[Rushaway]

Summary

• Wrong player could be kicked: the spoof-check DataPack carried the raw client index; when the async SELECT came back after the player left, KickClient(client) could hit whoever took the slot. The pack now carries GetClientUserId(client) and Query_SelectConnection resolves it with GetClientOfUserId (+ 0 check) (PlayerManager.sp).
• Spoof check skipped every map change: OnMapEnd deleted the connection and OnConfigsExecuted reconnected, so every player authorizing before the new connection landed hit the !g_hDatabase early return and was never checked. The connection is now kept across maps (SourceMod keeps it; the MySQL driver auto-reconnects), and once a connection is ready, DB_CheckPendingClients re-runs the check for connected, authorized players that missed it (tracked with g_bConnectionChecked, reset in OnClientConnected). This also covers late-load players, who were checked before the DB connected.
• Failed connect never retried: GotDatabase returned on failure with the state stuck at Connecting, so DB_Connect never tried again. It now schedules a retry (DB_ScheduleReconnect, same 15 s delay as a lost connection). The retry timer no longer uses TIMER_FLAG_NO_MAPCHANGE: a map change killing it would leave the state stuck in Wait.
• Any SQL error dropped the connection: DB_Conn_Lost deleted the handle on any results == null. It now only does so when the error text is a real connection error ("Lost connection", "gone away", "Can't connect", "Broken pipe").
• Charset: db.SetCharset("utf8mb4") is called once per new MySQL connection before any query is queued (survives the driver's auto-reconnect); the async SET NAMES is only sent as fallback if SetCharset fails.
• Dead retry code removed: DB_SetNames / DB_CreateTable had retry branches that could never run (DB_Connect() is always true right after connecting) and would have passed a possibly-deleted raw handle to a timer. Removed along with the sm_manager_query_retry convar, which only fed those branches.
• Queries: the INSERT / SELECT now use g_hDatabase.Format instead of plain Format (escapes the auth/address strings). The MySQL upsert now also sets timestamp=CURRENT_TIMESTAMP on duplicate key: the column had no ON UPDATE, so it kept the first-seen time while the SQLite INSERT OR REPLACE path refreshes it.
• Version bump 2.3.6 -> 2.3.7.

Not included

• LogPotentialSpoofer passes the raw client index to an async HTTP request (OnPlayerSummaryReceived / APIWebResponse); same class of bug but not SQL, left for a separate change.

Test plan

• Compiles with spcomp 1.12.0.7259 (same deps as CI)
• CI build passes
• Change map several times: only one "Connected to database." in the logs, no reconnect per map
• Restart the server with players reconnecting immediately: players authorized before "Connected to database." get their connection row inserted/updated after the connect
• Stop MySQL at startup: "Connecting to database ... failed" logged, retry every 15 s, connects once MySQL is back
• NoSteam client with a known Steam ID and different IP still gets kicked; with the same IP it is allowed
• MySQL: timestamp in connection updates on every rejoin
• sm_manager_query_retry can be removed from cfg/sourcemod/PlayerManager.cfg

🤖 Generated with Claude Code

[WareZ57]

Reviewed against master (merges cleanly, CI green). It compiles with spcomp 1.12.0.7259 and the same deps as CI (MultiColors, connect.inc, ripext, ProxyKiller, utilshelper), and _connect_included is defined so the SQL code is actually compiled. No warnings.

What I checked:

• The spoof-check pack now carries the userid, and Query_SelectConnection bails out on !client (PlayerManager.sp:865-875) before it can reach the KickClient at :902. The insert path just skips that cell. That's correct.
• Reconnect state machine: DB_ScheduleReconnect is only reached from GotDatabase failure (only while Connecting and only for the current lock) and from DB_Conn_Lost (only when not already Wait/Connecting). So there's at most one TimerDB_Reconnect in flight, and without TIMER_FLAG_NO_MAPCHANGE the Wait state can't get stuck across a map change anymore. OnConfigsExecuted -> DB_Connect() is a no-op while connected, so keeping the handle across maps doesn't open new connections.
• SetCharset runs before the handle is published and before DB_SetNames/DB_CreateTable queue anything (:662-671).
• Ordering in GotDatabase: the table creation is queued at normal priority and DB_CheckPendingClients queues its SELECTs at DBPrio_Low, so on a fresh DB the table exists first.
• DB_CheckPendingClients uses AuthId_Engine, which is what the OnClientAuthorized forward receives, so SteamClientGotValidateAuthTicketResponse matches the same key. Late-load players are picked up too, since their flag starts at false.
• g_hDatabase.Format escaping and the timestamp=CURRENT_TIMESTAMP on the MySQL upsert both look fine.

One small gap, not blocking: a check that was sent but failed on a lost connection isn't re-run. OnClientAuthorized sets g_bConnectionChecked[client] = true (:212) before the SELECT. If that SELECT comes back with "gone away"/"Lost connection", Query_SelectConnection logs and returns (:877-882), the handle is dropped and a reconnect is scheduled. But DB_CheckPendingClients skips that player after the reconnect because the flag is already true. So a NoSteam spoofer who joins right as MySQL restarts isn't checked until the next map. Setting g_bConnectionChecked[client] = false; in that error branch (you already have client there) would let the reconnect pick them up.

LGTM otherwise.