{"content":"GT#72 live-observation fix 2/2 — premature [STATUS:REVIEW] flip (P0): the chair's closing clock line \"…→ 实读放行 → [STATUS:REVIEW]。全组最后两棒。\" was a citation of a FUTURE step, but the parser's \"any position on the last line\" rule (task #52 heritage) applied it as an instruction: the task flipped to review 45 minutes early mid-assembly, three dispatches got suppressed, and the dispatch_held notes needed 4.5 minutes to reopen it. Task #70 replayed despite both prompt defenses — the LLM chair broke its own backtick rule.\n\nFix (branch fix/group-task-observe-r1, commit 52f59143):\n- the end-line instruction field now requires the tag to sit at an END of its line: leading it (\"[STATUS:REVIEW] — explanation…\", task #52/#63 shapes) or capping it (\"终检完成，进入验收 [STATUS:REVIEW]\", task #52 short verdict / task #47 illegal-directive shapes — audit + anomaly paths preserved)\n- a tag EMBEDDED mid-line with words on both sides is a citation: descriptive, never applied; the rate-limited descriptive note (task #63 heritage) tells the chair to re-send a bare tag, so an intended-but-malformed verdict self-heals in one turn instead of costing a reopen ceremony\n- chair playbook, runbook §5, and the group-task skill doc now teach the same lead-or-end shape\n\nDaemon suite 242/242 green (5 new parser tests + GT#72 replay).","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}