Scope writing context and verify document builds

[nzy1997]

Shared writing guidance could load unrelated context, apply blanket word-count/connective rules, or report a clean build when the compiler failed without printing the word warning. This PR scopes source loading to the requested evidence, preserves meaningful scientific contrasts, and makes the documented Typst check honor both compiler exit status and diagnostics.

The change follows OpenAI's Rethinking skills and prompts for GPT-6 Astra: keep shared rules narrow, preserve user intent, and define a verifiable completion condition. It is the shared-writing slice split from #62 after review feedback.

Validation

• python3 scripts/validate_skills.py: 16 skills pass
• targeted writing/repository tests: 60 passed
• python3 -m pytest -q: 272 passed, 5 skipped
• the build regression covers clean output, warnings, non-warning compiler errors, and an unavailable compiler, including temporary-log cleanup
• git diff --check: clean

[GiggleLiu]

Narrowed on top of your commit (3bb1c65) to the two things the title promises: context scoping in writing-workflow.md and the compile-check fix (which was a real bug — a failed compiler with no "warning" text printed "clean"). Dropped from this PR: the how-to-technical-writing guardrail and checklist rule changes (sentence-length band, "X not Y", notation table, inline calculations), the whole-report drafting mode, the recency-gate removal, and the Markdown output default. Those change rules merged in #60 and belong in their own PR if wanted. The check snippet no longer uses exit, so it is safe to paste into an interactive shell; your test still passes against it.