Standalone channel games — hosting the whole stack yourself

This is the self-hosted path: you build and run the on-chain referee (a libxayagame GSP), you run a WebSocket relay, you ship your own frontend, and you own every line of the consensus code. Nothing here is required to publish a game — if you want to ship without operating anything, read ARCADE.md instead; the hosted Arcade is the default execution of this same protocol and does every part of this file for you. Come here when you need a deployment nobody else controls, a seat model or presentation the hosted path cannot express, or you are porting an existing standalone game.

Read SKILL.md first. It carries the protocol facts this file assumes — the channel lifecycle, the strictly turn-based model, the off-chain sign prefix, dispute vs resolution, and the real-time reconciliation patterns. This file is the implementation half: the C++ referee, the browser-side machinery, consensus hardening, and the deploy. Wagering on top of a channel → WAGERING.md. Compiling your rules to WebAssembly → WASM.md.

Reference implementation, cited throughout: the pre-split standalone Xayaman (github.com/xaya/xayaman, branch main) — one game running its own GSP + relay + frontend, with the C++ referee in engine/gsp/ and engine/board/ and the browser half in src/lib/channel/, src/lib/crypto/ and src/lib/relay/. ⚠️ Archived — do not start here. It is also private during the curated phase — access on request, so the file paths below are citations telling you what the shape is, not files you can open today. Ask in Discord if you need to read it; /repos on the arcade site carries the current visibility of everything named here. It is frozen at the pre-split commit and is cited only because it is a complete, working, end-to-end standalone stack you can read; it is not a template, it is not maintained, and no new game should be forked from it. (xayaships_frontend / polyxayaships in EXAMPLES.md carry the same mark.) The copy-me template for new work is the Arcade one — see ARCADE.md.

That reference stack is strictly 2-participant, and every code sample below assumes a pair. Going wider means budgeting for N-seat turn rotation, proofs that carry every live seat's signature, and an N-way settlement policy — the protocol supports it (metadata.participants), the reference implementation does not exercise it.

Trust this file over memory for the relay auth string and the session-key storage prefix. Both are exact-match protocol details that silently break when approximated, and both are quoted below from their authority files — re-read them at the cited paths before you deploy.

1. What you own, and the two block constants

Two block windows govern the endgame. The reference GSP's compiled-in defaults are at xayaman/engine/gsp/logic.hpp: DISPUTE_BLOCKS = 10 and CHANNEL_TIMEOUT_BLOCKS = 12 (the repo prefix matters — arcade-platform carries the same two names in its own engine/gsp/logic.hpp). They are that GSP's values, not a platform constant — your GSP picks its own. Nor is a compiled-in constant the only shape: the Arcade's multi-game host GSP keeps the same two names only as fallbacks and reads a per-game window off its registry row instead (COALESCE (gr.dispute_blocks, …) / COALESCE (gr.channel_timeout_blocks, …) in arcade-platform/engine/gsp/logic.cpp, columns defaulting to 10 and 12 in arcade-platform/engine/gsp/schema.sql). The two are unrelated: DISPUTE_BLOCKS is the timer a dispute runs before the referee force-resolves the channel, while CHANNEL_TIMEOUT_BLOCKS only reaps channels that never got a second participant.

2. The GSP referee (C++) — create / join / dispute / resolve

The consensus half is a libxayagame GSP built on the xaya::ChannelGame base: the referee that watches on-chain channel moves and force-closes disputes. Everything below is verified against the reference GSP (engine/gsp/, engine/board/).

