; if the screen works but would benefit from new/better art, log the prompt need in ClaudeBrandingReview.md.

• SOFT EDGES — art must fade into the screen, not show a hard tile edge (mandatory): every displayed illustration should blend/feather softly into the background, not sit as a hard-edged rounded rectangle/card with a visible boundary or border line. Inspect each illustration's edges against the screen on both themes — a crisp tile edge, outline/border, or a pale block floating on the surface is a finding (C-ART-EDGE-001). (Fixed R11: BrandIllustration now feathers its 4 edges to transparent via Modifier.graphicsLayer{compositingStrategy=Offscreen} + drawWithContent BlendMode.DstIn linear gradients — clip+border removed — and EmptyState routes its illustration through BrandIllustration, so all tiled art melts into the surface. Recurring check: verify it still holds and that any NEW art helper / direct painterResource tile also feathers.) Fix pattern (if it regresses): feather the edges to transparent, or a vignette matching the surface, or ship transparent-edged art — applied in the shared BrandIllustration/EmptyState helpers so it's consistent everywhere.

• ⛔ CLAUDE — RUN THE AUTOMATED THEME SCAN FIRST (MANDATORY, BEFORE THE VISUAL SWEEP): Do NOT start the manual visual sweep until the automated scan has completed and you have reviewed its results. The scanner is scripts/theme-scan.sh. Run it from the project root and save the report:

  cd /home/kaspa/.openclaw/Projects/relationship-app
  ./scripts/theme-scan.sh > /tmp/claude-theme-scan-$(date +%Y%m%d).md
  cat /tmp/claude-theme-scan-$(date +%Y%m%d).md
  

  The script reports findings by severity and ends with a ## Summary section showing the exact counts. Record those counts in ClaudeQACoverage.md under Pass C before starting the visual sweep.

  • 🔴 CRITICAL — container/surface/background set to a hardcoded color. Will produce visible light/dark mismatches. Example: Surface(color = Color.White) inside a dialog in dark mode.
  • 🟠 MAJOR — component color overrides or direct painterResource that bypasses BrandIllustration. Likely to break theme adaptation or decoupled-theme art.
  • 🟡 REVIEW — hardcoded text/icon/border/gradient colors that may be correct on a branded container but must be verified in both themes.

  ⛔ CLAUDE: You are explicitly allowed to improve scripts/theme-scan.sh and this Pass C methodology whenever you discover a new light/dark failure mode. Examples: new Compose patterns that evade the current grep, a new color token that should be checked, better false-positive filtering, or converting the output to JSON/CSV. Keep the script runnable from the project root and update the script header with what changed. Do not remove existing patterns unless they are provably wrong.

  After running the scan: read the report, file all CRITICAL and MAJOR findings to ClaudeReport.md as Pass C theme defects, then proceed to the manual visual sweep. Any screen flagged CRITICAL/MAJOR must be verified in BOTH themes during the sweep. If you fix hardcoded colors during the QA round, re-run the scan to confirm they are gone.

  Tier 2 — Theme definition validation: scripts/theme-scan.sh also validates that darkColors in Theme.kt has every required Material3 slot explicitly defined. If a slot is missing, log it to ClaudeReport.md as a P2 theme defect.

  Tier 3 — Compose screenshot diff suite (endgame, not yet implemented): The true "catch everything" solution is an automated screenshot comparison pipeline that renders every route in light mode, renders the same route in dark mode, and pixel-diffs them — flagging any screen where the dark version has white backgrounds, invisible text, or wrong-variant art. This catches compositional and gradient-based mismatches that static analysis cannot. When implemented, use papAROS, Shot, or Roborazzi with a custom darkTheme = true test parameter for each route. Log this to Future.md as "Tier 3: Compose screenshot diff for visual regression".

• THEME-VARIANT ART must follow the IN-APP theme, not just the system (mandatory — RUN THE DECOUPLED STATE): the app has its own theme toggle (Settings → Appearance → Light/Dark/Device) that swaps Compose colors but does not change the Android config uiMode, while -night drawables (drawable-night-nodpi/) and painterResource resolve off the system uiMode. So art can mismatch the UI when the two disagree. Test the decoupled state explicitly, every round: force system light then set the app to Dark, and force system dark then set the app to Light, and on every screen that has a dark art variant confirm the illustration matches the in-app theme (no bright/light tile on a dark screen, no dark tile on a light screen). Commands: adb -s <serial> shell cmd uimode night no (system light) / … night yes (system dark); then toggle the in-app theme in Appearance. Screens with -night variants to check: Security (privacy_recovery), Memory Lane, Bucket List, Answer History, Date Match (empty + success), Connection Challenges header, Pairing success, Messages empty, Past Games, Quiet-hours, Account-deletion, + any new illustration_* added to drawable-night-nodpi/. Restore cmd uimode night auto after. Light art on a dark screen (or vice-versa) when the in-app theme is switched = bug (P2 theme-not-adapting; see C-DARKART-001). (Fixed R11: CloserTheme provides LocalAppInDarkTheme; BrandIllustration loads each drawable through context.createConfigurationContext(cfg) whose UI_MODE_NIGHT_* is set from LocalAppInDarkTheme, so the -night variant follows the IN-APP theme, not the system. Verified live R11 both decoupled directions. Recurring check: re-run the decoupled state and confirm it still holds, including any newly added -night art.) Fix pattern (if it regresses): drive the resource uiMode from the in-app theme as above, or AppCompatDelegate.setDefaultNightMode/config override, so painterResource picks -night per the app's own setting.

• EVERY image needs BOTH a light AND a dark variant matching the theme (mandatory — audit every image-bearing page). It is not enough that the -night mechanism works — the dark asset must exist. Go page by page through every screen that shows an illustration / hero / banner / empty-state / pack art and confirm there is a real dark variant in drawable-night-nodpi/ (and a light one in drawable-nodpi/) that matches the in-app theme — a light/pink image shown on a dark screen (because only the light asset exists) is a bug, even with feathered edges. Cross-check each page against the Image theme-variant coverage table in ClaudeBrandingReview.md; a missing variant is filed as a bug in ClaudeReport.md and the dark/light asset to create is logged in ClaudeBrandingReview.md as a prompt to be made. (2026-06-27 audit found all illustration_couple_* heroes, daily_question, partner_activation, tonight_partner_prompt, together_empty, and all 10 pack_art_* are light-only — these still need dark variants.) Only genuinely theme-agnostic transparent/celebration art is exempt, and only after you verify it reads on both.

• EVERY icon/glyph must be a CUSTOM Closer glyph — no generic Material icons, no generic hearts (mandatory). On every screen, inspect each icon: any generic Material icon (Icons.Filled.*/Icons.AutoMirrored.*/Icons.Default.* — ArrowBack, Favorite/FavoriteBorder, Person, Lock, Star, PlayArrow, Check, Close, Send, …) is a placeholder, not brand → a finding. File it as a brand defect in ClaudeReport.md and log the custom glyph_* to make in ClaudeBrandingReview.md (see its Icon/glyph audit). Reflex grep to find them: grep -rE "Icons\.(Filled|Outlined|Rounded|Default|AutoMirrored)\." app/src/main/java/app/closer/ — every hit is a generic icon that needs a bespoke Closer glyph (ImageVector.vectorResource(R.drawable.glyph_*) + Icon(tint=…)). (2026-06-27 audit: ~60 distinct Material icons across ~201 call sites still to replace.)

• States, not just happy path: empty / loading / error / not-paired / locked-premium / signed-out / stale-or-deleted-target / populated-with-many where they exist; many need data setup (seeding is user-gated) — note unreachable states in coverage rather than skipping silently.

• Text/data stress: test long names, long relationship labels, long question/answer text, emoji, multiline content, empty optional fields, many list items, and both partners having similar names. Verify no clipping, overlap, confusing attribution, broken sorting, or hidden actions.

• Readability at scale: default font size + spot-check largest system font scale on text-heavy screens. (The full accessibility sweep — large-font on every primary flow, TalkBack labels, touch targets, keyboard, reduce-motion — is Pass J; per-route performance/jank is Pass I.)

• Orientation / form-factor (the app is NOT portrait-locked — AndroidManifest.xml declares no screenOrientation, so it DOES rotate to landscape). Don't only check "rotation doesn't lose state" (that's Pass F) — verify the landscape layout actually renders correctly on the text-heavy / game / paywall / dialog screens (adb shell settings put system accelerometer_rotation 1 then rotate, or use the emulator rotate control): no clipped/cut-off content, no broken scrolling, dialogs and bottom CTAs still reachable. Spot-check a large-screen / tablet AVD too. If landscape is not a supported experience, the correct fix is to lock portrait in the manifest — file that as the finding (see the app-finding note) rather than shipping an unverified landscape layout.

