scrabble-game/PRERELEASE.md

# Pre-release plan — hardening before Stage 18

Living tracker for the pre-release hardening pass that runs **before Stage 18** (the
prod cutover). Same discipline as [`PLAN.md`](PLAN.md): one phase per session,
**interview the owner on the open details** at the start of each phase, bake every
decision back into `PLAN.md` / `docs/` / the affected `README`s / Go Doc comments in
the **same** PR, get CI green, then mark the phase done. Phases run as
`feature/* → development` PRs (the Stage 16 branch model); the owner approves+merges.

**Why now:** the system is feature-complete through Stage 17 and the test contour is
green, but there is **no prod data yet** — schema, wire labels and the dictionary
layout can still change for free. These phases spend that one-time freedom and harden
the edge before prod. Each phase maps back to the owner's raw pre-release TODO list
(numbers in the tracker).

## Phase tracker

| # | Phase | Raw TODOs | Status |
|---|-------|-----------|--------|
| R1 | Schema & naming reset | 1 + 10 | **done** |
| R2 | Stress harness + contour observability + early run | 9a | **done** |
| R3 | Edge hardening | 2 + 8 + 3 | **done** |
| R4 | Push enrichment + kill the last poll | 4 + 5 | **done** |
| R5 | Bundle slimming | 6 | **done** |
| R6 | Refactor + docs reconciliation + de-staging | 7 | **done** |
| R7 | Final stress run + tuning | 9b | **done** |
| UI | Tab-bar navigation redesign (drop the hamburger) | owner ad-hoc | **done** |
| MW | "Multiple words per turn" rule for Russian games (engine v1.1.0) | owner ad-hoc | **done** |
| MW2 | Single-word rule connectivity fix: the word must run along its own line through an existing tile (perpendicular-only contact no longer connects); single-tile direction picks the best legal word (engine v1.1.1) | owner ad-hoc | **done** |
| MW3 | Graceful replay degradation: a game whose journalled move became illegal under MW2 is closed as a draw (`end_reason='aborted'`) on open instead of erroring, with an impersonal organizer note in the history + GCG (migration `00002`) | owner ad-hoc | **done** |
| OW | Open auto-match: enter the game at once and wait inside it (robot after 90–180 s) | owner ad-hoc | **done** |
| DA | Dictionary admin: online release-archive upload → word-diff preview → install/activate; versioned dict volume; active version persisted in DB; resident label = release tag | owner ad-hoc | **done** |
| AB | Manual account block (admin suspension): permanent/temporary with an editable en+ru reason picklist; a block forfeits the player's active games + cancels their open ones; a backend gate refuses a blocked account with **403 `account_blocked`**; the UI shows a terminal blocked screen and stops all push/poll; manual unblock; temporary blocks self-expire (migration `00003`) | owner ad-hoc | **done** |
| AI | Honest AI opponent in quick game: an explicit 🤖 AI / 👤 random selector (AI default); the robot is seated and moves at once; 7-day inactivity loss (the per-turn timeout reused); chat/nudge disabled, no statistics; the opponent is shown as 🤖 everywhere | owner ad-hoc | **done** |
| → | Stage 18 — prod contour deploy | — | see [`PLAN.md`](PLAN.md) |

## Key findings (these reshaped the raw list — read before starting a phase)

- **R1 (TODO 1 + 10) is one cheap moment, now.** Squashing the 12 goose migrations is
  safe precisely because there is no prod data and the contour DB is wiped. Folding the
  new variant labels (`scrabble_ru`/`scrabble_en`/`erudit_ru`) into that single baseline
  makes the rename need **no data migration and no back-compat mapping**. Today's labels
  (`english`/`russian_scrabble`/`erudit`) are persisted in `games.variant`,
  `game_invitations.variant`, in `pkg/fbs` and the UI — ~100 files, but a mechanical sweep
  on a clean DB.
- **R4 (TODO 4 + 5): the app is already push-first.** Game state refreshes on
  `your_turn`/`opponent_moved`, the lobby on `notify`, chat on `chat_message`. The **only**
  genuine periodic server poll is `lobby.poll` (matchmaking, 2.5 s,
  `ui/src/screens/NewGame.svelte`). What remains is killing that one poll **and** enriching
  push events to carry payloads so the UI stops re-fetching after each signal.
- **R3 (TODO 2): identity forgery is already mitigated.** Identity is always derived from
  the session (`Authorization: Bearer` → `X-User-ID`); the client cannot inject identity,
  the backend re-validates resource ownership, Telegram initData is HMAC-checked. The real
  gaps are a missing **request-body size limit** (cheap DoS) and **invisible rate-limit
  rejections** (no log/metric/admin view — that is TODO 8). Static landing serving is **not**
  covered by the gateway token bucket (it only guards `Execute`).
- **R6 (TODO 7) scale:** ~431 `Stage N` references across ~104 files (incl. the file name
  `backend/internal/inttest/stage6_test.go`). Code is the source of truth; `docs/` describe
  current state; `PLAN.md` keeps the decision history.