Your logic class (engine/gsp/logic.hpp → YourLogic : public xaya::ChannelGame):

class YourLogic : public xaya::ChannelGame {
  YourBoardRules boardRules;
  const xaya::BoardRules& GetBoardRules() const override { return boardRules; }
  void SetupSchema(xaya::SQLiteDatabase& db) override;       // SetupGameChannelsSchema(db) + your tables
  void GetInitialStateBlock(unsigned& h, std::string& hash) const override;  // per-chain genesis (POLYGON case)
  void InitialiseState(xaya::SQLiteDatabase& db) override;   // empty (or the wager bootstrap — WAGERING.md)
  void UpdateState(xaya::SQLiteDatabase& db, const Json::Value& blockData) override;  // the dispatch, below
  Json::Value GetStateAsJson(const xaya::SQLiteDatabase& db) override;  // channels + stats → RPC JSON
};

ChannelGame provides ProcessDispute() / ProcessResolution(), which call VerifyStateProof() for you — replaying every transition through your board's ApplyMove and checking every signature. You never verify a proof by hand. main.cpp wiring is the same xaya::SQLiteMain(config, gameId, rules) as a non-channel GSP, with two channel extras: rules is your ChannelGame, and the RPC server needs the channel instance factory — xaya::ChannelGspInstanceFactory instanceFact(rules); config.InstanceFactory = &instanceFact; (engine/gsp/main.cpp; the class is gamechannel/gsprpc.hpp in libxayagame).

The move dispatch (UpdateState) — each on-chain move is one key; dispatch on it, ignore unknown or multi-key moves:

for (const auto& mv : blockData["moves"]) {
  const std::string name = mv["name"].asString();
  const auto id  = GetIdFromMove(mv);   // for `c`, the channel id = this move's txid/mvid
  const auto& d  = mv["move"];
  if (!d.isObject()) continue;          // not an object → not a move of ours
  if (d.size() > 1) continue;           // exactly one action per move
  HandleCreateChannel     (db, d["c"], height, name, id);
  HandleJoinChannel       (db, d["j"], name, id);
  HandleAbortChannel      (db, d["a"], name);
  HandleDeclareLoss       (db, d["l"], name);
  HandleDisputeResolution (db, d["d"], height, /*isDispute=*/true);
  HandleDisputeResolution (db, d["r"], height, /*isDispute=*/false);
}
// wager admin commands arrive SEPARATELY in blockData["admin"] — see WAGERING.md
ProcessExpiredDisputes(db, height);
TimeOutChannels(db, height);

On-chain move payloads — every channel move rides the normal envelope (XayaAccounts.move('p', playerName, '{"g":{"<gameid>":{ <one key> }}}', …)):

The joiner learns channelIdHex by reading the open-channels list out of the GSP's getcurrentstate (the same feed the channel manager bootstraps from, §5) — the create move's id is the channel id.

reinit (what it is). A channel's reinit is the id/state that scopes its signature domain for the current participant set. On create the channel is Reinitialise(meta, ""); on join, UpdateMetadataReinit(id, meta) bumps it (participants went 1→2) and the channel is Reinitialised with the initial board state. A StateProof's initialState is a reinit state; reinitId tags which reinit an off-chain message belongs to, so a proof from a superseded participant set is rejected. Off-chain code threads reinitId through (processOffChain(reinitId, proof), §5); on-chain, declare-loss carries the base64 reinit so the GSP confirms you are conceding the current game.

Board rules (engine/board/) — the bridge from your game logic to the protocol:

// libxayagame's proto-board template takes BOTH proto types (state AND move);
// alias it once, as the reference implementation does (engine/board/board.hpp):
using BaseProtoBoardState =
    xaya::ProtoBoardState<proto::BoardState, proto::BoardMove>;

class YourBoardState : public BaseProtoBoardState {
  bool IsValid() const override;                                   // ★ runs on UNTRUSTED input — §9
  bool ApplyMoveProto(const proto::BoardMove& mv,
                      proto::BoardState& newState) const override;  // your rules; false on any bad input
  int  WhoseTurn() const override;                                 // participant index, or NO_TURN when finished
  unsigned TurnCount() const override;                             // THE freshness counter — must be monotonic (PITFALLS.md #18)
  Json::Value ToJson() const override;                            // the shape your board-state converter reads (SKILL.md)
};

class YourBoardRules : public xaya::ProtoBoardRules<YourBoardState> {  // ONE template param
  // REQUIRED — pure virtual in BoardRules; return ORIGINAL:
  xaya::ChannelProtoVersion GetProtoVersion(
      const xaya::proto::ChannelMetadata& meta) const override;
  // Optional: override ParseState too if the empty pre-join state ("") needs
  // special handling (the reference implementation does — engine/board/board.hpp).
};

