Preserve scope across research handoffs

[nzy1997]

Survey, dialog analysis, and researcher indexing could repeat choices already supplied by the user or caller, and a full-text KB request could drift into an unrequested report. This PR carries the chosen topic, source, paper set, target KB, and requested endpoint across those handoffs. Open-ended exploration still asks for a material direction choice.

The change follows OpenAI's Rethinking skills and prompts for GPT-6 Astra: make decision boundaries explicit and continue through the authorized deliverable. It is the research-handoff slice split from #62 after review feedback. The survey edits are rebased on #64 and retain its Typst structure and template guidance.

Validation

• python3 scripts/validate_skills.py: 16 skills pass
• targeted repository/Typst tests: 56 passed, 4 skipped
• python3 -m pytest -q: 268 passed, 5 skipped
• git diff --check: clean

[GiggleLiu]

Closing. The survey's contract is that the user picks strategies and directions; this PR makes those agent-side decisions unless the agent judges the request open-ended, and drops the MCP server pick and the know-me-better next-step menu. Main already skips the clarification step when the topic is given. The one piece worth keeping, know-me-better reusing researcher/source/target supplied by a calling skill, can come back as a small standalone PR.