[Cylop]

Fixes the minor findings deferred from #55, #56 and #57, and adds a setup guide for a small team.

Fixes

• Profiles: a person with status: left expects nothing in profile check, even when a directory group still assigns them profiles.
• access review: refuses existing NNN.json/NNN.execution.json before discovery; escapes Markdown in --detail summaries; the old-index.json refusal names the folder and the CI remedy.
• Starter verify workflow: accepts only a finished post-merge permesh-plan.yml push run on the default branch of this repository, including runs that exited 4 or 6, which still hold the matched plans. The job exits like the CLI for the worst plan (errors, then 3, 4, 6). Behavioural tests run the workflow's shell steps with stand-in gh and permesh.
• Starter: workflows pass PERMESH_GOOGLE_CLIENT_SECRET and PERMESH_GOOGLE_REFRESH_TOKEN, because a Google access token expires within an hour, too short for weekly CI. The README uses env:// references instead of the keychain-only auth login. GitLab docs say not to mark provider variables Protected.
• permesh mcp:
  • a failure to reach the client exits 5 instead of blaming the protocol version;
  • a symlink that leaves the workspace reads as "File not found in the workspace", so its target's existence isn't revealed;
  • empty profiles/remove lists get a specific message, and the schemas set minItems: 1;
  • the instructions say report.complete;
  • docs list every command that is left out;
  • ADR 0035 gains "Rejected options".
• Stable fixtures: new access-plan-move.json and check-report-notifications.json (with notifications enabled; failure reasons normalized as human text). There's a renamed-field self-check on the plan fixture and a stronger reload check for the profile suggest draft. README and ADR 0036 now state coverage accurately.

Docs

• New docs/team-setup.md: about 10 people on GitHub and Google Workspace, from repository to weekly CI and reviewed joins, moves and leaves. Linked from docs/README.md and getting started.

Validation

• cargo fmt --check, cargo clippy --workspace --all-targets -D warnings and docs check are clean.
• cargo test --workspace: 734 passed, 0 failed.
• Final review on Opus found 0 Critical and 3 Important findings, all fixed in one pass, each with a test that failed first where code changed:
  • verify refused runs that exited 4 or 6;
  • the guide's credentials couldn't work in CI;
  • the guide's example profile included an undefined profile.

Rulings

• Leave plans on providers without a groups capability: completeness unchanged. Such a provider can't report group routes, and requiring memberships would make every leave there permanently incomplete. Documented in access-plans.md.
• Cancelled MCP call queued behind a running one: no change. rmcp drops responses for cancelled requests, and a new test shows the queued call starts no provider.
• Race between MCP path checks and file creation: documented, not fixed. A local writer that can swap symlinks can already write anywhere the user can.
• Notification failure reasons: treated as human text in fixtures, with their type still frozen.

Deferred minors

• access verify now applies left too: verifying an old join/move plan for someone since marked left lists all their access as extra (exit unchanged).
• The verify workflow compares .path exactly; if GitHub ever adds @ref, every run is refused.
• GFM still autolinks bare URLs and emails inside escaped --detail text; a leading - in an operation can nest a list.
• check_out silently passes when --out can't be listed; on Windows 000.JSON isn't matched.
• The guide redirects profile suggest straight into profiles/; a failed run leaves an empty file that breaks profile loading.
• The guide doesn't say where provider versions for ci/providers.txt come from.
• The access-plans sentence on providers without groups is hard to follow.