// The starting position is a plain free function your join/start handler
// serializes into the channel's reinit state (NOT a rules-class member):
proto::BoardState InitialBoardState(const proto::Config& config);

ParseState wraps IsValid, and ProcessDispute/ProcessResolution call it on attacker-supplied bytes before signatures are checked — so IsValid / WhoseTurn / ApplyMoveProto must never CHECK/abort on bad input. This is the most important consensus rule in the whole stack: read §9 before you write them.

Protobuf schemas — reuse the channel protos from libxayagame unchanged (gamechannel/proto/{metadata,signatures,stateproof,broadcast}.proto). Author your game's own proto/{boardstate,boardmove}.proto (+ optional gameconfig). Generate C++ with protoc (GSP side) and TypeScript with buf (target=ts, frontend side).

Building the channel GSP — identical to a plain libxayagame GSP Dockerfile except: also compile the gamechannel protos; link the channel libs -lgamechannel -lchannelcore -lxayagame -lxayautil -lethutils (a non-channel GSP links just -lxayagame -lxayautil); source layout board/ + gsp/ (logic/main/schema/gamestatejson) + proto/. The plain-GSP build recipe it extends is not in this skill — one sentence of it: a libxayagame GSP is a C++ daemon built against the libxayagame image, run with --xaya_rpc_url pointed at XayaX and its own --datadir; load the building-persistent-games skill for the full section.

Dispute lifecycle + timeouts. ProcessExpiredDisputes force-resolves a channel once the DISPUTE_BLOCKS timer elapses; TimeOutChannels reaps channels that never got a second participant. Neither may compute the deadline with an unsigned subtraction — on a low-height chain (a fresh fork or regtest) height - DISPUTE_BLOCKS underflows and every dispute looks already expired. Write the comparison in sum form (disputeHeight + DISPUTE_BLOCKS <= height): it avoids the subtraction entirely, which is why the Arcade's host GSP now uses it in both passes — its own comment there says the sum form "needs no manual underflow guard, unlike the old single-window (height - window) subtraction" (arcade-platform/engine/gsp/logic.cpp, ProcessExpiredDisputes and TimeOutChannels). A if (height < DISPUTE_BLOCKS) return; guard in front of a subtraction also works, but only for as long as nobody forgets it.

And do not assume a timeout simply awards the win to the disputer — that is one policy, not the protocol. The Arcade's host GSP asks the game's own blob (arcade_resolve_timeout) what a timeout means at every seat count and settles on its answer, falling back to the survivor close only when that answer is unusable; above two seats a non-finished answer becomes a reinit rather than a close, and the survivors keep playing. ARCADE.md §6 has that contract. A GSP you write yourself owns this decision entirely — decide it deliberately. Either way a dispute-timeout close never sets the board's finished flag; that UI trap, and what you may and may not infer from it, are in SKILL.md §5.

3. Session keys

Session keys let every off-chain move be signed instantly with no wallet popup. MetaMask (or equivalent) is used only for on-chain transactions: create, join, dispute, resolution.

Storage prefix (the key builder in src/lib/crypto/session-key.ts): xayaman_session_<channelIdHex>, with a lazy one-time migration from a legacy prefix left over from a pre-rename build — run once, best-effort, never blocking key access if localStorage is unavailable:

const STORAGE_PREFIX = 'xayaman_session_';
const LEGACY_STORAGE_PREFIX = 'racer_session_';

export function getSessionKey(channelIdHex: string): `0x${string}` {
  if (typeof window === 'undefined') return generatePrivateKey();
  migrateLegacySessionKeys();          // copies any racer_session_* forward, once
  const key = `${STORAGE_PREFIX}${channelIdHex}`;
  const existing = localStorage.getItem(key);
  if (existing) return existing as `0x${string}`;
  const pk = generatePrivateKey();
  localStorage.setItem(key, pk);
  return pk;
}

If you fork this pattern for a new game, pick your own <yourgame>_session_ prefix and drop the migration shim — it exists only because that game was renamed mid-project from an earlier codename. Prefix collisions between two games served from one origin are a real hazard, not a theoretical one.

