Group rooms: #1204 budgets inbound context, but bots still write long CI/merge essays into the log #1251

Observed (prod tip 7fdb8082, 交付组 session-e9b2bc34-…, seq 3112–3171)

After #1203/#1205 shipped in #1204 (room char budget + status collapse + board snapshot):

• Notice is already short (214 chars) and forbids CI/queue/Verdict novels.
• Last 60 lines: ~50k chars, avg ~840/msg; Chief 34/60; 38 long bot lines still match CI/merge/Verdict/enqueue patterns.
• Budget group-turn context and drive work from the board #1204 collapses status-shaped lines when building the next prompt, but the stored room log still accepts 1–2k-char CI/rebase/review essays. The next speaker (and every later FIRST_LOOK) still pays for them until they age out of the 60-line ring.

Gap

Inbound budget ≠ outbound gate. Product does not stop or stub a bot (including admin/Chief) from posting a status novel into the room.

Desired

When a bot posts to a group room and the body matches CI/merge/queue/Verdict / long tip-audit narration, either:

1. refuse / ask the model to retry as a one-line stub, or
2. store a collapsed stub in the room log (full text may live on the PR/issue), or
3. hard-cap outbound room text length for bots (people unchanged).

Prefer host-side enforcement over notice text (notice already says not to).

Acceptance

• Focused test: a bot room reply that is a CI/merge essay is stored short (or rejected) under tip on current main.
• Human lines and short handoffs (#1234 PASS @sha → @Next) still land verbatim.
• Related: Group turns: compress/load room context; progress from board not chat ping-pong #1203 (closed), Research: measure and cut token spend per Bot turn (prompt overhead, history growth, tool output, fan-out) #796.

Evidence

Prod dump /workspace/flybot-jiaofu-session-e9b2bc34-last80.* (u-682b8a, tip 7fdb808, 2026-10-26).