Failure messaging and runner console reliability

Goal: a failure a runner or user meets is named once, plainly, where they look. - Answer delivery: the latest failure and retry count show on the task; a stuck delivery never blocks the runner's settle. - Identical consecutive dispatch refusals collapse to one timeline row with a count. - Provider breaks are classified on first occurrence with the vendor's own sentence and no traceback (depleted Gemini credits = billing; a rejected Anthropic key ends the run at once). - Agent activity shows no CLI update nags; runner pane titles carry no escape payloads; control MCP tools reach Claude ACP sessions. - Settings > Providers gap text reads the same at phone and desktop width. - An Audit > Sessions row opens its transcript; Session from the slot drawer leaves the console header clickable. - Jarvis says when docs search is down and links `/docs`; its replies carry the answer, not tool narration. Out of scope: login flows; adding providers. Done means: - Each item reproduced on dev and shown fixed (screenshots at desktop and 390 px for UI items). - Tests pin provider classification, the settle path and the Jarvis replies; `./dc check quick` green.

All 7 failure-messaging items fixed and pinned by tests; PR #7045 merged into develop (b4507083b); quick gate green; review lane passed. UI checked at 1440/390 on the slot stack, not dev. Drafts: #10337 transcript viewer, #10350 branch cleanup. Memory: used mem-25288; proposed none.