The session key's address is what goes on chain, in the "addr" field of the create/join move. The GSP checks off-chain signatures against that address, never against the player's main wallet.

4. WebSocket relay auth

All Xaya channel games can share one relay codebase — github.com/xaya/xaya-relay, listed with its current visibility on the arcade's /repos index — but each game runs its own relay instance pointed at its own GSP (single GSP_URL; same Docker image, different env). Deploy shape in §10.

Auth message (verified exact format, src/lib/relay/ws-broadcast.ts — RELAY_AUTH_AUDIENCE for the audience constant and the authMessage template literal that follows it, client side; src/auth.ts getAuthMessage server side. Cited by symbol, not line: this file is the one that moves most, and the two repos that carry it disagree on line numbers):

Xaya Relay Auth
Audience: <aud>
Channel: <channelIdHex>
Timestamp: <unixSeconds>

const RELAY_AUTH_AUDIENCE =
  (typeof process !== 'undefined' && process.env.NEXT_PUBLIC_RELAY_AUDIENCE) ||
  'xayaships-relay-v1';

const channelIdHex = channelId.toString(16).padStart(64, '0');
const timestamp = Math.floor(Date.now() / 1000);
const authMessage =
  `Xaya Relay Auth\nAudience: ${RELAY_AUTH_AUDIENCE}\nChannel: ${channelIdHex}\nTimestamp: ${timestamp}`;
const signature = await sessionSigner.signMessage(authMessage); // Xaya-chain-prefixed — SKILL.md §3

The audience identifies the relay deployment, not the game. Its job is cross-relay replay protection: it binds a signature to this relay so one captured here cannot be replayed against another relay sharing the same channelId format. The relay does exact single-string equality (src/auth.ts hard-rejects a mismatched audience before spending CPU on signature recovery), so one value per deployment and every client connecting to it sends that same value — two clients with different audiences cannot both be served by one relay. Because each game runs its own relay instance, in practice that means one audience per game; xayaships-relay-v1 is only the client-side fallback baked into the snippet above, carried over from the relay's original game. The server has no default: with auth enabled and RELAY_AUTH_AUDIENCE unset it exits at startup (src/index.ts). Set it on the relay and set NEXT_PUBLIC_RELAY_AUDIENCE on the client to the identical string.

The server recovers the signer from the (chain-prefixed) auth message, checks it matches the claimed address, checks timestamp freshness against its AUTH_TIMEOUT window (src/auth.ts, env-tunable — read your deployment's value), and enforces single use: an accepted (address, channel, signature) triple is remembered for the freshness window and a replay of it inside that window is rejected, so a captured signature cannot re-trigger a seat takeover.

⚠️ The auth message must match your relay BUILD byte-for-byte. The current relay implements the Audience: line and treats it as mandatory: with auth enabled and no RELAY_AUTH_AUDIENCE set the server refuses to start, and there is no optional or dual-dialect mode. Relay builds predating that change sign Xaya Relay Auth\nChannel: <hex>\nTimestamp: <ts> with no audience line. Before wiring auth, open your relay's src/auth.ts and make the client sign exactly the string the server recovers; a one-line mismatch rejects every join, with no partial-credit failure mode.

Server → client message types: subscribed, gsp_state, gsp_status, joined (with participants), peer_joined, peer_left, replay (a header before N buffered messages), and error. Binary frames carry the actual StateProof / BroadcastMessage payloads; JSON frames are control messages — dispatch on typeof event.data === 'string'.

The relay caps addresses per channel at MAX_PARTICIPANTS (src/channel-router.ts), an env-tunable default that is no longer pinned at two — read it from your deployment rather than assuming.