• Navigation from every entry point: reach each screen from all the places that link to it and confirm it opens correctly each time — e.g. a conversation from the inbox AND from "Discuss" AND from a notification; a game from the Play hub AND from a notification; Paywall from each gated feature; Settings sub-pages; reveal from Today AND from history AND from partner_answered. A screen that works from one entry but breaks/duplicates from another = bug.

• Every link, CTA, and mission must prove its destination: actively hunt for dead buttons, wrong targets, generic Home fallbacks, no-op taps, stale routes, and confusing affordances. Example class: a Reveal card saying "Tiny Mission: Send one flirty text" must open the relevant Messages/conversation flow, not do nothing. For every button/card/chip/row, record the expected destination before tapping, then verify the actual destination, state, payload, and back stack. Broken/no-op/wrong-destination CTA = bug (usually P2; P1 if it blocks a core flow).

• All routes into a game / join-game state (verify each opens the correct game + session + partner-state + mode + premium/couple-entitlement + back stack): Play-hub cards (incl. premium-gated), active-session banners, Home/Today game prompts, game history, replay/results, waiting screens, notification-opened screens, in-app banners, "join/resume/continue/view results/end (their) game", deep-link/crafted intent, and bottom-tab return into an active game. Wrong/duplicate destination, double-back, stale-session join, dead-end, or a route that bypasses the premium/couple check = bug.

• TAKE EVERY AVENUE (exhaustive nav fuzzing — actively hunt for nav bugs, don't just walk the happy path): treat navigation as something to break. On every screen, tap every interactive element — each button, card, row, icon, chip, link, tab, header back-arrow, system back, and any "see all / history / edit / manage" affordance — and follow where it goes. Then try the combinations and sequences a curious user hits:

  • Every order: switch bottom tabs in many orders, mid-flow (open a game, jump to Messages, come back); enter a deep screen then tab away then back; open A→B→C then back-back-back.
  • Rapid / repeated input: double- and triple-tap navigation targets (especially "open game", "Play now", "Create/Start session", notification taps) to surface double-push/duplicate-screen/stale-route bugs (cf. B-004).
  • Interrupt mid-navigation: background/rotate/lock during a transition; tap a notification while already on that screen, on a different screen, and while logged-out/unpaired; cold-start straight onto a deep link.
  • Dead-ends & traps: from every screen confirm there's always a way out (back/close/home) — no screen that strands the user, needs two backs, exits the app unexpectedly, loops, or lands blank. Re-check the asymmetric-game waiting screens, replay/results screens, and paywall specifically.
  • Log every wrong/duplicate/dead destination with the exact tap sequence to reproduce. Wrong/double-back or dead-end = P2 (P1 if it traps the user or loses their progress).

• Back-stack / "double back": from every entry point, system back AND the in-app back arrow return to the correct previous screen — no dead-ends, no exiting the app unexpectedly, and no screen that requires pressing back twice (duplicate/stacked destinations on the back stack = bug). Bottom-tab reselection and deep-link/ notification entries must land with a sane back stack (back → Home, not off the app or a blank screen). Wrong/ double back or a dead-end = P2 (P1 if it traps the user).

• UI consistency / polish defects: compare each screen against sibling patterns in the same area and across the app. Headers, labels, status chips, partner names, connected-state copy, spacing, card treatments, and button hierarchy should feel intentional and consistent. Awkward or out-of-place UI such as a Settings relationship row where "Connected with ..." looks visually odd, cramped, misaligned, or unlike the rest of Settings is a finding: file as a bug if it looks broken/inconsistent; log to Future.md only if it is purely a product/content improvement.

Pass D — Security & encryption (cornerstone; findings default to P0)

Read first: manual's E2EE model · Firestore rules · Encryption versions. The cornerstone: every private field is ciphertext at rest, rules hold against non-members, keys/recovery are sound. D3 (live negative raw-API) is MANDATORY every round — never deferred to "only 2 emulators" (mint a non-member token via admin → Identity Toolkit signInWithCustomToken → Firestore REST). Run all of D1–D7:

• D1 At-rest coverage: admin-read RAW docs/objects, assert ciphertext for every private type — chat text + lastMessagePreview (enc:v1:), chat media bytes (Tink 01 69 59 51 f0…), answers (sealed:v1:/enc:v1:), date plans + date_swipes, Memory Lane capsules, Bucket List. Also: wrappedCoupleKey + recovery material never plaintext; invite code (KDF seed) never stored raw; no push payload carries private content.
• D2 Rules audit (static): member-only reads, author/server-only writes, ciphertext enforced on every private field, immutability, no premium self-grant, entitlements write:false; re-audit conversations/typing/reactions
  • entitlement partner-read; no catch-all match /{document=**}; list/query not enumerable; get()-rules don't over-expose; no legacy plaintext/downgrade path (coupleEncryptionEnabled holds; no disabled-encryption branch).
• D3 Negative access tests (EXECUTE LIVE via raw API — do not defer): a non-member account is denied reading messages/answers/dates/entitlements/sessions/capsules, writing plaintext to encrypted fields, self-granting premium, and any cross-couple access. Run it the raw-API angle: mint a non-member ID token (admin custom token → Identity Toolkit signInWithCustomToken REST) and issue Firestore REST GET/PATCH against the couple's docs — expect App Check 403 or rules PERMISSION_DENIED on every attempt. Also issue the same reads with a member token to characterize the enforcement layer (App Check vs rules). Any unauthorized 200 with couple data = P0.
• D4 Key exchange / management / recovery (E2EE crux): couple key client-generated, only leaves device wrapped (KDF from invite seed; server holds only wrappedCoupleKey+kdfSalt/kdfParams+encryptedRecoveryPhrase); KDF strength; Tink AEAD = AES-GCM/256 with AAD=coupleId, no weak/custom crypto/nonce reuse; keybox/sealed/commitment integrity; recovery-wrap server-blind; unpair revokes decrypt; invites CSPRNG + single-use + expiry.
  • NEW-DEVICE / LOST-PHONE RECOVERY — drive it end-to-end, don't just verify the phrase is revealed (the make-or-break data-continuity path for an E2E app). The keys are single-device (a known limitation); the recovery phrase is the only bridge. Infra: crypto/RecoveryKeyManager.kt, data/local/RecoveryPhraseStore.kt, ui/pairing/RecoveryViewModel.kt, crypto/CoupleKeyStore.kt. Exercise the full flow on a fresh install / second device: sign in → enter the recovery phrase → the couple key is rebuilt → prior enc:v1:/sealed:v1: messages and answers actually DECRYPT and render (not just new ones). Then the failure paths: a wrong/typo'd phrase fails gracefully (clear error, no crash, no corruption); a user who lost the phrase is told honestly what is/isn't recoverable; and throughout, the partner's device keeps working (one side recovering must never break the other). Confirm the server stayed blind (only wrappedCoupleKey/encryptedRecoveryPhrase ever transit — verify via admin read). Without this, "I got a new phone" silently loses the relationship history. (Also exercised from the account-lifecycle angle in Pass F and the Settings → Security flow in Pass M.)
• D5 App Check / Functions / secrets: App Check enforced; callables validate auth+membership; webhook authenticity; admin-only writes rejected from clients; service-account JSONs never committed; no plaintext/secrets in logcat; temp files deleted.
• D6 Leak vectors: no private content in analytics/crash; allowBackup=false + backup rules exclude sensitive data; deep links re-check membership; clipboard user-initiated; consider FLAG_SECURE; repo scan for committed secrets.
• D7 Encryption migration: test the encryptionVersion paths (0 plaintext → 1 migrating → 2 strict) on a legacy couple — migration completes without exposing plaintext or losing/garbling old content, and a half-migrated couple is safe (no mixed read failures, no downgrade). This is the riskiest data path for existing users.

Pass G — Account creation, validation & fake-account abuse (MANDATORY — both the happy path AND the attacks)

Cover every account-creation avenue a real user takes and every fake/abusive creation attempt an attacker would try. Use throwaway test accounts (sign-out → fresh sign-up; never pm clear). Report-first like every pass.

• Real creation flows (happy path + validation): sign-up (email/password and any social/anonymous path), profile creation, and pairing — both create-invite and accept-invite sides. Verify field validation (invalid/empty email, weak/short password, mismatched confirm, name length/emoji/unicode), the error copy is friendly (no raw SDK/Firebase error leaking — cf. A-OBS), loading/disabled states, and that a brand-new unpaired account lands on the correct "create or accept invite" home (not a broken/blank or paired view).
• Duplicate / conflicting creation: sign up with an already-registered email (clear "already in use", no crash, offer sign-in); create a second account while one is signed in; re-run onboarding after completing it; accept an invite while already paired (must be rejected cleanly); two devices accepting the same invite (single-use — the second must fail gracefully).
• Fake / malicious creation attempts (security — expect DENY, never crash or leak): create an account that is NOT a member of the test couple and attempt every cross-couple action (read messages/answers/dates/entitlements, write to the couple, self-grant premium/hasPremium, join/hijack pairing with a guessed/expired/reused invite code) — all must be denied by rules (this is the live execution of D3). Probe invite-code abuse: replay a used code, use an expired code, brute-force/guess attempts (CSPRNG entropy + single-use + expiry must hold). Probe App Check: a request without a valid token is rejected. Confirm a malformed/forged sign-up can't bypass profile or membership requirements. Any successful unauthorized create/read/write = P0.
• Account lifecycle around creation: sign-out → sign-in (state restores, no stale couple); delete account then re-create with the same email (clean slate, partner notified/unpaired); an unpaired/just-created account tapping a stale notification or deep link is handled gracefully (no crash, sane landing).
• Done = every creation avenue exercised (happy + duplicate + malicious) with each attack denied and each happy path validated end-to-end; findings filed with exact repro.

