CoderMCPServer delete and RBAC test follow-ups #378

Follow-ups from the final assessments of #368 (CoderMCPServer delete path), #376 (test S9) and #377 (test S11). None blocks the activation. Line numbers refer to main after #368.

Delete path

1. A marker-only object waits forever when its pinned Coder organization is deleted while the control plane lives. With only a create marker (no status.configID), the delete reads the organization's MCP server configs (findMCPServerSlug, internal/controller/codermcpserver_controller.go:492). Any list error, including Coder's own 404 for the deleted organization, gives CoderUnavailable, so the finalizer stays. Fix: treat Coder's own 404 on that list (isCoderResourceNotFound) as absence. A proxy 404 must still wait.
2. Missing tests:
   • No operator access on delete: the wait branch at internal/controller/codermcpserver_delete.go:61.
   • A marker-only delete when the operator is not a site owner: internal/controller/codermcpserver_delete.go:107.
   • A marker-only delete when the config list fails.
3. A manual finalizer removal leaves the object's sentCredentials entry in memory until the operator restarts. The entry holds only hashes, not values. The release path removes it, but a hand-removed finalizer skips that path.

RBAC test S9 (#376)

4. S9 reads only config/rbac/role.yaml. It must scan every config/rbac/*.yaml for rules that reach core Secrets, and also pin config/rbac/clusterrolebinding.yaml, so a new Role or binding cannot widen Secret access without a golden-copy change.

Admission test S11 (#377)

5. The comment of TestCoderMCPServerStatusPolicy says the policy "denies the status writes of every other test". It denies only status writes of CoderMCPServer objects. Correct the comment.

Refs #155

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high • Cost: $220.71