It also runs a liveness sweep, and that is where a silent drop's peer_left comes from. A link that dies silently — a phone losing signal, a laptop suspended, a killed process — never fires its own close event, so a socket with nothing to send can stay nominally open for minutes after the network under it is gone. The relay therefore pings every registered connection on a fixed interval and terminate()s any that did not answer the previous ping (HEARTBEAT_MS, src/limits.ts, env-tunable — read your deployment's value). A terminate fires the close event the dead link never sent, so the drop takes the ordinary path: leaveChannel, then peer_left to the rest of the channel. That means the detection window is between one and two intervals, and it is what starts the game's own disconnect grace — set the interval well below the stall timeout your client offers a dispute on, or the game's timeout gets there first and the sweep buys you nothing. Browsers answer a ping automatically, so nothing is required of the client.

The relay is a dumb binary forwarder: no game logic, no matchmaking, no tick loop, no timestamps of its own. It receives bytes from one peer and immediately ws.send()s them to the others. That is exactly what SKILL.md's timing-security argument rests on.

5. Channel manager — the off-chain state machine

One instance per browser session (src/lib/channel/channel-manager.ts). On the hosted path @xayaarcade/sdk implements this for you and you never write it; standalone, it is yours:

processOnChain(blockHash, height, metadata, reinitState, proof, disputeHeight)
  → bootstraps/updates from GSP on-chain state
  → triggers auto-moves if it's our turn

processOffChain(reinitId, stateProof)
  → processes the opponent's off-chain message
  → validates reinit match + strictly-increasing turnCount before accepting
  → triggers auto-moves in response

processLocalMove(moveBytes)
  → applies the player's own move (or bot's), signs, broadcasts via relay

Critical init order: call processOnChain() before attaching to the relay. The relay's replay buffer (§4) hands you buffered messages immediately on join — if the channel manager is not bootstrapped yet, those are silently dropped and the match starts a state behind.

Who sends the closing transaction. When the game ends the winner sends the resolution — never the loser, who has no incentive to spend gas declaring their own loss. Auto-detect it here:

// After auto-moves settle, if the game is DECIDED and we won.
// isFinished() is a separate question from whoseTurn() === NO_TURN, and the
// difference matters here: every finished state answers NO_TURN, but a NO_TURN
// state that is NOT finished is a real, reachable shape. Settle on the former;
// use NO_TURN only to choose resolution-vs-dispute (the GSP rejects a dispute
// naming no on-clock seat). `winner` is the board's own typed query, in the raw
// ABI convention: -1 undecided, -2 draw, else the seat index.
const finished = boardState?.isFinished();
const winner = boardState?.winner();
if (finished && winner === ourIdx) {
  txHash = await this.moveSender.sendResolution(this.currentProof);
} else if (opponentIsUnresponsive) {
  txHash = await this.moveSender.sendDispute(this.currentProof);
}

One exception: a draw has no winner, so nobody's win path fires — the lowest-index participant sends the closing resolution instead (src/lib/channel/channel-manager.ts), otherwise the channel hangs open forever. For a wagered game that draw close is also what triggers settlement — see WAGERING.md.

Duplicate-resolution prevention. A flag stops triggerAutoMoves from firing a second resolution tx when a pending-transaction check clears too early. It is cleared in exactly ONE place — when the reinit changes, i.e. a new game on the same channel re-arms it. Clearing it anywhere else re-opens the duplicate-tx bug; never clearing it blocks the next game's resolution and locks funds:

private resolutionSent = false;
// in processOnChain, on reinit change — a fresh game re-arms the close:
if (reinitChanged) this.resolutionSent = false;
// in triggerAutoMoves / the win path:
if (finished && winner === ourIdx && !this.resolutionSent) {
  const txHash = await this.moveSender.sendResolution(this.currentProof);
  if (txHash) this.resolutionSent = true;
}