Pass E — Full notification suite, deep-links & join-game navigation (every type, both clients, every app state)

Run the complete suite across both clients (QA→Sam AND Sam→QA). Each type verified end-to-end: trigger fires → delivered to the right partner (never self/non-member/ex-partner) → correct channel + copy with no private content → tap opens exactly the right item (loaded, not generic Home/dead-end) → sane back stack → privacy/authz re-checked on open. No duplicates; rate limiter (20/day, 100/week) doesn't drop legit ones.

• Notification chunk contract (small chunks, complete coverage): each chunk owns one notification type (or one explicit subchunk of that type, e.g. chat_message QA→Sam foreground/source-screen sweep, then chat_message Sam→QA background+killed+stale). Before starting, write the chunk's matrix in ClaudeQACoverage.md; after finishing, mark each cell pass | fail→id | blocked→id | not implemented→Future.md. A notification type is not complete until all applicable cells below are covered:
  • Directions: QA→Sam and Sam→QA; sender must not receive their own push unless intentionally designed.
  • Process states: foreground, background/warm, killed/cold-start, force-stopped if deliverable, screen locked, and resumed after rotation/process recreation when relevant.
  • Current screens: Home, Play hub, active game/waiting/results, Today/reveal, Messages inbox, exact conversation, Settings/sub-settings, Paywall, unrelated deep screen, logged-out, unpaired, and stale prior-partner context.
  • Entry surfaces: foreground in-app banner/head, Android system tray tap, any push action button, crafted deep-link/intent matching the payload, repeated/double tap, and tap after the target has changed.
  • Targets: fresh target, already-open target, completed target, stale/expired/deleted target, unauthorized target, wrong couple/session/item ID, malformed/missing extras, and no-network-on-open.
  • Assertions: correct recipient, correct channel/priority/copy, no private payload/log content, exact destination, membership/auth/entitlement re-check, no duplicate route/session, sane back stack, logcat clean, and coverage/docs updated before the chunk ends.
• Notification tap crash triage (mandatory): never conclude "the notification didn't open" from UI behavior alone. Before each notification/deep-link tap, clear or timestamp logcat; after the tap, inspect both devices for FATAL EXCEPTION, ANR, ActivityTaskManager errors, RuntimeException, navigation/deep-link exceptions, PERMISSION_DENIED, and swallowed repository/decryption errors. If the app returns Home, stays put, flashes, restarts, or silently fails, classify whether it was wrong routing, missing extras, stale data, permission denial, or a crash. Any notification tap that crashes (example class: tapping a game notification to open Spin the Wheel) is a filed bug with stack trace + exact payload/session/game type, not a vague "didn't open" note.
  • Test the REAL launch path, not a synthetic one. adb am start … --es type=… does not reproduce a real notification tap: the OS notification tap launches the activity through the SysUILaunch splash handover (reportSplashscreenViewShown → handOverSplashScreenView), which am start skips. A whole bug class (e.g. the splash-exit provider.iconView NPE — the handover delivers a splash view with no icon, SplashScreenView: Icon: view: null, on notification cold-starts only) crashes onCreate → "Force finishing activity" → the app opens-and-closes, yet am start AND the normal launcher icon both pass. Verdict: for cold-start/notification routing, a synthetic-intent pass is not a pass — confirm with a real push tapped from the shade on an am kill'd app.
  • "Opens and closes / flashes / returns to launcher" ⇒ assume a crash; pull the stack FIRST. logcat -c before the tap, then grep FATAL EXCEPTION|AndroidRuntime|Force finishing|getIconView. A real repro + the stack trace beats code-reasoning every time (this bug was misdiagnosed as deep-link routing until the live stack named MainActivity.kt + SplashScreenViewProvider.getIconView). Confirm crashes reach Crashlytics so field cold-start crashes surface.
  • Many notification types "broken" at once ⇒ suspect the SHARED entry path (splash/onCreate/launch), not each handler. When chat AND every game's results push all fail identically, the bug is in what they share (the cold-start path), not per-type routing. Re-run a cold-start smoke after ANY change to MainActivity / splash / theme / manifest / launchMode / branding-"loading state" commits — these cosmetic-looking changes broke the launch.
  • For "worked before, broken now": git blame / git log -L the crashing line/function to pin the introducing commit, then re-test that exact path on it.
• Both-client × app-state matrix (per type): QA→Sam and Sam→QA, each in foreground / background / killed (cold-start), plus already on the target screen, on a different screen, logged out, unpaired, with a stale/expired/completed/deleted target, and both users opening around the same time. Not a pass unless it works from both clients in every state that applies.
• Current-screen/source-screen matrix (per type): do not test notifications only from Home or only from a clean launch. For each notification type, vary where the receiving client is when the notification arrives/taps: Home, Play hub, active game/waiting/results, Today/reveal, Messages inbox, exact conversation, Settings/sub-settings, Paywall, an unrelated deep screen, app backgrounded from each major tab, and app fully closed/killed. Foreground banners, system-tray taps, warm-start onNewIntent, and cold-start launch must all route to the exact target. A tap that lands on generic Home, stays on the old screen, opens the wrong tab, loses extras, duplicates the destination, or needs a second tap is a bug.
• Permission/token health: cover Android POST_NOTIFICATIONS granted, denied, "don't ask again"/system-disabled, and re-enabled states; Settings notification toggles; sign-out/sign-in token refresh; same account on two devices; partner/account switch; stale token cleanup; app reinstall/update; and notification channel migration. Denied/system disabled notifications should fail gracefully with in-app state still correct, never with lost data or broken routing after permission is restored.
• Doze / battery-optimization / background-restriction delivery (real-device gate — emulators NEVER enter these states, so per-round emulator passes systematically miss the #1 real-world "notifications don't work" cause). Scheduling is entirely server-side (no client WorkManager), so the only thing standing between a fired push and the user is the OS power state. On a physical device, verify each push type still delivers when the recipient device is dozing / the app is battery-optimized or "Restricted": adb shell dumpsys deviceidle force-idle (then send a real partner action + a scheduled push), app set to Optimized then Restricted in battery settings, and App Standby buckets. Assert: high-priority FCM (partner actions) wakes the device and delivers; lower-priority/data-only pushes degrade predictably (document which); scheduled pushes (daily question, capsule unlock, reminders) still arrive within the expected window. Because our recurring setup is two emulators, keep this as a blocked→needs-device row in ClaudeQACoverage.md (with the device-matrix gate) rather than silently assuming delivery — and run it before any store push. OEM battery-killers (Xiaomi/Samsung/etc.) are even more aggressive; note them for the device matrix.
• Six assertions per notification: (1) trigger fires correctly — right event, not early, not twice, sender doesn't get their own (unless intended), retry/idempotency doesn't duplicate; (2) delivered to the right person — correct token, old tokens unused after sign-out/account-switch; (3) copy + channel correct — friendly, right channel/ priority, no raw Firebase error/raw IDs, no private content in text/payload/logs/analytics/crash; (4) tap opens the exact destination — specific conversation/session/capsule/match/question/settings/pairing, never blank, never a crash on missing/stale/malformed/unauthorized data, no duplicate/stacked copies, completed→results/replay, expired/deleted→ graceful fallback; (5) back stack sane — back returns sensibly (Home/prev context), no double-back, no unexpected exit/loop/blank; (6) deep-link re-checks auth + couple membership + pairing + entitlement + target ownership + session status + existence — a non-member/logged-out/stale/unpaired open must NOT reach private content and must fail gracefully.
• qa/qa_push.js is faithful to the PUSH, not the TRIGGER — assertion #1 needs ≥1 real in-app action per round. qa_push.js sends the FCM via admin (messaging().send), so it faithfully reproduces delivery + channel/copy + cold-start launch + tap-routing (use it for the bulk of the type×state matrix and the entrypoint_smoke.sh smoke). But it bypasses the Cloud Function — no onMessageWritten/onGameSessionUpdate/onAnswerWritten/createDateMatch actually ran. So a qa_push.js-only round can never satisfy assertion #1 (trigger fires correctly): a broken or un-deployed trigger (Firestore-path change, deploy regression, rules change) is invisible to synthetic pushes. Each round, drive ≥1 real in-app partner action (send a chat, finish a game, answer the daily Q) and confirm the matching push lands on the partner. (UI-automation tip: the chat composer's send button is the rightmost control in the composer row, content-desc Send; if uiautomator taps mis-fire, verify the action via admin read — the new message/answer doc exists enc:v1: — rather than claiming the trigger from a synthetic push.)
• Inventory (type → Cloud-Function trigger → recipient → destination) — verify each; mark any unimplemented type not implemented→Future.md (don't count as pass): chat_message(onMessageWritten → partner → conversation; foreground→chat-head bubble) · partner_started_game/partner_finished_game(onGameSessionUpdate → partner → game/join · results/reveal) · partner_completed_part(onGamePartFinished → waiting partner → game; fired when the FIRST player finishes an async game so the partner is told "your turn" — async games complete only when BOTH answer, so without this the waiting partner got nothing between first-finish and both-finish) · join_game/game_invite & partner_joined_game (if present → partner/starter → join screen · waiting-room update) · partner_answered(onAnswerWritten → partner → reveal) · game_abandoned/game_ended (if present → partner → safe ended state, not a stuck session) · daily_question(assignDailyQuestion)/daily_question_reminder/daily_reminder(dailyQuestionReminder → Today) · date_match(createDateMatch → match) · date_plan_update (if present → date plan/builder/match) · partner_joined+invite_created(acceptInviteCallable → pairing/home) · partner_left(onCoupleLeave)/partner_deleted_account(onUserDelete → home/relationship settings) · memory_capsule_unlocked(scheduled → capsule) & memory_capsule_created (if present → Memory Lane/locked capsule) · challenge_day_ready(→ Connection Challenges) & challenge_day_completed (if present → challenge progress) · outcome_reminder(scheduledOutcomesReminder) · reengagement(reengagement/gameRetention) · gentle_reminder(sendGentleReminderCallable) · spki(key identity/confirm → security/key screen) · subscription_entitlement_changed & security_recovery (if present).
• Game-notification suite (per game): A starts from Play hub → B gets the start/join push (if supported) → B taps and lands on the correct join/waiting/active screen → B can join from there → A sees B joined/answered → both finish → finish push opens the exact results/reveal → re-opening the push after completion opens replay/results (not a dead active session) → if A ends/quits, B is notified or shown a graceful ended state → a stale game push routes to results/history or a clear expired-session message → simultaneous start/join yields one session, neither stuck → premium gate holds (neither-premium push must NOT bypass paywall; either-premium unlocks for both). For each game type, including Spin the Wheel, notification taps must be paired with logcat review so crashes are caught even if the visible symptom looks like a no-op or generic Home fallback.
• Join-game navigation suite: every entry that leads to joining/resuming a game opens the correct game + session + partner-state + mode + entitlement + back stack — Play-hub card, active-game banner/card, Home active-game card, Today game prompt, notification tap, in-app foreground banner, game history/replay, partner waiting screen, results/ reveal, "End their game"/stuck-session recovery, deep-link/crafted intent, cold-start from push, bottom-tab return into an active game, any push action buttons, and any "join/resume/continue/view results/play again". No wrong game type, no accidental stale-session join, no duplicate session on double-tap, back returns correctly.
• Payload security (P0 on any hit): inspect raw payload + logs — no plaintext message/answer/capsule/date-plan/ bucket-list/swipe content, no raw invite code/seed, no recovery phrase, no wrapped/decrypted key material, no email/name unless intentionally public; payload carries only the minimum routing metadata. Any private content = P0.
• Malformed / stale intents: fire crafted deep-links with missing/unknown type, missing/wrong target or couple ID, wrong game type, expired/completed/deleted target, unauthorized couple/session, malformed params, duplicate/rapid taps, a push for another user/previous partner, while logged-out/unpaired, while on the target screen, and during a different active game → never crash/leak, always a graceful fallback + sane back stack.
• Scheduled/time-based: trigger manually (invoke callable/function or seed the due condition — user-gated).
• Foundations: FCM token registration on sign-in (TokenRegistrar) + onNewToken + token cleanup on sign-out/ account-switch; POST_NOTIFICATIONS prompt + denied path; channels (di/NotificationModule); deep-link routing (MainActivity.deepLinkRouteFromIntent → AppNavigation); foreground/background split (core/notifications/AppMessagingService); no duplicate local+remote notification.
• Coverage: record per row type × trigger × recipient × app-state × destination × back-stack × privacy × both-client in ClaudeQACoverage.md; only pass when delivery + routing + back-stack + privacy + both-client are all verified. Missed delivery or wrong deep-link = P1; private content in any payload = P0.

