Fix all findings from the technical-writing review Correctness: the provider registry section across four pages was written before #458-#469 landed (oMLX is now expected to be a real registered provider row, key "omlx", not an absent/null upstream); migration range V1..V7 -> V1..V8; single-run lock misattributed to RunBenchSupport.java (the real lock is WorkerService.guard()'s - RunBenchSupport.acquireLock() is unused); the 14-routed-views list wrongly included DiffView, which has no @Route; the Vaadin LOC figure was stale; Groups' pooling key was wrong in two different ways on two different pages (no `mode` dimension - it's (task, model, key_hash)); User-Manual-Providers.md's own routing-rules section contradicted its own intro screenshot caption, since the UI moved from a free-text model field to a cascading provider/model picker. Completeness: added a Design-Decisions.md entry for budget-exhaustion rescue (previously only ever mentioned as a dependent clause, never explained), a matching Features.md entry, and cross-linked both from the User Manual's existing Budget rescue section. Clarity: clarified that L1's "no Docker-runtime checks" means no C1/C2, not no Docker at all (B1/B2 still run in a container); added the two missing per-task-dialog screenshots (History, Live) so all seven tabs now have one.
Add full ACE wiki: architecture, features, design decisions, user manual Covers the job queue/worker model, the benchmark runner, the oracle and Docker-outside-of-Docker scoring, the LLM provider registry, Judge and cross-task lessons, and a screenshot-driven walkthrough of jobs, providers, Preflight, runs, groups, compare, and experiments on the live instance.