v0.86.1

Added

• The MCP endpoint is documented where an operator looks for it. README.md had no MCP section at all, and docs/API.md described the Bearer <handle>.<key> header and said wiring a client up was the operator's deliberate act without ever showing the one command that does it. Both now carry claude mcp add --transport http.

Changed

• One name for the test scratch directory. tests/configs/logging-*.php pointed at /tmp/event-logger-nodes-test while the baseline config used /tmp/newspack-event-logger-nodes-test, so run-coverage.sh carried a second rm -rf for the unprefixed name and a comment explaining the split. Both are the prefixed name now, declared once as Tests\TestCase::TEST_DIR — four classes each held a private copy of the same constant, and a fifth spelled it as a bare literal — and the cleanup is one glob that covers the scratch base and every make_temp_dir() sibling beside it. Test-only.

Fixed

• The Overview breakdown no longer strands itself off Server on a hub. OverviewSection dropped the server option and WROTE the selection back to status whenever serverNames was empty. Empty means "the overview reply has not landed" as often as it means "one server", and the reset effect ran on the FIRST render pass — the ! overview early return sits below the hooks — so every page load demoted the selection before a name could arrive, and nothing promoted it back. A hub opened on Status Codes with a populated Server selector beside it. The fallback is now DERIVED per render in PerformanceDashboard, as canBreakDownByServer, and feeds the chart, the option list and the overview request alike, so nothing latches: a reply that has not landed keeps the default, a reply carrying one server (or only the Other overflow key) falls back, and a hub whose second server reports a window late gets the axis back instead of staying on status for the session. Regression from v0.84.1's "no Server breakdown outside hub mode" (050973f9).

Security

• allowed_users now gates the dashboards, not only the settings page. Admin::current_user_allowed() — the manage_options baseline plus the allowlist — was called from three places, all of them the Settings screen. The top-level Event Logger menu and the Performance, Errors, Gyroscope and Request Log submenus are registered by the main plugin file's own admin_menu closure, which asked for manage_options alone, so an operator who set allowed_users to restrict the admin UI still handed every dashboard to any administrator. It takes the same gate now. Access to the underlying data is unchanged and stays with the service CI's capability roles: this closes the UI gap the setting names, not the command endpoint.