## Locked decisions (owner interview)

- **Stress test (TODO 9):** **early + final** runs. Driver = **edge protocol** (Connect/FB
  through the gateway, moves generated by the solver) **plus a separate gateway-hammer**
  saturation test. Pacing = **realistic (under limits) + saturation (ramp to the knee)**.
  Resource metrics = **add cAdvisor + postgres_exporter to the contour** (today only
  Go-runtime metrics exist). The harness stays in the repo for repeats.
- **Push (TODO 4 + 5):** **both** — kill `lobby.poll` (use the existing `match_found`, keep
  poll as the ws-down fallback) **and** enrich push events with payloads.
- **Refactor (TODO 7):** **hygiene + structural changes by a reviewed list** —
  behaviour-preserving, test-gated, contentious items surfaced to the owner before applying.
- **Landing (TODO 3):** **separate static container** behind the project caddy
  (`/` → landing, `/app/` + `/telegram/` → gateway); drop `landing.html` from the gateway
  `go:embed`.
- **Rate-abuse (TODO 8):** metric + Grafana + admin view **plus a conservative auto-flag** —
  a *soft, reversible* "suspected high-rate" marker for operator review, tunable threshold,
  **no auto-ban**.
- **Open auto-match (owner ad-hoc):** a quick game **enters a real game at once and waits inside
  it** (status `open`, the opponent seat empty); a second human searching the same variant+rule
  joins it, or a robot fills it after a **90 s + random 0–90 s** wait, pushing the in-app
  **opponent_joined** event. While open, the starter may move on their turn but resign, chat and
  nudge are disabled, and the lobby + opponent card read "searching for opponent". Matchmaking is
  now **DB-backed open games** — the in-memory pool, `lobby.poll` and `lobby.cancel` are gone. The
  schema is edited in the baseline (no prod data); `game_players.account_id` is nullable for the
  empty seat.

## Phases

Each phase: read this tracker + the relevant `docs/`, **interview the owner on the open
details below**, implement within scope, then update the tracker + docs/code and get CI
green before marking it done.

### R1 — Schema & naming reset *(TODO 1 + 10)* — first
Squash `backend/internal/postgres/migrations/00001..00012` into one `00001_baseline.sql`
(method: `pg_dump --schema-only` from a fully-migrated DB → wrap as the goose baseline →
prove a fresh migrate yields a schema identical to the 12-migration chain via the
integration suite → delete the old files; keep goose). Bake the new variant labels into the
baseline. Propagate `scrabble_ru`/`scrabble_en`/`erudit_ru` through the backend
(`engine.Variant`/`ParseVariant`, `registry.dictFiles`, the CHECK values), the wire
(`pkg/fbs` `variant:string`, regenerate FB) and the UI (`lib/model.ts` union, `variants.ts`,
fixtures, premium/alphabet keys, tests); i18n display keys stay display-only. Tidy
`../scrabble-dictionary` to a single source→dawg build point and align the dawg artifact
names to the new labels (crosses into `../scrabble-solver`'s committed fixtures — keep them
byte-identical). After merge, **wipe the contour DB** (drop the volume) so it re-provisions
on the next deploy.
- Critical files: `backend/internal/postgres/migrations/`,
  `backend/internal/engine/{engine,registry}.go`, `pkg/fbs/scrabble.fbs`,
  `ui/src/lib/{model,variants}.ts`, `../scrabble-dictionary/{Makefile,cmd/builddict,…}`.
- Open details to interview: the exact dawg filename scheme; whether the dict-repo tidy is
  one PR or split; how to script the contour DB wipe in the deploy.

### R2 — Stress harness + contour observability + early run *(TODO 9, part 1)*
Build the reusable load harness as a new `loadtest` module in `go.work` (reuses `pkg/fbs`,
`connect-go`, and `scrabble-solver` for legal-move generation): a seeder that inserts
**1000 guest + 10000 durable** accounts with pre-created sessions (token hashes) directly in
the DB and hands the plaintext tokens to the client; a driver that runs N virtual users,
each in 3–5 concurrent 2–4-player games, exercising submit-play / pass / exchange / nudge /
chat / check-word / draft-move / profile-save through the **edge protocol**, in
**realistic** (under rate limits) and **saturation** (ramp) modes; plus a separate
**gateway-hammer** that deliberately exceeds limits to verify the limiter holds and measure
its cost. Add **cAdvisor + postgres_exporter** to `deploy/docker-compose.yml` and a Grafana
resource dashboard. Run the **early pass** against the freshly-wiped contour; produce a
**trip report** (logic/concurrency bugs + a resource baseline) that feeds R3 and R6.
- Critical files: new `loadtest/`, `deploy/docker-compose.yml`, `deploy/observability/*`,
  `docs/TESTING.md`.
- Open details: the scale ramp steps; the move-selection policy (a mid-ranked solver move
  for realistic game progress); run duration; the pass/fail bar.

### R3 — Edge hardening *(TODO 2 + 8 + 3)*
Add a **request-body size cap** at the gateway h2c mux / `Execute` (e.g. ~1 MB). Add
**rate-limit observability**: a `gateway_rate_limited_total{class}` counter + a structured
log per rejection; an **aggregate** Grafana panel (request rate + rejection rate — spikes
visible without per-user label cardinality, honouring the Stage 12/17 discipline); an
**admin-console view** of recently throttled users/IPs (in-memory ring buffer, single-
instance, reset-on-restart, like the `active_users` gauge). Add the **conservative
auto-flag**: when a user is *sustained*-throttled past a tunable threshold, set a soft,
reversible `account.flagged_high_rate_at` marker (baked into the R1 baseline) surfaced in the
admin user list/detail — **no auto-ban**; the operator clears it. Split the **landing** into
its own static container (`deploy/` + a Caddyfile route `/` → landing) and drop
`landing.html` from the gateway `go:embed`.
- Critical files: `gateway/internal/connectsrv/server.go`, `gateway/internal/ratelimit/`,
  `gateway/internal/connectsrv/metrics.go`, `backend/internal/adminconsole/`,
  `deploy/caddy/Caddyfile`, `deploy/docker-compose.yml`, `gateway/internal/webui/`.
- Open details: the auto-flag threshold/window + whether the marker is persisted vs
  in-memory; the landing image base (caddy vs nginx).