Pass F — Resilience, concurrency, lifecycle & time (cross-cutting; a 2-user realtime app needs these)

• Concurrency / realtime races (two partners at once): both answer the daily question simultaneously; both start/join the same game; both swipe a date / react at once; one quits while the other submits; both tap a notification at once; partner acts while you're mid-flow. No lost writes, no stuck state, no duplicate sessions, reveal still correct. (This is where a couples app breaks.)
• Lifecycle / process death: background mid-flow + return; force-kill the app and relaunch (Android may kill the process) — state/auth/draft restore sanely; deep-link/notification after process death still loads (verified for chat — extend to all). Rotation/config-change doesn't lose Compose state. Low-memory.
  • Deterministic state-restoration ("Don't keep activities" — do NOT rely only on am kill). am kill is non-deterministic; enable Developer options → Don't keep activities (adb shell settings put global always_finish_activities 1) so the Activity/process is destroyed on every backgrounding, then walk each primary flow (sign-up, pairing, a game mid-answer, an unsent message draft, capsule/Date Builder in progress, paywall) and background→return at each step. Assert no lost form input, scroll position, draft, in-progress game state, or nav back-stack — i.e. rememberSaveable/SavedStateHandle actually persist it. Restore with adb shell settings put global always_finish_activities 0 after.
• Interruptions mid-flow (the OS or another app steals focus): incoming phone call, alarm, another app taking the foreground, screen-off/on, split-screen / multi-window, and picture-in-picture during a game/answer/message-compose → returning resumes cleanly with no lost state, no crash, no duplicate submit, and audio/camera (voice note, photo) releases + re-acquires sanely.
• Cold-start launch integrity from EVERY entry point (Pass F OWNS this — it's the shared path no other pass owned, and where the splash-crash hid): the app must open AND stay (no crash, no "opens-and-closes", lands off the launcher) when cold-started from: the launcher icon, each notification type tapped from a killed (am kill) app, a deep link, and any widget/quick-action. This is the MainActivity/splash/onCreate/auth-bootstrap path; a crash here (e.g. splash-exit iconView NPE) breaks all notifications at once. Run qa/entrypoint_smoke.sh here every round and after any MainActivity/splash/theme/manifest/nav/notification change. Reproduce via the REAL push tapped from the shade (not am start); "opens-and-closes" ⇒ pull the FATAL stack (see Pass E crash-triage).
• Network resilience: offline / flaky / airplane mid-action across answers, games, dates (not just chat media) — graceful failure + retry/queue, no crash, no silent data loss, recovery on reconnect.
• Idempotency / rapid input: double-tap send/submit, rapid nav, double-start, double-join, repeated paywall-unlock taps — guarded (no double-send, no duplicate session, no crash).
• Time-dependent behavior: daily-question rollover (6 PM CST assignment), streak day-boundary + repair window, capsule unlock times, reminder schedules, challenge-day availability, timezone change — test across a date change (manipulate device clock / trigger functions).
• Account/couple lifecycle: brand-new (empty) account; unpaired state; pair → unpair → re-pair; partner leaves mid-session; account deletion cascade; same account on two devices; stale notifications after unpair/delete are graceful; invite accepted while already paired is rejected cleanly. No orphaned/broken state.
• Install/update/migration lifecycle: fresh install, update over an existing signed-in install, app data retained, Room/DataStore/SharedPreferences migrations, notification channel migration, cached encryption/key material, pending deep links/notifications across update, and version-skew between partners if one device updates first. No sign-out loops, stale build routing, lost local state, broken permissions, or migration crashes.
• Crash reporting: confirm crashes/ANRs are actually captured (Crashlytics) so field issues surface.

Pass H — Branding & artwork (every screen: could it carry more of the brand? where would art help?)

Branding review is a MANDATORY part of QA every round (not an optional polish pass) — its findings + the assets to create are logged in ClaudeBrandingReview.md. A consumer-mindset pass focused on brand presence and delight AND two hard brand standards. Walk every screen and surface and ask: does this feel like Closer (private, warm, equal, intentional — a ritual for two)? Could brand color, the heart mark, a brand message, or an illustration make it warmer or clearer without clutter? Output is artwork descriptions written as ready-to-paste ChatGPT image-generation prompts — the user generates the images; we only describe them.