6. Auto-move re-entrancy

Signing is async, so a naive auto-move loop can re-enter itself while the previous signature is still pending. Guard with a simple lock, and chain inside the lock hold — needed whenever a game's protocol can run multiple auto-moves back-to-back (bot turns, multi-step reveals):

private autoMoveInProgress = false;

async triggerAutoMoves(): Promise<void> {
  if (this.autoMoveInProgress) return;
  this.autoMoveInProgress = true;
  try {
    let madeMove = true;
    while (madeMove) {                 // chain auto-moves inside ONE lock hold
      madeMove = false;
      const move = await this.channel.maybeAutoMove(this.boardState);
      if (move) {
        await this.applyAndBroadcast(move);
        madeMove = true;               // loop: check for a chained auto-move
      }
    }
  } finally {
    this.autoMoveInProgress = false;
  }
}

Do not chain by recursing into triggerAutoMoves() from inside the lock: the re-entrant call sees the lock held and returns immediately, so chained auto-moves silently never run. The loop form above is what the reference implementation ships (src/lib/channel/channel-manager.ts). A strictly single-input-per-turn game rarely chains, but the guard costs nothing and prevents a subtle re-entrancy bug the moment any auto-move path is added later.

7. Opponent disconnect + resync

The relay emits peer_joined / peer_left (§4) on WebSocket connect/disconnect. Wire these to the UI and to auto-resync:

broadcast.setPeerCallbacks(
  () => { setOpponentDisconnected(false); channelManager.resendCurrentState(); },
  () => { setOpponentDisconnected(true); },
);

resendCurrentState() re-broadcasts the current proof on rejoin — the reconnecting peer's processOffChain() accepts it via the existing monotonic move-count check, so no new protocol message is needed:

resendCurrentState(): void {
  if (!this.currentProof || !this.offChainSender) return;
  this.offChainSender.sendNewState(this.reinitId, this.currentProof);
}

Edge cases, all handled by that same monotonic check: both peers reconnect simultaneously (higher count wins), a peer reconnects before the opponent advanced (the resend is not newer, silently dropped), rapid repeated reconnects (idempotent).

8. Proof optimization and client-side verification

Trimming the proof. A naive StateProof accumulates every transition since channel open — for a long match that is a lot of protobuf messages and an unbounded resolution transaction. optimizeProof trims to the minimal suffix that still carries every participant's signature somewhere in the proof (src/lib/channel/state-proof.ts, optimizeProof):

1. Walk backward from the last transition, tracking which participant signed each newState.
2. Stop as soon as every participant has signed at least one state in the walked suffix.
3. Promote that boundary transition's newState to become the new initialState.
4. The remaining (kept) transitions prove the current state from that signed checkpoint forward.

BEFORE (full history):
  initialState (reinit, unsigned)
  → t0 (P0 signed) → t1 (P1 signed) → ... → tN (P1 signed)

AFTER optimizeProof (2 transitions, regardless of N):
  initialState = t[N-2].newState   (already P1-signed)
  → t[N-1] (P0 signed)
  → tN     (P1 signed)             ← current state

Valid because the GSP's proof verification accepts an initialState that is not the reinit state, as long as every participant has signed somewhere in the proof — the trimmed suffix still satisfies that. If the walk cannot find every participant's signature, it returns the proof untrimmed rather than shipping an unprovable one. Call it after every extendStateProof(); the reference implementation's extendStateProof() does this automatically before returning. For a two-player game it always trims to 2–3 transitions no matter how long the match ran.

Verifying a received proof. Both clients verify every proof received from the peer before applying it — never trust the wire:

1. Walk every transition in the proof.
2. For each, re-execute the move through the (same) WASM applyMove() and compare the resulting state hash against the claimed newState.
3. Verify each signature via ecrecover, with the Xaya chain sign prefix (SKILL.md §3).
4. Drop invalid proofs entirely — never partially apply. Surface the event as a divergence indicator in the UI rather than silently ignoring it: a divergence that only shows up as a stalled board is unreportable, and it is your first signal that native and WASM disagree.

A cheater cannot forge state this way: their WASM must produce the identical result as yours, on identical inputs, or the proof is rejected locally before it ever reaches the chain.

9. Untrusted state is a pre-auth consensus DoS — the three-layer IsValid ★

Before real money — or any adversarial use — rides on a channel game, treat the GSP as a consensus system: every node must compute the same state from the same moves, and must never crash on attacker-crafted input.

