v2.3.0 — doc-honest glob, CI, hardening, mechanical eval

This release tightens the gap between the documented protocol and the actual behavior of the guardrail scripts, adds continuous integration, makes the install path crash-safe, and adds mechanical structural checks for AGENT_HANDOFF.md and SKILL.md.

Added

• CI matrix (P2): new .github/workflows/ci.yml with two jobs.
  • shellcheck (Ubuntu) lints every *.sh in the repo. SC2254 is excluded because check-ownership.sh deliberately uses dynamic case patterns as globs — that is the parser's contract.
  • fixtures matrix on ubuntu-latest and macos-latest runs the ownership parser fixtures, the install rollback fixture, and the new mechanical eval checks (Python 3 + PyYAML provisioned in the job).
• Atomic install rollback (P3): install-skill.sh now exits with code 3 and restores the pre-install backup if cp -R fails. The system is left in either the "old target + no backup" or "backup + new target" state, never an intermediate state. Documented in usage().
• Install rollback fixture (P3): new scripts/test-fixtures/install-rollback-test.sh PATH-shadows cp with a failing stub to assert the rollback contract.
• Glob-crosses-slash fixture (P1): new scripts/test-fixtures/handoff-glob-crosses-slash.md and corresponding run-tests.sh case that pins the actual behavior of bash case glob matching across /, so the documented semantics cannot drift from the parser again.
• Mechanical eval checks (P4): new evals/run-mechanical-checks.sh driver with an explicit "mechanical guards only" head comment, running the ownership fixtures, the install rollback fixture, the structural lint over AGENT_HANDOFF.md, and a PyYAML-based check that SKILL.md frontmatter parses with name and description present.
• Structural lint (P4): new evals/lint-handoff.py verifies required top-level headings, the three ownership subsections (### agent-owned, ### user-reserved, ### frozen), ISO 8601 timestamps in ## History, and that ## Next agent starts from is non-trivial.
• Supported Platforms section (P2): README.md and README_IT.md now document the supported target as Bash on Linux and macOS (scripts rely on Bash features like mapfile and [[ ... ]], not strictly POSIX sh); native Windows is unsupported and WSL is not part of the test matrix.
• Version metadata (P5): skills/cli-collaboration/SKILL.md frontmatter now carries version: "2.3.0". The same version is surfaced in README.md, README_IT.md, and CHANGELOG.md so the source of truth is not solely the frontmatter.
• Italian contributors record (P5): new CONTRIBUTORS_IT.md mirroring CONTRIBUTORS.md for symmetry with README_IT.md.
• Python ownership parser (PR #4 review cycle): new skills/cli-collaboration/scripts/parse-ownership.py. See the matching entry under ### Changed for the migration rationale.

Changed

• Doc-honest globs (P1): corrected the documented behavior of * in ownership patterns. * in a bash case pattern matches any sequence including /, so scripts/* matches both scripts/foo.sh and scripts/sub/foo.sh. The previously asserted "single directory level" behavior was factually wrong against the parser.
• Ownership parser migration to Python (PR #4 review cycle): check-ownership.sh was reduced to a thin Bash wrapper (exec python3 "$SCRIPT_DIR/parse-ownership.py" "$@") and the full parser + matcher + conflict logic moved to a new scripts/parse-ownership.py. Python 3 is now a runtime dependency of the ownership check. The migration was forced by GitHub Actions fixtures (macos-latest) failing on Bash 3.2.57 whenever a handoff ownership line contained multibyte UTF-8 bytes (en-dash, em-dash). Seven Bash-side mitigations were attempted and ruled out before the migration. CLI contract (flags, env var, exit codes, conflict-message wording) is preserved verbatim. The new Python parser uses fnmatch.fnmatchcase for glob matching, preserving the P1 contract in which * crosses /.
• Shell hardening (P3, selective): install-skill.sh and sync-skill.sh use set -euo pipefail. check-ownership.sh keeps set -u and set -o pipefail (no -e).
• evals/evals.json honesty (P4): top-level status field changed from "tri-cli-complete" to "mechanical-guards-only", with a status_note explaining what CI actually verifies. Each scenario carries a boolean "mechanized" field: scenario C is true (covered by the ownership parser fixtures); A, B, D, E, F are false (would require an LLM-judge harness, out of scope).
• AGENT_HANDOFF.md history (P5): per the project rule in SKILL.md, the ## History section was compacted: the last three detailed entries are preserved verbatim and older entries are summarized.

Out of scope (deferred to v3, gated by docs/future-architecture.md)

• Custom segment-bounded glob parser (** token, single-directory *).
• POSIX flock-based concurrency, .agent/state.json sidecar, MCP transport.
• LLM-judge harness for behavioral evaluation of scenarios A, B, D, E, F.
• Native Windows support.

---

Full changelog: https://github.com/Spe1977/cli-collaboration/blob/v2.3.0/CHANGELOG.md