Preserve author control through manuscript workflows

[nzy1997]

Manuscript workflows need different defaults for new papers, local edits, broad polish, and already approved changes. This PR keeps venue, story, and outline checkpoints for a new guided manuscript; routes broad polish through a marked proposal; and applies explicit direct edits or accepted findings without asking again. Figure and notation detail moves to an on-demand reference.

The behavior follows OpenAI's Rethinking skills and prompts for GPT-6 Astra, but retains model-independent author-control defaults established during #62. This is the manuscript slice split from #62 and is stacked on the shared-writing PR because it relies on that guide's scoped source and prose rules.

Validation

• python3 scripts/validate_skills.py: 16 skills pass
• targeted manuscript/shared-writing tests: 61 passed
• python3 -m pytest -q: 273 passed, 5 skipped
• the independent-install test follows the moved figure/notation reference
• bounded Claude file checks from Scope research skills to requested work and preserve authorized continuation #62 confirmed that broad polish leaves the original unchanged while accepted replacements apply directly; those samples are behavioral evidence, not a broad model benchmark
• git diff --check: clean

[GiggleLiu]

Closed (GitHub closed it automatically when #72's branch was deleted). Reason it would not have merged as-is: it removes write-paper's Iron Rules and Pre-Submission Checklist, makes venue-before-story optional, weakens the von Delft/Martinis figure rules (error-bar discussion, straight-line theory, minimum line weight), and lets how-to-review-figure edit sources, which is report-only by design. The review-paper application-mode table is a good idea on its own; a PR adding only that table would be welcome.