### R4 — Push enrichment + kill the last poll *(TODO 4 + 5)*
Replace `lobby.poll` with the existing `match_found` push (keep the poll as a ws-down
fallback). Enrich `your_turn`/`opponent_moved`/`notify` to carry the state payload so the UI
renders from the event without a follow-up `game.state` (removes the lobby↔game nav latency
the owner noticed). Wire-contract change: `pkg/fbs` event payloads → backend `notify` emit →
UI stream consumers (`ui/src/lib/app.svelte.ts`), with the per-game cache as the landing
spot; regenerate FB.
- Critical files: `pkg/fbs/scrabble.fbs`, `backend/internal/notify/events.go`,
  `ui/src/lib/{app.svelte,transport}.ts`, `ui/src/screens/NewGame.svelte`.
- Open details: which events carry full vs delta payloads; the fallback-poll cadence when the
  stream is down.

### R5 — Bundle slimming *(TODO 6)* — done
Analysed the bundle against the 100 KB-gzip budget; **no code slimming was warranted**, and the
budget metric was retargeted to measure the app correctly. The build already minifies +
tree-shakes; the dominant cost is the Connect/FlatBuffers transport runtime + generated bindings
+ the Svelte runtime (≈⅔ of `main`'s source is third-party/generated) — irreducible within scope.
**Lazy-loading was rejected**: `bundle-size.mjs` sums every emitted chunk, so code-splitting yields
no total-size win and adds request latency (+N gateway fetches on first navigation to a split
screen). i18n lazy-load was skipped (the catalogs are a sliver of a Svelte-runtime-dominated shared
chunk, and `en` must stay bundled as the `MessageKey` type source + fallback). Instead,
`bundle-size.mjs` now measures **per HTML entry**, with three independent gates on the natural chunk
boundaries — **app entry ≤ 100 KB, the Svelte+i18n shared chunk ≤ 30 KB, the landing's own chunk
≤ 5 KB** — since the app's real payload is its entry chunk plus the shared chunk (≈97 KB), while the
landing (≈24 KB) is reported separately and kept minimal. Same CLI + exit-code contract, so the CI
step is unchanged.
- Critical files: `ui/scripts/bundle-size.mjs`; no app code changed.

### R6 — Refactor + docs reconciliation + de-staging *(TODO 7)* — done
Behaviour-preserving only. Three separable, separately-committed passes: (a) mechanical
**de-staging** — remove `Stage N`/`TODO-N` references from code, comments and service
READMEs (rename `stage6_test.go`); (b) **docs↔code reconciliation** — reconcile
`docs/ARCHITECTURE.md` / `docs/FUNCTIONAL.md`(+`_ru`) against the code-as-truth, fixing drift
and Go Doc comments; (c) **structural changes by a reviewed list** — surface a list of
proposed optimizations / test-suite consolidations to the owner, apply only the approved,
behaviour-preserving, test-gated ones. The full suite + the final stress run (R7) are the
regression gate. Incorporates the early-run (R2) bug fixes not already shipped.
- Open details: the structural-changes list itself (owner-approved before applying); the test
  consolidation targets.

### R7 — Final stress run + tuning *(TODO 9, part 2)* — done
Re-run the R2 harness against the final, refactored system on a clean contour; analyse
resource consumption across **all** components (gateway, backend, Postgres, the
metrics/observability stack, docker log volume) and agree the tuning (pool sizes, rate
limits, cache TTLs, container limits, GOMAXPROCS, log levels). Apply the agreed tuning; record
the methodology + results in the repo.

→ **Stage 18** (prod contour) then proceeds per [`PLAN.md`](PLAN.md).

## Sequencing rationale

`R1` first (cheapest now; everything builds on the final schema/naming and the stress test
must run against it). `R2` builds the harness and runs the **early** pass to surface bugs and
a resource baseline that feed `R3` and `R6`. `R3`/`R4`/`R5` harden and improve the system.
`R6` (de-stage + reconcile + structural) runs near the end so it sweeps settled code once and
benefits from all accumulated bug knowledge. `R7` validates the final system and tunes it.
Then Stage 18.

## Regression-safety discipline (cross-cutting)

- Every phase is a `feature/* → development` PR; CI (`unit` + `integration` + `ui` behind the
  `CI / gate` check) must be green before the owner merges; watch the post-merge contour
  deploy with `gitea-ci-watch.py`.
- `R6` structural changes are behaviour-preserving, test-gated, and split from the mechanical
  sweeps; contentious items are owner-approved first.
- The two stress runs (`R2` early, `R7` final) are the system-level regression gate.

## Verification (per phase)

- `go build ./<module>/...`, `go vet`, `gofmt -l .` clean, `go test -count=1 ./<module>/...`;
  UI: `pnpm check && pnpm test:unit && pnpm build`; the integration suite
  (`-tags integration`) for DB/schema changes; `docker compose config` for deploy changes;
  green CI on the PR + a healthy contour deploy.
- `R1`: prove the squashed baseline yields a schema identical to the 12-migration chain
  (integration suite on a fresh DB) **before** deleting the old files.
- `R2`/`R7`: the harness runs end-to-end against the contour; the trip report lists concrete
  defects + a resource profile from the Grafana cAdvisor/postgres_exporter panels.

