newproject: extract scaffold logic into scripts/scaffold.sh

[nsmiller2501]

Summary

Follow-on to #4. Hoists the deterministic project-scaffolding logic out of SKILL.md and into a standalone shell script at .claude/skills/newproject/scripts/scaffold.sh. SKILL.md now invokes the script and focuses on the conversational/CLAUDE.md authoring half of the protocol.

What changed

• New .claude/skills/newproject/scripts/scaffold.sh: 168-line shell script that creates the project root, populates the directory tree, writes the requirements.txt/config.{do,py,R} stubs, and emits a README on first run.
• SKILL.md shrinks from ~135 → ~50 lines. The deterministic scaffolding steps are replaced with a single bash ~/.claude/skills/newproject/scripts/scaffold.sh <project-name> call. The interactive portions (CLAUDE.md authoring, collaborator prompts, philosophy framing) stay in SKILL.md.

Why

The mkdir / heredoc / file-template logic is purely deterministic — there's no benefit to running it through an LLM each invocation. Extracting it to a shell script makes the scaffold step:

• Cheaper (no LLM tokens spent on emitting the file tree).
• Independently testable (bash scripts/scaffold.sh /tmp/test-project exits cleanly or doesn't).
• Re-runnable without re-prompting Claude.

This mirrors the script-extraction pattern used elsewhere in the skills tree (e.g. /bibcheck/scripts/split_bib.py).

Testing

• Ran bash scaffold.sh /tmp/newproject-smoke-test and confirmed the resulting tree matches what the previous inline protocol produced.
• Confirmed SKILL.md's reference to the script path is correct and that no inline scaffolding logic was left behind.