v7.0.5 — chat resilience and scoped worker failures

• Routine chat-read and acknowledgment maintenance handles specifically classified local admission delays and transport timeouts with bounded backoff. Failed cursors do not advance and provider pacing remains authoritative.
• A message positively known not to have been dispatched can wait in a bounded room/channel queue. An uncertain sent message is not blindly repeated; messages from an old room are not delivered into its replacement.
• Joining-player records and ready state remain available while narrowly classified profile lookups wait. Matching undispatched score reads defer without turning a temporary lookup delay into a room crash.
• A delayed status snapshot alone is not treated as proof that a worker exited. Freshness requirements for website observations, emptiness and update admission remain strict.

The native supervisor can park a failed worker after positively confirming its room closed and its process exited, leaving healthy occupied siblings running. Cleanup has a bounded reconciliation period. Invalid authority, unresolved ownership and unreconciled failures still require escalation; this is not complete isolation from every shared-service failure.

A later incident exposed a remaining ordering bug: an ordinary room shutdown could close its local connection before process exit, causing the shared controller to escalate too early. v7.0.7 repairs that owner path; it is queued separately and is not claimed live here.

The release retains the cumulative playlist repairs, compact help, full feedback names, map-selection tuning, saved room settings, OAuth/Discord integrations, catalogs, ratings and enrollment. The owner intentionally runs fixed Standard, Mania 4K and Mania 7K; the disabled automatic Standard room is not an outage.

Affected source and frozen-runtime checks passed. Installed four-worker donor tests with full Standard/4K/7K catalogs preserved occupied matches during a chat-acknowledgment timeout and during a positively closed alias failure, retained pending votes, and shut down cleanly. Provider traffic was invented for these tests. Unchanged whole-repository/calculator suites and deliberate production fault injection were not repeated.

Operator automatic-update bootstrap · Queued owner repair · Release history