chore: add CLAUDE.md, /release skill, and permission allowlist additions

[TakiTake]

Summary

Encodes what this session learned shipping pall8t v0.1.0 into the repo so future sessions/teammates don't rediscover it from scratch:

• CLAUDE.md (repo root): build/verify commands (mise isn't installed in the dev container — run cargo directly; cross-target lint required for #[cfg(target_os)] branches), the git worktree workflow, and which publish-facing actions the permission system gates to the human user (tag pushes, gh repo create, pushes to the Homebrew tap, gh pr merge) — framed as "publishing anything public-facing may need the user," not a closed list, since a push to an existing public repo (the tap) was gated too. Links out to README/docs rather than duplicating them.
• .claude/skills/release/SKILL.md: an operational, step-by-step /release skill derived from docs/release.md, with explicit agent/user actor labels per step, the artifact-verification commands (checksum, tarball layout, Mach-O ARM64 magic bytes) that were actually used to verify v0.1.0, and a failure-recovery note for the case where a tag gets pushed against a release-prep state that turns out to be broken. Cross-linked from docs/release.md.
• .claude/settings.json: appended read-only gh allowlist entries (gh pr diff, gh pr checks, gh run list/view/watch, gh release list/view) that prompted repeatedly this session. Nothing mutating.

Review notes

Ran /code-review and /skeptical-review on the diff. The skeptical review caught real issues, fixed in this PR:

• CLAUDE.md originally stated the permission gates as an absolute, closed list; reworded to reflect that it's an externally-enforced, empirically-observed pattern, not a repo invariant — and to stop implying that pushes to existing public repos are safe (they aren't; the tap push was gated too).
• The release skill had no guidance for what happens if the release workflow fails after a tag is already public — added an explicit "if the release workflow fails" recovery path, plus a pre-tag sanity check (confirm Cargo.toml/CHANGELOG.md on main actually match the version before handing the tag command to the user), since release.yml only validates the tag after it's already pushed.
• CLAUDE.md implied the pre-commit lint hook runs automatically; clarified it's opt-in (mise run setup-hooks / git config core.hooksPath .githooks).
• The digest cross-check in the skill's artifact-verification step was framed as a hard requirement but the field can be null on some gh versions; downgraded to a bonus check with the .sha256 file as the authoritative one.

A /simplify pass then deduplicated the tag/push command snippet (was written out identically in both docs/release.md and the skill — now only in the skill, with docs/release.md pointing to it) and removed a session-dated self-reference ("gated the session this skill was written") in favor of a stable statement.

Test plan

• python3 -m json.tool .claude/settings.json — valid JSON
• cargo test — all tests pass (docs/config-only change, included per the mandated gate)
• /code-review + /skeptical-review — findings addressed, described above
• Verified every technical claim in CLAUDE.md/SKILL.md against the actual .github/workflows/*.yml, scripts/lint.sh, mise.toml, and .githooks/pre-commit

🤖 Generated with Claude Code