fix(cli): preserve registry inputs when writing reports

[YMuskrat]

Fixes #756.

A registry scan such as skillspector scan registry.json --mcp-registry --format json --output registry.json currently replaces its input with the generated report. The registry branch returns before the existing same-file guard. A comparison report can also be overwritten when passed as both --mcp-registry-compare and --output.

Extract the existing destination check into a shared helper and apply it to both registry inputs before scanning. Existing Path.samefile() behavior covers relative paths, symlinks and hardlinks; distinct output files continue to work.

Validation: make lint, make format-check, and make test pass (9,966 unit tests and 124 integration tests). Regression coverage checks all six input aliases, comparison-input protection, unchanged input bytes, and that rejected scans never call the scanner.

[rng1995]

[SkillSpector Review]

@YMuskrat thank you for spotting that registry scans skip the same-file output guard, and for extracting the existing check into a shared helper instead of copying it.

Value and readiness

Ready for final maintainer review. On main, skillspector scan registry.json --mcp-registry --format json --output registry.json returns from the registry branch before the same-file guard runs, then replaces the input with the report. --mcp-registry-compare previous.json --output previous.json destroys the comparison report the same way. The PR moves the guard into _validate_report_destination and calls it for the registry input and the comparison report before scan_registry runs. It keeps the existing semantics (is_file() + exists() + samefile()), so relative paths, parent traversal, symlinks, symlinked parents and hard links are all rejected. The ValueError is caught by the registry branch's existing except Exception and exits 2 before any scan or write. URL registry inputs are unaffected because Path(url).is_file() is false. The local-skill path calls the same helper with unchanged behavior.

Material findings

None.

PIC tradeoffs

None identified.

Verification and gaps

• Read the scan() registry branch and the local-input guard on main b07f181d and on the merge of this head with main. git merge-tree is clean, and the helper is called at all three sites after the merge.
• Tests: the existing alias test now also runs in registry mode (six aliases, JSON). It asserts exit 2, the error text, unchanged input bytes, and scan_registry.assert_not_called(). A new test covers the comparison report via same path and hard link. The existing success test now asserts the payload is untouched when the output is a different file. Each rejection case would fail on main, where the registry branch never reaches the guard.
• CI is green on 1ac08a2b (6/6). The branch is about 220 commits behind main. Main changed cli.py in that time but not these blocks, so I expect no semantic drift; a branch update and fresh CI before merge would confirm it. I did not run the tests (static review).

---

Decision: Approved (reviewed head 1ac08a2b25699f2ff78ed8038f13e8c30dab3626)