## Refinements logged during implementation

- **R1** (interview + implementation):
  - **Variant labels** `english`/`russian_scrabble`/`erudit` → **`scrabble_en`/`scrabble_ru`/`erudit_ru`**
    across the backend (`engine.Variant.String`/`ParseVariant`; the `games`/`game_invitations` `variant`
    CHECK in the baseline; GCG `#lexicon` and the `variant` metric attribute both flow from `String`),
    the wire (`pkg/fbs` `variant` is a `string` field — values change with **no FlatBuffers regen**) and
    the UI (`model.ts` union, `variants.ts` records, `codec`/`premiums`/mocks/tests, the admin
    `dictionary.gohtml`). **Kept:** the Go enum identifiers (`VariantEnglish`…, internal) and the i18n
    display keys (`new.english`/`new.russian`/`new.erudit`, display-only). `complaints.variant` stays
    free-text (no CHECK, as before).
  - **dawg filenames kept descriptive** (`en_sowpods`/`ru_scrabble`/`ru_erudit`) — only the registry's
    `Variant` key carries the rename, so `registry.go`, the published `scrabble-solver` fixtures and the
    dictionary release artifact are untouched (decouples the three repos).
  - **Migrations squashed** 12 → one hand-written `00001_baseline.sql`. Verified by a
    `pg_dump --schema-only` diff (the chain vs the baseline are **identical** but for the two intended
    variant-CHECK values) plus the green integration suite. **No data migration** (no production data).
  - **Done (cross-repo + contour):** the **`scrabble-dictionary` tidy** merged (PR #2) and was re-cut as
    the **byte-identical `v1.0.1`** release for clean provenance (the backend stays on `v1.0.0` — same
    bytes, no rewire; the backend pulls a version-pinned release artifact, not master). Post-merge the
    contour `backend` schema was wiped (`DROP SCHEMA backend CASCADE` + restart, not a volume drop) and
    re-migrated to the baseline — verified the new variant CHECK (`scrabble_en/scrabble_ru/erudit_ru`),
    `games`=0 and a clean boot.

- **R2** (interview + implementation):
  - **Locked decisions:** game assembly via **invitations** (real path, no robots; not direct game-row
    inserts); **moderate** ramp **50 → 200 → 500** at 10 min/step; **diagnostic** pass bar (no SLO gate);
    run as a **one-shot container on `scrabble-internal`** in this PR.
  - **Harness** = new `scrabble/loadtest` module (`use ./loadtest` + a `replace scrabble/gateway` for the
    dot-free edge-proto import). It seeds 1000 guest + 10000 durable accounts + sessions **directly in
    Postgres** (token hash mirrors `backend/internal/session`), drives players over the **edge protocol**,
    generates **mid-ranked legal moves locally** with the embedded `scrabble-solver` by replaying
    `game.history` (the edge carries no board — mirrors `engine.ReplayBoard` via the public API), and a
    **gateway-hammer**. Compact CLI (`run` / `cleanup`), distroless Dockerfile (DAWGs baked), Go unit tests.
  - **Adding the module broke the other images' builds** — backend/gateway/telegram Dockerfiles reduce the
    workspace but still referenced `./loadtest` (not in their context); each now also
    `-dropuse=./loadtest` (backend/telegram additionally `-dropreplace` the gateway replace). Caught by the
    first deploy run; verified by building all four images.
  - **Harness payload fixes found by the smoke pass:** the draft DTO's `rack_order` is a string (was sent
    as `[]` → `bad_request`); the display-name validator forbids digits/colons, so the cleanup marker
    became a letters-only `Zzloadtest` so `profile.update` resends the seeded name. `chat_not_your_turn` /
    `nudge_own_turn` are **by-design** turn gates, correctly exercised.
  - **Observability:** added **cAdvisor + postgres_exporter** + the **Scrabble — Resources** dashboard +
    two Prometheus jobs. **Finding:** cAdvisor yields only the root cgroup on the contour host (separate
    XFS `/var/lib/docker` breaks its layer-ID resolution — the existing galaxy deploy has the same limit),
    so per-container CPU/RSS for the early pass was captured via `docker stats`. **R7:** adopt the otelcol
    `docker_stats` receiver (already the contrib image) for per-container metrics in Grafana.
  - **Early run (2026-06-09):** ramped clean to 500 players, no crash/deadlock, cleanup removed all 11000
    accounts. 1.2 M edge calls, 48 870 plays, 2 798 games finished; the per-user limiter held under the
    hammer (99.97 % rejected, p99 2 ms). **Top finding:** ~14 % `transport_error` on `game.state` at 500
    players, under CPU saturation (backend/gateway/Postgres each ~1 core) and amplified by the harness's
    single shared `http2.Transport`; the harness itself peaked at 86 % of a core on the same host, so the
    figures are pessimistic. Full trip report in [`../loadtest/REPORT-R2.md`](../loadtest/REPORT-R2.md);
    it feeds R3 (h2c `MaxConcurrentStreams`/timeouts, body-size cap), R6 and R7 (per-player transports,
    separate hardware, pool/limit sizing).
  - **CI:** `./loadtest/...` added to the path filter + vet/build/test; `go.work.sum` carries the new deps.