• MANDATE 1 — every image has a light AND a dark variant (theme-matched). Cross-check every image-bearing page against the Image theme-variant coverage table in ClaudeBrandingReview.md; a light-only image shown on dark (or vice-versa) is a bug → ClaudeReport.md and the missing variant is a prompt to make → ClaudeBrandingReview.md. (Shares the per-page audit with Pass C; H owns producing the prompts + tracking the coverage table.)
• MANDATE 2 — every icon/glyph is a CUSTOM Closer glyph (no generic Material icons / generic hearts). Audit all icons in use (grep -rE "Icons\.(Filled|Outlined|Rounded|Default|AutoMirrored)\."); each generic icon is a brand defect → ClaudeReport.md + a custom glyph_* to make → ClaudeBrandingReview.md (the Icon/glyph audit table). The bar for ship: zero generic Material icons — every icon is bespoke and on-brand.
• Existing art integration check: judge the art as part of the whole page, not as a standalone asset. Confirm each image supports the screen's job, aligns with the surrounding typography/actions, has enough breathing room, and uses the right light/dark treatment. Art that looks generic, unfinished, randomly placed, or visually disconnected is a finding even if the bitmap itself is technically valid.
• Soft edges (art melts into the surface): illustrations should fade/feather into the screen background, not read as a hard-edged tile/card with a crisp boundary or outline. Confirm edge treatment on both themes; a hard tile edge is a finding (C-ART-EDGE-001). Generated art should carry transparent/feathered edges (no baked-in rounded-rect block); if rendered, the shared helper should fade the edges to the surface. Record the desired edge treatment in each prompt.
• First, lock the house style (do this once per round, refresh if the art evolved): read docs/brand/visual-identity.md
  • docs/brand/asset-system.md AND open 2–3 existing illustrations (illustration_couple_onboarding, illustration_reveal_celebration, pack_art_*) to capture the actual look. New screens/features since the last brand review must be folded in. Keep the canonical house-style prompt prefix + palette in the branding deliverable (ClaudeBrandingReview.md) so every prompt reuses it and all generated art matches the existing artwork.
• House style (must hold for every prompt): flat 2D pastel vector illustration; soft rounded shapes, no harsh outlines, gentle gradients; palette aubergine #24122F / deep purple #56306F / lavender #B98AF4 / soft pink #F7C8E4 / soft lavender #D9B8FF / blush white #FFF8FC; motifs = two-equal-halves heart, paired/sealed cards, floating hearts + petals, candle/mug/lavender-sprig warmth, moon/quiet-hours, calendar/date-card, capsule; mood = warm, quiet, equal, intentional. Couple figures balanced + inclusive, faces simple. Never show readable answer/ prompt/message text, invite codes, emails, dating-app clichés, stock photos, alarm/urgency/surveillance imagery.
• Per screen, decide the brand opportunity (pick the lightest that fits — don't over-decorate):
  • none needed (already on-brand, or a dense list/form where art would clutter) — say so;
  • color/typographic brand touch (palette, heart mark, a rotating privacy message);
  • small glyph (brand glyph for a relationship concept — describe it for the glyph set);
  • hero/empty-state/celebration illustration (the high-value case → write the full ChatGPT prompt).
• Each artwork item records: screen/route · placement (hero / empty / header / card / celebration) · why it helps · filename to match the existing scheme (illustration_*, pack_art_*, glyph_*, particle_*) · the ChatGPT prompt (house-style prefix + the specific scene) · aspect ratio/size + light/dark behavior. Cross-check the brand doc's "Needed additions" / empty-state list and mark which already have assets vs still need art (e.g. Android may still lack illustrations that iOS has).
• Prioritize the screens a user feels most: onboarding/pairing, Home, paywall/subscription, reveal/celebration, empty states (no messages/dates/capsules/history), Memory Lane, Connection Challenges, date match, quiet-hours.
• Branding defects (mis-colored, clipped, off-brand, low-contrast art) are bugs → ClaudeReport.md. Pure "works but could be warmer / a feature idea" → Future.md ## QA. New art to create → ClaudeBrandingReview.md.

Pass I — Performance & route efficiency (jank, redundant reads, caching) [Future.md P14]

Before store polish, profile every top route and every high-cardinality list for jank, repeated Firestore reads, missing cache use, and slow navigation. Drive each route as a user and instrument reads/frames.

• Frame / jank: scroll every long list (Messages inbox + conversation, Answer History, Question Packs, Past Games, Wheel History, Bucket List, Date deck, Activity/Progress) and open every top route while watching adb shell dumpsys gfxinfo <pkg> framestats (or Perfetto / Studio Profiler) — flag dropped/janky frames, slow first frame, and Choreographer: Skipped N frames / main-thread stalls in logcat. Transitions/animations stay smooth (~60fps).
• Redundant Firestore / network reads: count listeners/gets per screen. Switching bottom tabs and returning must not refetch unchanged data; opening a screen twice must not double-read; snapshot listeners detach on leave (no leaked/stacked listeners — a 2-user realtime app accumulates these fast). Watch for N+1 reads on lists.
• Memory leaks (beyond listener leaks): add LeakCanary in the debug build (or take heap dumps) and navigate in→out of every heavy screen (conversation with media, game, image viewer, Memory Lane) repeatedly — flag retained Activities/Composables/bitmaps/Contexts. A leak that grows per navigation = bug (P2; P1 if it OOMs).
• StrictMode in debug (catch main-thread I/O + leaked closeables cheaply): enable a StrictMode thread + VM policy in the debug Application (detectDiskReads/Writes/Network, detectLeakedClosableObjects); any violation logged on a primary flow is a finding (disk/network on the main thread → jank/ANR risk).
• Caching / lazy-load: static question/category data is cached locally (Room) and not re-fetched each entry; large lists use lazy paging (LazyColumn/paging, not load-all); images cached (Coil); offline reads serve from cache.
• Latency: measure cold-start-to-interactive (splash→loader→Home) and tab/route transition latency; flag anything perceptibly slow (>~300ms).
• Deliverable: a reusable route smoke-test checklist (every top route × {load time · jank · read count}), captured as a runnable script so each round re-checks cheaply.
• Remediation when found: lazy-load/page large lists; cache local question/category data; dedupe + scope snapshot listeners; skip redundant fetches on tab switches; add skeleton/loading states (cf. Future.md P8) over blocking spinners.
• Findings: real jank/leak/redundant-read = bug → ClaudeReport.md (P2; P1 if it ANRs or leaks listeners, P0 if it drops data); "could be smoother / add skeletons" → Future.md ## QA.

Pass J — Accessibility (font scale · contrast · screen reader · targets · keyboard · reduce-motion) [Future.md P15]

Every primary flow must be usable with accessibility settings on. Enable each setting and walk the core flows (auth, onboarding, pairing, Home, a full game, daily question + reveal, Messages, Paywall, Settings) end to end. This is the deep home for a11y; the Pass C contrast/font spot-checks feed into it.

• Font scaling: adb shell settings put system font_scale 1.3 (then 1.5, 2.0) — every primary flow stays usable: no clipped/overlapping text, no cut-off or hidden buttons/actions (scroll where needed). Acceptance: all primary flows usable at increased font scale without clipped buttons or hidden actions. Restore font_scale 1.0 after.
• Screen reader (TalkBack): every interactive element has a meaningful semantics/contentDescription (icon-buttons especially: back, send, like, close, the brand-mark loader, game option cards); decorative images are silenced (clearAndSetSemantics {} / null desc); reading order is logical; no unlabeled "Button"; custom controls (spin wheel, date swipe deck, answer cards) are operable + announced; no focus traps.
• Contrast: body text + essential icons meet WCAG AA (4.5:1 body / 3:1 large) in both themes — measure, don't eyeball; re-check the known dim spots (game answer text, muted captions, the C-DS-001 area).
• Don't rely on color alone (color-blind / WCAG 1.4.1): any state conveyed by color must also carry a non-color cue (icon, label, shape, position). Audit the match/mismatch rendering (e.g. AnswerRevealScreen), status chips, selected/disabled states, and any red=bad/green=good signal — they must be distinguishable in grayscale / with a color-blindness simulation (adb shell settings put secure accessibility_display_daltonizer_enabled 1). Color-only status = bug.
• Touch targets: interactive targets ≥ 48dp (icon buttons, chips, nav, close/back, reaction buttons, swipe-deck actions). Flag anything smaller.
• Keyboard / external input: with a hardware keyboard, forms (sign-up, message, capsule, profile) tab in a sane order, IME/Enter actions work, focus is visible, no traps.
• Reduce-motion: with "Remove animations" (adb shell settings put global animator_duration_scale 0), the loader, celebration particles, reveals, splash handoff, and transitions degrade gracefully and no motion-gated content becomes unreachable (the loader/particles already honor this — verify everywhere). Restore to 1 after.
• Remediation: add semantics labels, raise touch targets, fix contrast tokens, guard motion behind the reduce-motion flag.
• Findings: missing label / clipped-at-large-font / sub-48dp / failing contrast = bug → ClaudeReport.md (P2; P1 if it blocks a primary flow for assistive-tech users); polish → Future.md ## QA.

