[TiantianFlow]

Follow-up to #81 (issue #80), from an audit of the personas against the skills, requirements, and READMEs.

What changed

• Lead skill, README (en and zh-CN), requirements lead row: they now say what the lead persona says: when seats' results disagree, the lead chooses by their evidence or writes the task that settles the point, and never puts its own answer in their place.
• Lead persona: integrating branches is the lead's job; a design isn't accepted, and the goal isn't done, until a reviewer has tried to break it (as the skill requires). Two independent proposals are two tasks, so they fit the skill's rule for running seats side by side. Two leftover Foil terms are now plain words.
• Reviewer persona: it reports pass or fail, matching the result contract, instead of "changes needed".
• Researcher persona: the lead decides, not "the lead and the designer".
• Memory curator persona: it writes replacement lessons in its result for the lead to file, since a worker's command list has no --replaces.
• Documentation writer persona: the placeholder tool init example is now a real command.
• Operator skill: it no longer names an outside persona catalog.

Verified

• uv run --frozen --extra dev pytest -m "not e2e_live" --deselect tests/test_view_browser.py: 566 passed on this branch's head.
• ruff check .: clean.
• git grep -i agency: no hits.
• Independent audit (consistency with skills, docs, and requirements; Foil mechanics leaking into personas; privacy; publish safety): its findings are fixed in this PR.

Not verified

• tests/test_view_browser.py locally (the sandbox's Chromium does not match Playwright's version); CI runs it.

---

Generated by Claude Code