This is the single most dangerous surface in the stack. During dispute/resolution the GSP verifies a StateProof by parsing an attacker-supplied board state and calling IsValid() before any signature is checked. So IsValid runs on fully attacker-controlled bytes. If a malformed state passes IsValid and later reaches a CHECK()/abort(), the GSP crashes — and because every node crashes identically on the same on-chain move, that is a consensus-fatal, pre-auth denial of service: one cheap move halts the game chain.

Rules (engine/board/board.cpp — a three-layer defence, not one function):

1. IsValid() rejects every malformed state by returning false — never lets later code reach a CHECK/abort.
   • Compare proto integer fields unsigned (they are uint32): a value at or above the signed bound must be rejected, not silently sign-flipped negative past a signed comparison.
   • Bound every value used as an array index (cell indices → per-cell arrays; owner indices → the players vector).
   • Guard near-maximum counters against overflow.
   • Cap the total count of attacker-sized collections (bombs, fires, entities, …) so proof-iteration cost is bounded.
2. WhoseTurn() degrades an out-of-range or unvalidated turn to NO_TURN (the move is rejected), never CHECK/abort.
3. ApplyMove() guards return false on bad input, never CHECK/abort — it too runs on adversarial pre-signature states.

Gate it. A board-validity test feeds hand-crafted malicious states through your ParseState (which wraps IsValid) and asserts each is rejected. In the reference implementation that test is baked into the GSP Docker image build (docker/Dockerfile.xayamand compiles and runs engine/test/board_validity_test.cpp as a build step), so a failure fails the build and it gates deploys, not just npm test. Annotate each malicious case with the specific downstream CHECK it defuses, so a future edit knows what a bound is protecting.

A rules blob on the hosted path has the same problem in a different shape — parse-vs-validate, bound every count, and the fuel/memory/output caps. That version is in ARCADE.md; do not port this C++ advice to a blob unread.

10. The four-service dockerized deploy

Four services on one compose network. Templates ship with this skill — scripts/docker-compose.template.yml and scripts/env.template — and are the copy-paste authority for exact flags, ports and build args:

Channel-shaped gotchas, in the order they bite:

• One audience string, two consumers. Feed the relay's RELAY_AUTH_AUDIENCE and the frontend's NEXT_PUBLIC_RELAY_AUDIENCE build arg from the same compose variable, or client and server drift and every join is rejected (§4).
• Declare the build arg. Docker silently ignores an undeclared --build-arg, so your Dockerfile.frontend must ARG (and ENV) every NEXT_PUBLIC_* the compose file passes. Miss NEXT_PUBLIC_RELAY_AUDIENCE and the client quietly falls back to its built-in default audience — a working build that cannot join a channel.
• NEXT_PUBLIC_* are inlined at build time, so changing one means docker compose up -d --build, not a restart. The relay URL is the exception: derive it at runtime from window.location and the frontend stays portable across hosts.
• The GSP's --game_id is the move namespace and must match the frontend's NEXT_PUBLIC_GAME_ID and, if you wager, the wager contract's game namespace. Three places, one value.
• Bind the GSP and the relay to loopback; expose only the frontend. The relay is an unauthenticated fan-out to anyone who can reach its port until auth is on, and the GSP RPC has no auth at all.
• Wagering needs a fresh GSP datadir or the queue bootstrap silently skips — see WAGERING.md.

Two rules here are general Docker/GSP operations rather than channel-specific: pin the compose network's subnet explicitly (never let Docker auto-pick on a shared host) and give a fresh GSP datadir a Polygon start height just below the current chain head or it never catches up — load the building-persistent-games skill for the full deploy and day-2 operations section.

Verified against github.com/xaya/xayaman@main (engine/gsp/logic.hpp, engine/gsp/logic.cpp, engine/gsp/main.cpp, engine/board/board.cpp, docker/Dockerfile.xayamand, src/lib/channel/, src/lib/crypto/session-key.ts, src/lib/relay/ws-broadcast.ts) and github.com/xaya/xaya-relay (src/auth.ts, src/index.ts, src/channel-router.ts).