Pass K — Billing & subscription lifecycle (the REAL money path, not the admin toggle)

Pass A tests the GATE (couple-shared unlock via an admin entitlement toggle); Pass K tests how the entitlement is actually earned, kept, and lost. This is the revenue path and it is almost entirely unexercised by the admin toggle. Read the manual's Billing section first. Needs real services (Google Play Billing sandbox + a Play license tester + RevenueCat) — emulators can't do real IAP, so run on a physical device with a sandbox account, or mark each money-path row blocked→needs-device (the admin toggle is not a substitute for these).

• Purchase, end to end: Paywall → select a plan → Play billing sheet → buy as a sandbox tester → RevenueCat → revenueCatWebhook/syncEntitlement → users/{uid}/entitlements/premium flips active → features unlock for both partners (couple-shared, Pass A) → onEntitlementChanged fires the partner push → the one-time Premium-unlock modal (PremiumUnlockOverlay) shows once for each partner.
• Restore purchases: Paywall "Restore" on a reinstall / second device / after sign-out→in → entitlement restored, no double-charge, features unlock.
• Plan switching: monthly ↔ annual upgrade / downgrade / crossgrade → correct proration + entitlement continuity.
• Trial / intro pricing (if configured); price + currency are displayed from the store, localized, never hardcoded; plan list + benefits render; offline/SDK-error paywall is friendly (A-OBS), Continue hidden until plans load.
• Cancel → expiry → RE-LOCK: cancel keeps access until period end (expiresAt); at expiry, CouplePremiumChecker reports inactive → premium features re-lock for BOTH and the premium-unlock "celebrated" flag re-arms. Test the expiresAt boundary (admin: set it just-past) — the couple-shared checker must treat a lapsed entitlement as inactive.
• Billing retry / grace period / account hold / pause (Play states) → entitlement + UI reflect the state; no hard crash, clear messaging.
• Refund / revocation: RevenueCat CANCELLATION/EXPIRATION/refund webhook → entitlement removed promptly → re-lock.
• Security (overlaps D3/D5): server-only entitlement writes (client self-grant → 403); webhook authenticity (forged/replayed RevenueCat webhook rejected); no client-trusted entitlement; receipt validated server-side.
• Error/abuse: cancel the billing sheet mid-flow, kill network mid-purchase, double-tap buy, rapid unlock taps → no false unlock, no duplicate purchase, retry recovers.
• Settings → Subscription reflects the live status; "Manage subscription" deep-links to Play.
• Done = purchase + restore + switch + cancel→expiry-relock + refund all verified on a real device (or each explicitly blocked→needs-device with the admin-toggle gate covered in Pass A).

Pass L — Messaging & chat (E2E, both clients, the whole feature)

Chat is a core couple feature with no functional home until now (Pass C covers its visuals, Pass E its chat_message push). Drive the main couple conversation AND the per-question "Discuss" threads QA↔Sam, both directions. Read the manual's E2EE model.

• Send every type, both directions: text, emoji, image (gallery + camera), voice note → arrives on the partner's device decrypted, correct attribution/timestamp/ordering, day separators.
• E2E at rest (overlaps D1): every sent item is ciphertext (enc:v1: / Tink media bytes), lastMessagePreview encrypted, decrypts only on member devices; raw-API read by a non-member = 403.
• Interactions: reactions (add / change / remove), read receipts ("Seen"), typing indicator, message ordering under rapid exchange.
• Failed send & offline: airplane mid-send → failed-message row → retry / dismiss (the 48dp controls), offline queue flushes on reconnect, no duplicate on retry (idempotency, overlaps F); double/triple-tap send guarded.
• Delete / moderation: delete a message (own / both) + deleted-message rendering; block/report a partner if such a flow exists.
• Media: gallery + camera + mic permission granted and denied → graceful; premium-gated media is couple-shared (Pass A); oversized image handled; image viewer opens/zoom/back.
• Inbox: conversation list, unread badge, decrypted last-message preview, recency sort; open a conversation from inbox and from "Discuss" and from a notification (Pass C/E) — all reach the same thread with a sane back stack.
• Foreground chat-head bubble for an incoming message while the app is open but that thread isn't on screen (MessageBubbleOverlay) → tap opens it (Pass E overlap).
• Realtime + perf (overlaps I): snapshot listener detaches on leaving the conversation; long-history scroll pages, no jank/leak. Quiet hours suppress the chat push (Pass M). Long/emoji/multiline/RTL text renders without clipping.

Pass M — Settings & account management (functional: settings PERSIST and TAKE EFFECT)

Pass C checks Settings looks right; Pass M checks each control does something, persists across relaunch, and takes real effect. Read Authentication and pairing flow.

• Appearance theme (Light / Dark / Device) → applies app-wide immediately, persists across process death + relaunch, and the decoupled-art behavior holds (Pass C).
• Notification toggles (daily reminder · partner answered · chat · streak) → toggling one OFF actually suppresses that push (verify by triggering it), ON re-enables; survives relaunch.
• Quiet hours → set a window covering "now" → partner-triggered pushes are suppressed/deferred during it and deliver outside it; the partner-action vs promotional rate-limit split holds. MUST test with the recipient BACKGROUNDED/KILLED, not just foreground (RETROSPECTIVE — M-001): a partner push carries a notification block the OS renders directly when the app isn't foreground, so any client-side QuietHoursManager.isInQuietHours check (which only runs in onMessageReceived, foreground-only) is bypassed exactly when quiet hours matters. Verdict bar: with QH on + recipient backgrounded, send a real partner action (chat/answer/game) → assert 0 notification in the shade AND the Cloud Function log says it suppressed (recipientInQuietHours); then QH off → same action delivers. Generalize: any "don't notify when X" setting (quiet hours, snooze, DND, per-type opt-out) must be enforced server-side where the push is SENT — verify the setting reaches Firestore and the sender honors it, not just the client. (Reminder: the users/{uid} update rule is a field allowlist — a newly-synced pref field is silently denied until added to it; confirm the write actually lands via an admin read, not just the UI toggle.)
• Biometric app-lock → enable → background-return / cold-start prompts for biometric; correct unlock proceeds, cancel keeps it locked, disable removes the lock. (Security-relevant: no bypass.)
• Edit profile → name, sex/gender (inclusive options), photo upload → persists, reflects on the partner's side, ciphertext/storage correct at rest.
• Relationship / unpair → unpair returns both to the unpaired state, revokes decrypt (D4), notifies the partner (partner_left), makes couple data inaccessible; re-pair works cleanly.
• Delete account → confirmation → account + couple data cascade (onUserDelete), partner unpaired + notified, re-create with the same email is a clean slate (overlaps G).
• Security → recovery-phrase reveal for both accepter and inviter (C-SEC-001), server-blind (D4); regenerate if supported; and the full new-device recovery flow — enter the phrase on a fresh install → existing history decrypts (canonical steps + failure paths in D4). Subscription → "Manage subscription" → Play (Pass K). Privacy & Terms / data export links open (the export contents are verified in Pass O).
• Analytics / funnel-event correctness (not just the leak check in D6). The app ships a real analytics tracker (core/analytics/FirebaseAnalyticsTracker.kt, wired via di/ObservabilityModule.kt); D6 only asserts no private content leaks into it — nobody verifies the events actually fire correctly, so the business funnel can silently break. Enable Firebase DebugView (adb shell setprop debug.firebase.analytics.app app.closer) and confirm the key lifecycle events fire once, at the right moment, with correct params: signup, pair, paywall_view, purchase/restore (Pass K), game_complete, daily_answer/reveal. Also confirm analytics honor any consent / opt-out (privacy): if a toggle or first-run consent exists, opting out must actually stop collection. Wrong/missing/duplicated events = bug; still no private content in any event (D6).
• Every toggle survives process death + reinstall-with-data (overlaps F).

Pass N — Daily question, reveal, check-ins & the other interactive features

⛔ CLAUDE: Run scripts/wiring-scan.sh BEFORE driving these features (review /tmp/claude-wiring-scan-<date>.md, record counts in ClaudeQACoverage.md). Every 🔴 dead-setter / 🟠 orphan-reader is a likely silent dead feature — this is the exact class that hid N-001 (Bucket List) + N-002 (Date Builder) behind innocent-looking empty states.

The non-game interactive surfaces that have no functional home (Pass B is games only). Read Daily question lifecycle.

