tlms enforcement for hub-triggered llmmsg-srv replies

tlms (too long make it shorter) shorthand, per-agent setting in settings.json/settings.local.json. monitor-llmmsgsrv-cc built the CLI-side UserPromptSubmit hook enforcement; that hook does not fire on a hub-triggered turn (no UserPromptSubmit event exists for it). Mechanism: (a) CLAUDE.md/llmmsg-protocol.md rule instructing self-formatting to the tlms shape (line1 verdict <=15 words, line2 decision if any, <=3 evidence lines <=12 words each) at reply-composition time when the setting is on for that agent - the setting itself is a static file read, available regardless of trigger source, only the hook enforcement mechanism is CLI-only. (b) Backstop: extend kpi-coach rubric (pm-llmmsgsrv-cc lane) with a judgment-based rule flagging a tlms-on agent's oversized hub replies, same class as existing Rule 1/9 density coaching - a semantic verdict/shape check isn't wire-gateable like the ack-opener regex is.