fix(generate): refuse to sweep orphans through a link out of the output root

[dyoshikawa]

Summary

Fixes item 7 of #3294: the --delete orphan sweep could follow a symlinked output directory out of the project and delete files there.

Reproduction (on main)

mkdir -p proj/.cursor notes && echo secret > notes/precious.md
ln -s "$PWD/notes" proj/.cursor/commands
# proj/.rulesync/commands/a.md exists, targets: ["*"]
cd proj && rulesync generate --targets cursor --features commands --delete
# => Deleted: .../proj/.cursor/commands/precious.md   (notes/precious.md is gone)

The same happens for rules (.claude/rules -> ~/somewhere). subagents and skills were already refused by their own guards, so only the shared file-feature path was affected.

Candidate enumeration uses followSymbolicLinks: false, which only covers links inside the scanned directory, not the directory itself. FeatureProcessor.removeOrphanAiFiles then removed whatever was listed with no inside-root check.

Fix

• removeOrphanAiFiles now asks refusesDeleteOutsideRoot before deleting each candidate. It uses the existing writablePathEscapesRoot on the candidate's parent directory, so:
  • a linked directory that leads outside the output root: refused, with a Refusing to delete "<path>": it resolves outside "<root>" through a symbolic link warning (same shape as the write-side refusal);
  • a linked directory that stays inside the root (dotfiles layout): swept as before;
  • an orphan that is itself a symlink: still unlinked (removing a link never touches its target).
• The check runs before the dry-run branch, and refused files are not counted or reported, so --check / --dry-run agree with a real run instead of reporting a deletion forever.
• Documented in docs/reference/cli-commands.md (--delete section); src/generated/docs-content.ts regenerated.

Tests

Unit tests in src/types/feature-processor.test.ts for each of the cases above, including the dry-run case. I also re-ran the reproduction against this branch: precious.md is kept and the warning is printed.

The other follow-ups from #3294 (items 1-6, about orphan-sweep.ts) go in a separate PR.

Refs #3294

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y4MwhqMc6s4ct5MzMReGJj

[dyoshikawa]

@dyoshikawa Thank you!