Document the defects found by this review The wiki review found nine defects the test suite could not see. Each page that described the old behaviour is corrected, and the pages that documented the defects as open gaps now document the fixes: - Real-Time-Layer: the four event-path gaps (snapshot payload, unreachable offline state, dropped greeting, missing autonomy broadcast) are now a "defects and fixes" table rather than a known-gaps list. - Data-Model: documents link_code_issued_at and why the link code expiry needed a new column to be enforceable at all. - API-Reference: run-agent is caller-scoped with a truthful 429; snapshots no longer 500 on a malformed limit; the webhook fails closed. - Operations: scoped debounce keys, dispatch_market_sweep, last-use token purge. - Telegram-Bot, Configuration: corrected three claims that the fixes made false. - Quality-and-Testing: 359 tests, with the new regression module documented. - Incident-Log: a closing note that a documentation pass found eight more.
Rewrite the wiki as full engineering documentation Replaces the default stub with 16 pages covering architecture, the agent debate, the execution guard, the data model, the real-time layer, the API, the MCP tool server, the Telegram bot, configuration, operations, quality and testing, engineering decisions, an incident log and the roadmap. Adds a custom _Sidebar for navigation between all of them. INC-003 documents the WebSocket flap with a reproduced root cause: redis-py 8.1.0's implicit DEFAULT_SOCKET_TIMEOUT of 5s is identical to channels_redis's brpop_timeout of 5s, so the read deadline races the blocking window. Measured 4/4 failures at the old value and 4/4 clean at the shipped value.