- **R3** (interview + implementation):
  - **Locked decisions:** the flag column lands by **editing the R1 baseline** (+ a contour schema
    wipe after merge — no migration chain accrues before prod); auto-flag defaults **1000 rejected /
    10 min** (`BACKEND_HIGHRATE_FLAG_THRESHOLD`/`_WINDOW`, rolling window, set-once, operator clears,
    no auto-ban); landing image = **caddy:2-alpine**; throttle data flows **gateway → backend** (a
    30 s per-key summary POST to the new `/api/v1/internal/ratelimit/report`, the existing trusted
    direction) with the episode window + flag rule in the backend (`internal/ratewatch`); rejection
    logging = **Warn summary per key per window + Debug per rejection** — a deliberate deviation from
    the phase's "structured log per rejection" (the R2 hammer would have logged ~522k lines in
    minutes); all three R2-report tails included (explicit h2c sizing, the session-resolve failure
    cause at Warn, reviving the admin limiter).
  - **Body cap:** `GATEWAY_MAX_BODY_BYTES` (default 1 MiB) as both the Connect per-message read limit
    and an `http.MaxBytesReader` wrap of the public mux; an oversized Execute is `resource_exhausted`.
  - **Dead config found:** `AdminPerMinute`/`AdminBurst` were never wired — the gateway `/_gm` mount is
    now 429-guarded per IP ahead of its Basic-Auth. The caddy-fronted contour path stays unlimited
    (stock caddy has no limiter) — an accepted gap, recorded in `docs/ARCHITECTURE.md` §12.
  - **Landing split:** a `landing` target in `gateway/Dockerfile` (the UI build stage is shared;
    identical compose build args keep it one cached build); the gateway drops `landing.html` from the
    embed and 308-redirects `/` → `/app/`; the contour caddy routes `/app/`, `/telegram/` and the
    Connect path to the gateway and the catch-all to the landing container; the CI deploy probe now
    checks both `/` (landing) and `/app/` (gateway).
  - **Observability:** `gateway_rate_limited_total{class}` (user/public/email/admin, aggregate-only)
    + a rate-vs-rejections panel on the Edge/UX dashboard; the admin console gains the **Throttled**
    page (the in-memory episode window, reset-on-restart like `active_users`, plus the flagged-account
    queue) and the flag badge / clear action on the user list / card.
  - The jet regen also restored the previously missing `game_drafts`/`game_hidden` generated models
    (their tables were added after the last jetgen run; no behaviour change).

- **R4** (interview + implementation):
  - **Locked decisions:** **delta-first**, not full snapshots — an event carries only the new move and
    the UI applies it to its per-game cache, keyed on `move_count` (idempotent + gap-safe: a gap or the
    actor's own move falls back to a `game.state` + `game.history` refetch). `match_found` /
    `game_started` carry the recipient's **initial `StateView`** (instant lobby→game); the fallback
    refetch stays the existing two calls (no merged endpoint); the matchmaking poll runs **only while
    the stream is down** (2.5 s); **all** UI-state-changing events carry their payload (incl. lobby `notify`).
  - **Enriched events** (`pkg/fbs` trailing fields — backward-compatible, no FB regen of *values*, only
    the schema): `opponent_moved` (+`move`/`game`/`bag_len`), `your_turn` (+`move_count`), `match_found`
    (+`state`), `game_over` (+`game`), `notify` (+`account`/`invitation`/`state`). The pre-R4
    `opponent_moved` scalars (`seat`/`action`/`score`/`total`) stay for wire back-compat, now redundant
    with `move`/`game` — slated for the R6 de-stage.
  - **Encoding placement:** the `notify` package keeps ownership of the FlatBuffers encoding (a new
    `encode.go` mirrors the gateway transcode but reads wire-agnostic `notify.*` input structs +
    `engine.MoveRecord`); the game/lobby/social services map their domain types to those structs, so the
    wire schema stays out of the domain. **Flagged for R6:** this partly duplicates the gateway encoders
    (different source types) — a candidate consolidation.
  - **Actor self-fetch killed too** (beyond literal "push"): the `submit_play`/`pass`/`exchange`/`resign`
    **response** (`MoveResult`) now returns the actor's refilled rack + bag size, so the mover renders the
    next turn from the response — `Game.svelte`'s `commit`/`pass`/`exchange`/`resign` drop their `await load()`.
  - **`match_found` enrichment** needs a per-seat initial state: `lobby.GameCreator` gained `InitialState`,
    and `game.Service.InitialState` builds the `notify.PlayerState` (rack re-encoded to wire indices, the
    variant alphabet embedded for a first-seen variant).
  - **UI:** a pure `lib/gamedelta.ts` reducer (`applyMoveDelta` / `applyGameOver` / `seedInitialState`,
    unit-tested) advances the cache; `app.svelte` seeds it on `match_found` / `game_started`; `Game.svelte`
    applies the delta (falling back to `load()` while composing, on a gap, or on its own move's new rack);
    `NewGame.svelte` polls only when `app.streamAlive` is false and guards its teardown so a push-delivered
    match is not cancelled.
  - **notify (friends/invitations) scope:** the backend carries the full account / invitation payload on the
    wire (per "all events → push"); the UI seeds the game cache from `game_started` but keeps its lightweight
    **authoritative** badge refresh (`refreshNotifications`, on the rare `notify` event + on foreground) rather
    than adding client-side friend/invitation caches — the per-move hot path is fully de-fetched, which was the
    goal. Deeper lobby-cache consumption is an easy follow-up.
  - **No schema change** (no migration); the contour needs no DB wipe. Tests: `notify` FB round-trips +
    `emitMove` delta + the `gamedelta` reducer; the e2e mock now emits the enriched delta.

