[LE0-Lin]

Summary

• Resolve explicit guardrail selection without walking/discovering the whole repository. An unrelated broken/external/nonregular instruction source no longer blocks a selected-file advisory review or confirmed append.
• Share the existing supported-file and directory-exclusion predicates with normal discovery; retain relevant ignore rules, selected ancestry, canonical filename preference and selected-file boundary/read/append guards.
• Reject directory aliases/reparse ancestors and ambiguous directory casing instead of guessing a whole-tree case-folded winner. Selection errors are sanitized; no unrelated instruction contents are read, repaired or certified.
• Keep normal scan/doctor behavior fail-closed. Review success does not turn an unsafe/incomplete whole-tree check into a clean score; confirmation tokens, scored rules, JSON fields and policy defaults are unchanged.
• Add 15 selection/scope contracts, an installed-wheel scoped preview/append check, and synchronized English/Simplified/Traditional core guidance.
• Source-only change: no package/version/tag/v0 update, frozen research rewrite or external maintainer contact.

Verification

• Reproduced the old whole-tree dependency before implementation; new scope fixtures failed under that behavior.
• Full Windows suite: 477 tests, 467 passed and 10 platform skips.
• Linux selected-scope suite: 15 passed, including case-sensitive and FIFO controls.
• Wheel/sdist build and installed-wheel acceptance passed; independent full-scan rejection remains covered.
• Demo remains F 56; self-scan remains A 100. These are engineering checks, not accuracy estimates.
• Runtime source identity: 50d40d5e4ad28deb26407d4cba2486714107e6fa608546e49221aec338aa83c6.
• Exact-head PR CI and the maintained pinned-repository replay run separately before merge. Expected operational rejection is not a successful clean scan or user adoption.