• Daily-question loop (the core daily ritual): assignment (6 PM CST, assignDailyQuestion) → answer (each answer type) → both-answered gate (neither sees the other's answer until both submit) → mutual reveal → per-question Discuss thread (Pass L) → Answer History → streak increment + milestone celebration (streak_milestone) → reveal isRevealed retry (the onAnswerRevealed push). Verify the premium daily-question fallback (DailyQuestionResolver per-user) does not desync the couple's shared daily Q.
• Relationship check-ins / Your Progress (outcomes): baseline check-in (gated to show once), 30/60/90-day follow-ups, slider inputs persist (submitOutcomeCallable), the progress view renders patterns/milestones, scheduledOutcomesReminder fires, "No baseline yet" → check-in dialog (C-DARK-UI-002 area). Submit + Skip both work.
• Bucket List: add / check-complete / edit / delete an item; empty state; both-device sync; at rest encrypted (D1); premium state if applicable (A).
• Plan a Date / Date Builder: build a plan (shape/steps) → save → persists + the partner sees it; date plan + date_swipes ciphertext at rest (D1); submit-outcome path.
• ACTUALLY PERSIST + verify via admin read — an empty list can be a DEAD feature, not an empty one (RETROSPECTIVE — N-001/N-002). For every interactive feature, create real data through the UI and confirm it lands in Firestore (admin read) AND renders back; don't accept the empty/initial state as "works." Bucket List looked like an empty list but was fully non-functional (coupleId never set → every op silently returned); Date Builder's "Create Plan" silently no-ops (dateIdeaId never wired) and writes to a collection no screen reads. Reflex: any VM that gates on if (someId.isEmpty()) return and expects the screen to call setX(...) is suspect — grep for the setX caller; if none, it's dead. Also confirm there's a display surface for whatever a "save/create" writes (a save into an unread collection is an incomplete feature, not a working one).
• Activity / Together feed: shared activity entries render + sort, unread count, navigation in/out.
• Each feature: empty / loading / error / not-paired states, two-device realtime sync, no stuck/orphaned state.

Pass O — Release build, store readiness & pre-launch security

Everything above runs on the DEBUG build; the shippable artifact is the minified RELEASE build — test THAT. This is a pre-ship gate, not a per-round pass (run it before any store push and after build-config / dependency / keep-rule changes).

• Release/minified build (R8 + resource-shrink): build the release APK/AAB and run qa/entrypoint_smoke.sh + a representative slice of A–N on it. R8 can strip/obfuscate classes that Firebase/Firestore/Tink/RevenueCat/Gson/ kotlinx-serialization/Compose need via reflection → crashes that never appear in debug. Verify keep-rules; 0 FATAL on launch + each core flow; upload the ProGuard mapping to Crashlytics so release crashes deobfuscate.
• Signing & packaging: release signing config + upload key; build the App Bundle; install the signed AAB via bundletool / Play internal-app-sharing and smoke it; 64-bit + target-SDK compliance.
• App Check enforcement (pre-launch — currently OFF in dev per standing instruction; do NOT enable the dev project): in staging, enable enforcement on Firestore + Functions → a valid-token app works, a raw/no-token request → 403 (extends D3/D5 beyond rules-only); confirm Play Integrity on a real device vs the debug provider.
• Deep links / Android App Links: closer:// and any https App Links (assetlinks.json) open the correct screen with auth/membership re-checked (overlaps E).
• Permissions & manifest: the manifest declares only what's used; runtime prompts (POST_NOTIFICATIONS, camera, mic, Android-13+ photo picker / READ_MEDIA_IMAGES) appear and degrade gracefully when denied; allowBackup=false holds (D6); and the screenOrientation decision is explicit — today the manifest sets none, so the app rotates to an unverified landscape layout (Pass C). Either lock portrait or certify landscape; don't ship it undecided.
• Age gate / content rating / maturity (the app has adult/intimacy content — Desire Sync — and currently NO age gate). Confirm an appropriate 18+ / age-appropriate gate exists where required, the Play content/maturity rating questionnaire matches the actual content, and any IAP/intimacy content complies with store policy. A missing age gate on adult content is a store-rejection + legal risk — file it (see the app-finding note).
• Localization & formats (i18n): strings are externalized (no hardcoded user-facing text), the longest translations don't clip (overlaps C/J), RTL mirrors correctly, dates/numbers/subscription prices+currency format per locale (overlaps K). Even if English-only today, confirm there's no layout that assumes English length.
• Play Store readiness: the Data Safety form matches the actual data flows + E2E encryption; privacy-policy URL live; version code/name bumped; store listing/screenshots are the brand pass (H); min/target-SDK device matrix (Methodology) covered.
• Data-rights compliance (GDPR/CCPA — verify the CONTENTS, not just that a link opens). Pass M confirms the export / privacy links resolve; here, confirm right-to-access actually returns the user's real data (and that E2E content is handled correctly — exported decrypted to the owner, or documented as unrecoverable), and right-to-erasure (delete account, Pass M/G) genuinely cascades server-side (onUserDelete). A privacy policy that claims flows the app doesn't do (or omits ones it does) is a finding.

Pass P — Content, copy & language quality (voice, grammar, inclusivity, the question bank)

Wrong language is a BUG, not a "nice-to-have." Typos, grammar/punctuation errors, off-brand or cold/salesy voice, non-inclusive or assumptive wording, leaked placeholder/dev text, raw SDK/Firebase/RevenueCat errors shown to users, copy that doesn't match behavior, and broken/duplicate/low-quality questions are all defects → ClaudeReport.md. Only genuinely-working copy that could be warmer/clearer goes to Future.md. Read first: docs/brand/visual-identity.md (Store voice) and seed/questions/QUESTION_CONTENT_GUIDE.md (v3 — readability test, no-AI-writing, duplicate prevention, variety, fun/relationship-first/premium rules). This is a recurring pass.

• UI microcopy audit (every screen + state): read ALL visible text — titles, labels, button verbs, helper text, empty states, dialogs/confirmations, toasts, loading + error copy, and notification text — for: typos, grammar, punctuation, capitalization/casing consistency, consistent terminology (feature names, "partner"/the partner's name, "couple"), and copy that matches the actual action/state (a button says what it does; "Day N of 7" matches the real state; correct names/counts/attribution). A label that misstates its destination or effect is a bug (overlaps C's CTA check).
• No raw / placeholder / dev text ever reaches a user: no Lorem, "TODO", debug strings, untranslated keys, or raw exception/Firebase/RevenueCat error text surfaced to users (the A-OBS class) — always friendly copy.
• Brand voice & tone (against visual-identity.md Store voice): copy is warm, quiet, equal, calm, specific — a private ritual for two. Off-voice = finding: cold/clinical, salesy/hype, urgent/alarmist, guilt- or streak-shaming, competitive, surveillance-y, or "we'll FIX your relationship" promises. Scrutinize paywall, notification, and streak copy.
• Inclusive & non-assumptive language: no heteronormative or relationship-structure assumptions, no assumption about who initiates or about bodies/ability/culture; gender-neutral where the design calls for it (the de-gendering effort — seed/degender_*.py); sensitive topics (Desire Sync, intimacy) phrased with care + a consent framing, never crude or clinical. Assumptive/exclusionary wording = bug.
• Question-bank content QA (against QUESTION_CONTENT_GUIDE.md v3): spot-check questions across every category, depth, and answer type as the user SEES them (live-rendered, not just the DB) for: passes the readability test; no AI-writing tells; no duplicates / near-duplicates; sensible variety + emotional mix; answer options complete + mutually exclusive + sensible (no overlapping, joke-only, or placeholder options); the answer type fits the prompt; fun-rule / relationship-first tone; and no broken/empty/garbled/offensive/unsafe prompts (cf. seed/fix_depth5_grammar.py, seed/validate_question_variety.py). A shipped question that's broken, duplicated, off-guide, or unsafe is a bug.
• Legal / store / monetary copy accuracy: paywall benefit claims are truthful (no over-promise); subscription terms + renewal/price wording present and accurate (overlaps K); Privacy & Terms links resolve; store-voice rules hold.
• Localization correctness (overlaps O): no clipped/awkward strings from concatenation, correct pluralization, dates/ numbers/currency per locale, RTL grammar intact — even if English-only today, flag any English-length/grammar assumption.
• Method: harvest strings from res/values/strings.xml + in-code literals AND read them in context on-device (a string can be fine in isolation yet wrong/cramped/ambiguous in place). Routing: incorrect/off-voice/non-inclusive/ placeholder/inaccurate/unsafe = bug → ClaudeReport.md (P2 default; P1 if it misleads, blocks, leaks a raw error, or is offensive/unsafe; P3 for pure nits); "could be warmer/clearer" → Future.md; new copy/voice work that's out of scope → note it.

