Stop three answers depending on where the data sat

Self-review of the review fixes, which re-broke narrower versions of two of the things they had just fixed.

The status fold spelled "nothing seen yet" the same as "unknown", which was harmless only while no placement could state an empty status — until the fix that stopped baking a display dash into the field made exactly that reachable. From then on ["", "ready"] agreed on ready while ["ready", ""] disagreed into "mixed": the same facts, two answers, decided by position. An unstated status is evidence neither way, so it is skipped; a real disagreement still surfaces.

The native placement sort moved onto the host so the rows would follow the CLUSTER column — but placementCluster falls back to the SLUG when the catalog has no host, so for exactly those rows the sort key and the printed cell were different strings. It sorts on the cell now.

The cluster filter resolved hosts and slugs in one pass over a map, so a value that is one cluster's slug and another's host was accepted or refused according to Go's randomised iteration order. Hosts get a pass of their own.

Each has a regression test, and each test was run against the reverted code to confirm it fails there — the two order-dependent ones would otherwise pass for months and fail once, remotely.