- **R5** (interview + implementation):
  - **No code slimming — by analysis.** A gzip measure + sourcemap attribution of the real `dist` showed
    the app bundle is already minified + tree-shaken and dominated by the Connect/FlatBuffers transport
    runtime + generated FB/PB bindings (≈⅔ of `main`'s source) and the Svelte runtime — all
    third-party/generated, irreducible within R5's scope. App-authored code carries no hand-trimmable fat.
  - **Lazy-load rejected** (screens *and* i18n): `bundle-size.mjs` sums every emitted chunk, so
    code-splitting moves bytes between chunks for **zero total-size win** while adding request latency (+N
    gateway fetches on first navigation to a split screen). i18n lazy-load additionally buys ≤3 KB (en-only
    users) at the cost of an async `t()`, and `en` must stay bundled (it is the `MessageKey` type source +
    fallback). **Chunk-collapsing rejected** too — keeping the near-static Svelte runtime in its own
    cacheable chunk is the recommended practice (an app deploy then re-busts only `main`, not the runtime),
    and HTTP/2 makes the extra preload request negligible.
  - **Metric retargeted to the app.** The two-entry build (`index.html` app + `landing.html`) makes Rollup
    hoist the code shared by both (Svelte runtime + i18n + `aboutContent`) into one preloaded chunk, so the
    app actually loads its entry chunk **+ the shared chunk** (≈74 + ≈23 = **≈97 KB**), never `landing.js`
    (≈1.6 KB). The old script summed all three chunks (98.8 KB), over-counting the app by `landing.js`.
    `bundle-size.mjs` now parses each built HTML for the JS it eagerly loads and gates three parts
    independently — **app entry ≤ 100 KB, shared (Svelte+i18n) ≤ 30 KB, landing-own ≤ 5 KB** — reporting the
    app total (≈97) and landing total (≈24.5). Same CLI + exit-code contract, so the CI step is unchanged.
  - **No app/source/build change** (`App.svelte`, `lib/i18n/`, `vite.config.ts` untouched); no schema
    change, no contour wipe. The stale "~82 KB" figure was corrected in `bundle-size.mjs` and `ui/README.md`.

- **R6** (interview + implementation):
  - **Locked decisions:** apply **both** wire/code structural changes (**B** + **A**) and **only C1+C2** of
    the test consolidation (not C3/C5); strip the `*(Stage N)*` tags from **all current-state docs**
    (ARCHITECTURE / FUNCTIONAL+`_ru` / TESTING / UI_DESIGN), keeping PLAN.md / PRERELEASE.md / CLAUDE.md as
    history; **split `stage6_test.go`** by domain. The `h2cMaxConcurrentStreams` sizing stays an **R7**
    concern (tuning, not behaviour-preserving); the R2 early run forced no code fix, so nothing was carried in.
  - **(a) De-staging:** removed the `Stage N` / `TODO-N` / `(RN)` references across code, comments, service
    READMEs and the current-state docs, rewording narratives to present tense (no technical content lost).
    Renamed the only stage-named identifiers (`registerStage8`→`registerSocialOps`,
    `registerStage11`→`registerLinkOps`) and split `stage6_test.go` (`TestEmailLoginFlow`→`email_test.go`;
    `TestGuestAutoMatchLeavesNoStats`+`provisionGuest`→`account_test.go`). De-staged the `.fbs`/`.proto`
    comments and regenerated: only the `.proto`-derived Go docstrings (`*_grpc.pb.go`, `push.pb.go`) changed —
    flatc strips schema comments, so the FB Go/TS bindings were untouched.
  - **(b) Reconciliation:** the docs were accurate (each R-phase baked its own); the one drift was a stale
    "guest-reaping deferred (TODO-3)" note in `ARCHITECTURE.md` §3 — guest reaping is implemented, so the
    note was replaced with the current behaviour (FUNCTIONAL/TESTING already described it).
  - **(c) B — dead `opponent_moved` scalars:** removed `seat/action/score/total` from `OpponentMovedEvent`
    (`pkg/fbs/scrabble.fbs` + the `notify` emit + the round-trip test); regenerated FB Go + TS. No reader
    used them (the UI codec/mock take `move`/`game`/`bag_len`; the gateway forwards the payload verbatim).
    A pre-release wire-slot renumber — free with no prod data, no DB change.
  - **(c) A — shared FB builders:** new `scrabble/pkg/wire` holds the single definition of the nested wire
    tables (GameView / MoveRecord / StateView / AccountRef / Invitation) shared by the backend `notify`
    encoder and the gateway `transcode`; both map their own source types to neutral `wire.*` structs and
    delegate. **Honest tradeoff:** the verbose `Start/Add/End` + reverse-prepend boilerplate is now written
    once, but the field *set* is still mapped per side, and the new package makes the change net **+~145 LOC**
    — a single-source / anti-drift win for the fiddly mechanics rather than a line-count cut. Behaviour-
    preserving: the two sides' field sets were verified identical and the round-trip tests pass unchanged.
  - **(c) C1+C2 — inttest fixtures:** moved the cross-file service/game fixtures (`newGameService` was used by
    10 files) into `backend/internal/inttest/helpers.go`; single-file helpers stay local. Pure relocation.
  - **No schema change → no contour DB wipe.** Regression gate: the full unit + integration + UI suites plus
    the R7 stress run.

