Adversarial multi-agent review of the launch candidate produced 22 confirmed
findings (12 more were refuted). This closes the critical, high, and the
mediums/lows that are cheap and real, then does the iso/3D/Tavern art pass.
CRITICAL — unauthenticated remote kill switch
Socket.IO does not wrap listener invocation, and this process treats
uncaughtException/unhandledRejection as fatal, so ONE unguarded throw in ANY
listener exits the server. `socket.emit('debug_ping')` with no argument did
exactly that, and systemd restarts straight back into it. Every registration
in handleConnection now goes through a safe-dispatch wrapper that contains
sync throws and rejected promises to the offending socket; handleDebugPing
also stops trusting its payload.
HIGH
- NIP-98 login now binds the signed event to this deployment's host. The path
check alone let an event signed for any other origin authenticate here, and
linking ADOPTS the account owning that pubkey — full account takeover.
- Fog of war is enforced server-side. getState() shipped absolute
entrance/exit/treasure coordinates and the monster's exact position in every
frame, with concealment implemented only in the browser. Discovery is sticky;
the monster is reported only while in view.
- NODE_ENV must now be explicit and recognized. Unset, every production gate
read false while DebugManager reported production and silenced logging.
- Tavern chat applies the lobby's ban list, rate limiter and user_id
attribution before entering global chat.
- userRegistry is deleted by socket.id (it was keyed by socket.id and deleted
by socket.client.id, so nothing was ever freed).
MEDIUM / LOW
- Payouts: a legitimately split transfer_split kept its transaction hashes
instead of quarantining already-broadcast funds with tx_hash NULL, and
provably pre-broadcast failures return to 'pending' rather than needs_review.
- RateLimiter key parsing fixed — cleanup() freed nothing and expired windows
never reset, because ids contain colons.
- Enter no longer submits a paid entry from behind a modal.
- Lobby chat no longer double-escapes server-escaped text.
- Rate limits for identity:get / address:prompt / early_entry.
- Spectators released when a game ends by disconnect, not just settlement.
- movementManager / addressManager per-socket maps cleaned on disconnect.
Renderers
Iso: opaque shading instead of translucent fog, additive torch/lantern/beacon
light pools, crevice AO, embers, cold memory wash, and it stops reallocating
the whole canvas every animation frame.
3D: instanced level geometry (3 draw calls), procedural stone/flagstone
materials, torch lights + shadows, real meshes for stairs/chest/monster
(they were humanoid capsules), dust motes, wall cutaway, and lighting
rebalanced now that fog is measured from the camera.
Tavern
Block height and hash fragments drift around the stage edge and ripple when a
block lands, fed by a cosmetic chain tip added to the blockheight broadcast.
Speech bubbles work again (they keyed off a socketId the server stopped
sending).
126 suites / 1005 tests pass.
Spent outputs (including change) lock for ~10 blocks on Monero. When a
payout batch has no unlocked balance to cover it, attempting transfer_split
fails noisily and, for batches, gets parked in needs_review for an operator.
Add a pre-flight unlocked-balance check that reverts the batch to 'pending'
(re-checked every new-block tick and batch interval) so it sends as soon as
outputs unlock — no manual intervention. Also classify insufficient-funds
RPC errors as pre-broadcast/safe-to-retry: single -> pending (not failed),
batch -> pending without an operator alert (not needs_review), since no tx
went on-chain and there's no double-pay risk.
transfer_split sends one on-chain tx to many destinations, so payout rows
in a batch share a tx_hash. The old UNIQUE index on payouts.tx_hash made
the 2nd+ row fail to record, leaving funds sent but rows stranded in
'processing' and re-sendable by the retry service.
- Migration 015 drops the payouts tx_hash unique index (idempotency is
enforced by idx_payouts_one_per_game) and extends that guard to cover
'processing' so no second payable row can be inserted mid-flight.
- Batch completion now updates the whole batch in a single transaction via
id = ANY(...), so rows can't be left half-completed. Per-user stats use
BIGINT arithmetic in SQL (no JS float).
- Ambiguous batch failures are marked 'needs_review' (which the retry
service does NOT pick up) instead of 'failed', preventing a double-send
when transfer_split may have already broadcast. Operator is alerted.
Adds batchPayout.test.js covering the shared-tx_hash completion and the
needs_review failure path. Full suite: 12 suites / 115 tests.
jwinterm