Add a --preview mode that runs Skill synthesis without writing files

[chigichan24]

Why

• npx crune --dry-run is intentionally local-only: it lists candidate topics with their reusability score and never calls claude -p (verified in scripts/cli.ts:176-188 — execution exits before the synthesis branch).
• That makes --dry-run cheap and offline, which is great for a first-pass look.
• But users who are deciding whether to actually emit Skills often want to see what the LLM-synthesized output would look like before committing to writing files. Today the only way is to run synthesis for real and let it write to --output-dir.
• We need a richer preview path that mirrors the real run (synthesis included) but emits to stdout instead of touching the filesystem.

What

• A new flag (proposed: --preview) that:
  • runs the same pipeline as a normal invocation: build graph → pick top candidates → synthesize via claude -p
  • skips the final mkdir / writeFileSync step (scripts/cli.ts:230-236)
  • prints each synthesized Skill to stdout with a clear separator so users can pipe / less / diff
• Honors --model, --count, --sessions-dir exactly like a real run.
• Combining --preview with --skip-synthesis falls back to today's heuristic-only output (i.e. equivalent to --dry-run).
• Keep --dry-run as-is (local, no LLM) for backward compatibility.

How

• Refactor the synthesis loop in scripts/cli.ts:193-237 so the synthesis step and the file-writing step are decoupled. They currently live in the same for loop.
• Thread a writeToDisk: boolean (or equivalent) through the per-candidate handler. --preview sets it to false.
• Confirm the documented behavior in README and --help output.

Notes

• Once Add an automatic Skill evaluation pipeline (skill-creator-style quality checks) #20 (Skill evaluator) lands, --preview should also surface the evaluator's score / hints alongside each synthesized Skill so users can judge quality before installing.