- **R7** (interview + implementation):
  - **Locked decisions:** run the harness **same-host** (one-shot container on `scrabble-internal`, capped
    `--cpus=3` so the contour keeps spare cores); **apply container limits + `GOMAXPROCS` now** (not just a
    prod recommendation); **replace cAdvisor with the otelcol `docker_stats` receiver** (it resolved only the
    root cgroup on this host); keep rate-limit / h2c knobs **compiled-in** (change values only if the data
    demands — it did not).
  - **Harness refinements (pre-run):** each virtual player builds its **own `edge.Client`** (its own h2c
    connection for its Subscribe stream + Execute calls) instead of all players sharing one `http2.Transport` —
    the R2 `transport_error` artifact; and `playTurn` now reports a **finished** game so the player drops it
    from rotation. Effect, measured: `game.state` `transport_error` 14 % (R2) → **2.49 %**; `game_finished` on
    chat ≈ 3 900 → **35**.
  - **Observability:** added the `docker_stats` receiver to `otelcol` (`api_version: "1.44"` — the daemon's
    minimum is 1.40; the receiver defaults to 1.25 and crash-looped until pinned), mounted the docker socket
    read-only with `group_add` (the contrib image runs as UID 10001), dropped the cAdvisor service + its
    Prometheus job, and retargeted the **Scrabble — Resources** dashboard to the docker_stats metric names
    (`container_cpu_utilization`/100 == cores). Cross-checked against `docker stats` within sampling error.
  - **Profile (final run, 500 players, limits in force):** the **gateway is the binding constraint** — with
    one connection per player it bursts into its 2-core cap (the residual 2.49 % `transport_error`); backend
    ~0.85 core and postgres ~1.4 cores had headroom; **tempo reached its 1 GiB cap**; the backend pool sat at
    its `MaxOpenConns=25` cap (28 backends); docker logs were unbounded (~14 MiB / 30 min on the backend at
    info). Full write-up in [`../loadtest/REPORT-R7.md`](../loadtest/REPORT-R7.md).
  - **Round-2 tuning (owner-agreed, all in `deploy/docker-compose.yml`, no code change):** gateway **2 → 3
    cores + `GOMAXPROCS=3`**; tempo memory **1 → 2 GiB**; backend `MAX_OPEN_CONNS` **25 → 40**; a json-file
    **log-rotation** default (10m × 3) applied contour-wide via a YAML anchor (level stays info).
    backend/postgres kept at 2 cores / 512 MiB (headroom is cheap on the shared host).
  - **Validation:** the same gradual ramp on the tuned contour cut `game.state` `transport_error` to **0.72 %**
    (gateway ~2 cores, now under the 3-core cap, no throttle; tempo ~1.27 GiB, under 2 GiB). A separate
    **burst** run (a single 100 → 500 jump) pegged the gateway at 3 cores (≈296 % sustained, 9.27 % error),
    confirming it is **connection-CPU-bound** — a true arrival spike is a **horizontal-scaling** lever, not
    more cores per node (recorded in the prod-sizing recommendation).
  - **No schema change → no contour DB wipe.** Bake-back: `loadtest/REPORT-R7.md` (new), `loadtest/README.md`,
    `docs/TESTING.md`, the telemetry/observability section of `docs/ARCHITECTURE.md`, the repo-layout line in `CLAUDE.md`.

- **UI — Tab-bar navigation redesign** (owner ad-hoc, not on the raw TODO list): drop the hamburger
  `Menu.svelte` everywhere (it fought the Telegram-fullscreen layout, where it had to be re-centred).
  - **Locked decisions (interview):** the in-Settings sub-nav is a **bottom TabBar with the active tab
    highlighted** (icon-only); **Export GCG** moves to the left slot of the move-history header (free in a
    finished game, where 🏁 *leave* does not apply); the lobby **⚙️ badge counts incoming friend requests
    only** (invitations keep their own lobby section); unread chat is badged on **the score bar and the 💬**.
  - **What shipped:** a ⚙️ **Settings hub** (`screens/SettingsHub.svelte`) over the existing
    Settings/Profile/Friends/About bodies and an in-game **comms hub** (`game/CommsHub.svelte`) over
    chat + dictionary, both with in-place tabs and a fixed back target; the game's menu items relocate into
    the open move history (🏁 leave / 📤 export + 💬 comms header) and the player cards (🤝 add-friend); a
    shared **TapConfirm** (`components/TapConfirm.svelte`, `lib/tapconfirm.ts`) — tap → fading ✅ → tap —
    replaces the Skip/Hint press-and-hold popovers and drives the add-friend confirm. Fixed the move-history
    "jump" bug (the slid board is now inert and the stage can't scroll, so a swipe up genuinely closes it).
    `Menu.svelte` + `HoldConfirm.svelte` removed.
  - **No schema/wire change → no contour DB wipe.** Bake-back: `docs/UI_DESIGN.md`, `docs/FUNCTIONAL.md`
    (+`_ru`). Regression gate: UI `check` + unit (`tapconfirm`) + build + bundle budget + e2e (Chromium &
    WebKit), all green.

- **UI — Merge Exchange/Pass; drop the dead Tournaments tab** (owner ad-hoc, not on the raw TODO
  list): the lobby's 🏆 *Tournaments* tab was an inert `lobby.soon` toast — removed (the lobby is back
  to three tabs, matching `docs/FUNCTIONAL.md`). In-game the separate 🥺 *Skip* (pass) tab folds into
  the 🔄 tab, now **Exchange/Pass**, whose dialog passes when no tile is selected and exchanges when
  tiles are.
  - **Decision — a pass is NOT an exchange of zero (verified against the rules + GCG):** the merge is
    **UI-only**. Pass and exchange stay distinct game actions end-to-end — wire (`GameActionRequest` vs
    `ExchangeRequest`), engine (`ActionPass` vs `ActionExchange`), and the GCG Poslfit dialect (a pass is
    a bare `-`, an exchange is `-TILES`). The engine forbids a zero-tile exchange (`ErrNothingToExchange`)
    and allows an exchange only with a full rack left in the bag (`ErrNotEnoughTilesToExchange`), while a
    pass is always legal — collapsing them would lose a real distinction. The dialog dispatches the
    existing `gateway.pass` / `gateway.exchange`.
  - **What shipped:** `Lobby.svelte` (tab removed); `Game.svelte` (one 🔄 Exchange/Pass tab no longer
    gated on an empty bag; the dialog disables tile selection while the bag is below a full rack
    (`bagLen >= RACK_SIZE`), its confirm button reading **Pass without exchanging** / **Exchange N**);
    i18n (`game.draw` → Exchange/Pass, new `game.passNoExchange`, dropped `game.skip` /
    `lobby.tournaments` / `lobby.soon`). No backend/wire/history/GCG change.
  - **No schema/wire change → no contour DB wipe.** Bake-back: `docs/UI_DESIGN.md`, `docs/FUNCTIONAL.md`
    (+`_ru`). Regression gate: UI `check` + unit + build + bundle budget + e2e (Chromium & WebKit).

