[LE0-Lin]

Changing only a Cursor rule's glob can stop a known instruction from attaching to a task even while both ordinary scans remain 100/100 and the existing score gate sees no new error.

Add scope compare BASE HEAD and Git-backed scope diff [BASE_REF], using the same declared target set on both sides. Capture each source once, then distinguish file lifecycle, lost trigger matches and aggregated exact-directive availability. Remaining multi-file triggers or another active copy retain availability; manual/model selection and potentially active opaque bodies remain unknown. Reports include original source/metadata lines, captured source hashes, tool identity and the resolved Git baseline commit.

Comparisons are advisory by default. Caller-owned --fail-on-lost-directives and --fail-on-unknown flags opt into separate review decisions; source/Git/encoding/identity failures return exit 2 without partial success JSON. Existing scan/diff scoring, score-policy JSON and Action defaults are unchanged. This is bounded attachment/exact-recognition analysis, not semantic equivalence, runtime enforcement or whole-repository context coverage.

Includes four runnable synthetic controls, installed acceptance through both aliases and the module entrypoint, a CI recipe pinned to the tested tool source, and synchronized English/Simplified/Traditional core docs. SOP/CV, published v0.24.0 artifacts, rolling Action refs and frozen evaluation data are unchanged.

Validation:

• Windows full suite: 592 tests, 580 passed and 12 platform skips.
• WSL Linux: all 20 new change-impact tests passed, including actual Git worktrees and cleanup after source failure.
• Built wheel + sdist; wheel-only acceptance passed all 18 workflow checks, including comparison/Git behavior across three entrypoints.
• Four actual CLI example decisions: narrow = review, remaining multi-file trigger = pass, renamed active source = pass, manual loading = unknown/review.
• Existing demo remains F 56; repository self-scan remains A 100; staged whitespace check passed.
• Maintained pinned-repository replay on a8244bd758b82dfc06d8f55a12543527ce5e477c: five scans and one expected typed source rejection, no acquisition errors, all maintained expectations matched. The report runtime fingerprint 4e63e4d2d451d3e24bf8a8b4454185b775e2589770ddfa39a947aa5d3e48069d matches this implementation; the subsequent change only refines the CI documentation.
• Pinned tool commit dd208c9302737f6d61b4f327de87b197479f8d0d additionally installed via local Git transport. Its installed runtime fingerprint matches the tested wheel/current source; installed narrowing and rename controls returned 1 and 0 respectively.

These are project-authored development and compatibility checks, not independent accuracy or adoption evidence.