Reporting → ClaudeReport.md (living QA report)

• Header: date, build, devices, round number + run-state header.
• One section per pass (A–P), each a table: ID | Area | Screen/Route | Mode | Severity | Description | Repro | Evidence | Suggested fix | Status.
• Summary: counts by severity. Report only during passes — no fixes recorded until the fix phase.

Report hygiene — keep it CLEAN, lean, and never dangling (the report is a current-state doc, not an archive)

The report's job is to show, at a glance, what's wrong right now — not to accumulate a history of everything ever fixed. Stale fixed rows and stacked old run-states make it unreadable and hide the real signal. So:

• A Fixed row survives exactly ONE confirmation round, then it's removed. When you fix an issue, mark its row Fixed and keep it through the next re-QA round. Once that round re-verifies it, delete the row — the durable root-cause/fix detail lives in the Engineering Manual landmine (mandatory for any escaped/deep bug) + git history (the row cites the landmine ID, and the commit hash once the user commits), so nothing is lost. Don't rely on "the commit message" as the only home — you don't commit (the user does, often batched), so the manual landmine is the reliable record. Don't carry confirmed-fixed issues across multiple rounds.
  • Make the archived-ID line a usable duplicate-fix lookup, not bare IDs. When you prune a row, attach a 2–4 word tag to its archived ID (e.g. C-PW-001 dark paywall pills) so the Fix-phase Regression triage can search it by symptom. Any fix whose class could plausibly recur gets at least a one-line Engineering Manual landmine entry — not only the "escaped/deep" bugs the MANDATORY-retrospective requires — so a future regression check never lands on an ID with no description. (This is why a separate fix-history file is unnecessary: the manual landmines + this tagged archived line + git already are the fix history.)
• One run-state header, always. Keep only the current Round N | Pass X | Chunk Y | NEXT ACTION block pinned at the top. Don't stack prior rounds' headers — collapse finished rounds into at most a single one-line history entry each (e.g. R6: branding regression — 0 new), or drop them entirely once their fixes are confirmed-and-pruned.
• Open issues first; resolved issues compact. Order every pass section open (P0→P3) on top; keep a short Resolved & confirmed (archived — detail in git) line listing only the IDs of older fixed-and-verified issues (not their tables). The big per-issue tables exist only for currently-open and fixed-this-round-pending-confirm issues.
• Severity board reflects NOW. One board, current counts; Open is the number that actually matters. When Open hits 0 at every level, the report should be short — current run-state, a 0/0 board, the archived-ID line, and the operational constants (devices/accounts, standing-auth, playbook pointers). If it's long while everything is fixed, it needs pruning.

Coverage-matrix hygiene (ClaudeQACoverage.md — a current-status matrix, not a per-round changelog)

• Flip, don't stack. When a fix is confirmed, change that row's fail→id to pass and move the ID to an archived line — never leave a confirmed-fixed fail→id dangling, and never keep a contradicting "still owed" note next to a completed row.
• One status per cell, current. Each screen/feature/game/notification shows its latest status only; collapse prior rounds' narration into a single one-line round history. Keep an at-a-glance pass-status table at the top.
• Keep the resume signal sharp. What a returning session needs is what's left — surface todo/deferred/ blocked items plainly; don't bury them under superseded prose.

Extremely-easy-to-read mandate (applies to ClaudeReport.md, ClaudeQACoverage.md, and Future.md)

Optimize every QA doc for a reader who has 5 seconds to find the current state:

• Lead with the answer. Top of the file = current round + the one-line verdict (e.g. "0 open P0–P3; security clean") before any detail.
• Tables over prose for issues; short rows. Put long root-cause analysis in the Engineering Manual landmine (the durable home), not the row — the row gets a one-sentence description + repro + the landmine ID (and commit hash once the user commits).
• No walls of text. Break run-state into scannable lines; bold the few words that matter; no multi-paragraph headers. If a paragraph is longer than ~3 lines, it's probably manual/landmine material, not report material.
• Consistent shape every round so a returning reader (or a post-compaction resume) finds things in the same place.

Fix phase (only AFTER all passes of the round complete)

• Work strictly by severity: all P0 → P1 → P2 → P3.
• ⛔ Regression triage — DIFF & history-check BEFORE you write a fix (every bug, not just crashes — don't fix blind). First answer "is this NEW, or did we break/relapse something that worked?" — fixing without this risks re-fixing a known issue a different way (divergent fixes) or masking the real regression:
  1. Have we fixed this before? (duplicate-fix / regression check.) Search the fix history for the same symptom/area/ID — the canonical home is the Engineering Manual Known landmines and recent fixes (root cause + the guard that should hold it) plus ClaudeReport.md's Resolved & confirmed archived-ID line. A match ⇒ this is a REGRESSION, not a new bug: re-open under the original ID, and fix why the guard lapsed (a scanner/ test/pass-step that was supposed to catch it) — do not re-implement a fresh fix from scratch.
  2. What changed? (diff before you fix.) git log / git diff / git blame / git log -L the failing area to pin the introducing change — including OTHER agents' recent commits (this repo is co-edited by Codex / kimi / Ripley, so "what changed" is frequently not your own work; git log --since / git log <file> across authors). Read that commit's diff and fix the actual cause it introduced, not the surface symptom. ("worked before, broken now" ⇒ always bisect to the change first.)
  3. Only after you know new-vs-regression and what introduced it do you design the fix.
• One issue at a time: implement → ./gradlew :app:assembleDebug → install both → verify THAT fix live (correct device/theme) + regression smoke (launch/no-crash, send text, inbox loads, a game opens, content still ciphertext in Firestore, ./gradlew testDebugUnitTest + functions npm test still green — a fix that reds a test isn't Fixed) → flip its row to Fixed + capture the durable substance in the Engineering Manual landmine → next (the user commits per issue/cluster — never run git yourself; see Guardrails). Don't start the next until the current is verified.
• Real-path verification gate (do NOT mark Fixed without it): verify the fix through the same path the user hits, not a synthetic shortcut. A crash/launch/notification fix is only "Fixed" once reproduced-then-cleared via the REAL channel (real push tapped from the shade on an am kill'd app; real launcher cold-start) — am start/am force-stop passes don't count. For any cold-start/notification/launch fix, the gate is qa/entrypoint_smoke.sh green. (This session's miss: a routing "fix" was declared on am start evidence while the real bug was a splash crash on the FCM cold-start. Don't repeat it.)
• Couple-shared premium fix: replace direct isPremium() gates with CouplePremiumChecker.coupleHasPremium(partnerId) in every gated VM/screen (partner-entitlement read rule deployed). High regression risk — re-verify each feature in BOTH self-premium and free states.
• Re-run associated scanners after fixing. If the fix touches UI colors/surfaces (Pass C), re-run scripts/theme-scan.sh and confirm the relevant CRITICAL count dropped. If the fix touches launch/splash/notifications, re-run qa/entrypoint_smoke.sh. Update the coverage matrix with the new counts; a "Fixed" row is only valid when the scanner (and the live visual sweep) both agree.
• Gated actions (entitlement toggles, deploys) are user-authorized per occurrence.
• New issues found while fixing are logged (new ID), not silently fixed beyond scope — next re-QA round catches them.

Definition of done: a pass is done when every coverage row is pass/fail→id/not implemented→Future.md/ blocked→id; a round is done when all recurring passes (A–N + P) are done; flawless = one full round with zero open P0–P2 and Passes D + E + L + P fully clean (no open P0/P1 in I/J), every game fully played through, every notification type verified or explicitly not implemented→Future.md, chat (L) + the couple-shared premium gate (A) + settings-take-effect (M) + interactive features (N: daily-Q/reveal, outcomes, Bucket List, Date Builder work end-to-end — created data persists AND is read back, scripts/wiring-scan.sh 🔴=0) + content/language (P: no typos/ off-voice/non-inclusive copy, question bank on-guide) verified, all join-game navigation paths and all back-stack checks verified, the unit + functions test suites GREEN (./gradlew testDebugUnitTest + functions npm test), and qa/entrypoint_smoke.sh GREEN on both emulators (0 FAIL — every entry-point cold-start opens and stays). Then stop (P3s optional). Pass O (release build + store readiness) and Pass K's real-money path are pre-ship / real-device gates — they don't block a per-round "flawless" but must be GREEN before any store submission. Don't re-open a clean pass within the same round.

Re-QA loop (until flawless)

After the fix phase, re-run Passes A–N + P (regression + confirm fixes; Pass K money-path when a sandbox device is available, Pass O when prepping a release). Repeat fix → re-QA rounds until a full round yields zero P0–P2 and Passes D+E fully clean.

• Prune on confirmation (Report hygiene): the moment a re-QA round re-verifies a Fixed issue, delete its row from ClaudeReport.md (move its ID to the compact Resolved & confirmed (archived — detail in git) line) and collapse that finished round's run-state header. A fixed issue lives in the report for one confirmation round only — never let confirmed-fixed rows or old run-states accumulate. See Report hygiene under Reporting.