GHSA-74p7-6h78-gw8p

## Impact Following the path-safety patches in [GHSA-wx3m-whqv-xv47](https://github.com/umanio-agency/skillctl/security/advisories/GHSA-wx3m-whqv-xv47) (v0.1.2), a comprehensive multi-angle audit surfaced five further vulnerabilities, now patched in v0.1.3: 1. **`source_sha` argument injection in `git ls-tree` (CRITICAL).** `InstalledSkill.source_sha` deserialized from `.skills.toml` (committed, PR-mergeable) flowed unvalidated into `git ls-tree -r -z <refspec> -- <path>`. Because the refspec sits before `--`, an attacker who slipped a malicious `.skills.toml` into a PR could set `source_sha = "--name-only"` / `--abbrev=0` / `--output=…` and corrupt the diff classifier (which drives `pull` / `push` destructive decisions), or forge a divergence state to trick `push --on-divergence overwrite` into clobbering the wrong content. 2. **FIFO / device / socket denial-of-service in `copy_dir_all` (CRITICAL).** The file-type branch only checked `is_dir()` / `is_symlink()`; a FIFO inside a skill folder fell through to `fs::copy`, which blocks indefinitely waiting for a writer. A character device like `/dev/zero` would read until OOM. Reachable on `skillctl add` against any adversarial library. 3. **`add --dest` arbitrary-directory wipe in agent mode (HIGH).** `--dest` accepted absolute paths and `..` traversal without validation, so a single invocation `skillctl add --dest /Users/victim/.ssh --on-conflict overwrite --skill <maliciously-named>` would `remove_dir_all` arbitrary directories — no `.skills.toml` round-trip required. Reachable in any non-interactive / agent-driven workflow where flag values may be attacker-supplied. 4. **Commit-message trailer forgery via skill names (HIGH).** Skill names were spliced verbatim into `git commit -m "update skill: <name>"` and into the `commit.message` field of `--json` output. A skill named `foo\nCo-Authored-By: evil@x` produced a forged commit trailer that downstream tooling (Linear, GitHub commit-bot, release-notes scrapers) treat