- **AI — Honest AI opponent in quick game** (owner ad-hoc, not on the raw TODO list): a second quick-game
  opponent the player *knowingly* chooses, distinct from the disguised robot of the random/open path
  (which is kept as-is). New Game's quick-game mode replaces the "auto-match" subtitle with a two-button
  selector **🤖 AI / 👤 Random player** (the `.seg`/`.opt` segmented style, AI the default); for AI the
  move-clock line reads "Loss after 7 days of inactivity" and the "searching" hint is hidden.
  - **Locked decisions (interview):** AI move is **event-driven** (the robot replies the instant the
    player's move commits; the 30 s driver is the fallback); AI games **do not touch `account_stats`**
    (practice, like guests); the **Stage 5 strength logic is reused unchanged** (`playToWin` 40 % from the
    seed + margin band); **no per-move timeout — a 7-day inactivity loss** instead; the 7-day line lives on
    the New Game screen (the in-game screen has no move-clock line); chat + nudge **disabled**, word-check
    kept, add-friend never drawn, opponent shown as **🤖** everywhere.
  - **The 7-day rule reuses the existing per-turn timeout:** an AI game is created with
    `turn_timeout_secs = AIInactivityTimeout` (7 days) and the existing timeout sweeper resigns the overdue
    seat — since the robot moves at once, only the human is ever on the clock, so the per-turn timeout *is*
    the abandon rule (no new column, no new sweeper).
  - **One game flag drives everything:** `games.vs_ai` (edited into the R1 baseline — pre-release, so a
    contour DB wipe after merge). It is set **only** on AI-started games, so a robot-filled random game keeps
    `vs_ai=false` and the disguised opponent is never revealed; the UI derives 🤖 / the gates **from the flag,
    never from the opponent account**. New backend path `Matchmaker.StartVsAI` (picks a pooled robot via the
    existing `Pick`, creates an **active** seated game via `game.Service.Create`, random seat order) — the AI
    request never enters the open pool, so the open-game reaper never touches it. The robot driver gains a
    `vs_ai` branch (no sleep, no proactive nudge, zero delay) and a focused `DriveGame`/`TriggerMove` fast
    path wired from the game service's after-create/after-commit hook (`SetAITrigger`, a func value so the
    game package never imports the robot package). Chat/nudge gated by a new `social` `VsAI` check
    (`ErrGameVsAI` → 409 `ai_game`); statistics skipped in `commit` when `vs_ai`.
  - **Wire:** `EnqueueRequest` += `vs_ai`, `GameView` += `vs_ai` (trailing FB fields, regenerated Go + TS),
    threaded through the backend DTO, the gateway transcode and the `pkg/wire` + `notify` builders.
  - **Tests:** `lobby` unit (StartVsAI seats a robot + flags the game; empty pool leaves no game); backend
    integration (`ai_game_test.go`: active+seated+vs_ai+7-day clock, robot moves immediately, stats skipped,
    7-day timeout resigns the human, chat/nudge rejected); UI codec round-trip (`vs_ai` on enqueue + game
    view); e2e (an AI game shows 🤖, no "searching", chat disabled, the dictionary still works) + the
    existing quick-match e2e updated to pick **Random player** (the default is now AI).
  - **Schema/wire change → a contour DB wipe** after merge (`DROP SCHEMA backend CASCADE` + restart, the
    R1/R3 pattern). Bake-back: `docs/ARCHITECTURE.md`, `docs/FUNCTIONAL.md` (+`_ru`), `docs/UI_DESIGN.md`,
    `backend/README.md`, Go Doc comments.
  - **Post-review refinements (owner, same PR):** (1) the **GCG export labels the robot seat "AI"** rather
    than its human-like pool name (`ExportGCG` overrides the name via `accounts.IsRobot`; the in-app 🤖 is
    unchanged); (2) honest-AI games **emit no `your_turn`** — the robot replies instantly, so the signal
    would arrive with the move and be pointless; `opponent_moved` still advances the UI; (3) the **admin
    console surfaces the AI flag** — a **🤖 column** in `/games` and an "AI game" line on the game card
    (`GameRow`/`GameDetailView` gain `VsAI`); (4) `games_started_total` / `games_abandoned_total` gain a
    **`vs_ai`** attribute and the Grafana *Game domain* dashboard splits started/abandoned into **human**
    and **AI** panels.
  - **Follow-up (separate PR — strategy deviation):** the robot now plays **≈20%** of opening/midgame moves
    *against* its per-game `playToWin` intent (toward the opposite margin band — a winning robot eases off, a
    losing one surges ahead), tapering linearly to **0 over the last 14 bag tiles** and **0 once the bag is
    empty**, so the endgame follows the chosen strategy strictly while earlier outcomes can swing the human's
    way. Deterministic from the seed (`mix(seed,"deviate",moveCount)`), applied to **both** robot paths via
    the shared `selectMove`; the per-game intent (and the admin card) is unchanged. Tests: `robot` unit
    (taper bounds + monotonicity, never-in-endgame, determinism, ~20% distribution). Bake-back:
    `docs/ARCHITECTURE.md` §7, `docs/FUNCTIONAL.md` (+`_ru`), `backend/README.md`, `PLAN.md` Stage 5.