This is an old issue, still not resolved and wrongly documented, unfortunately. The workaround is that you dont dump all the samples into the spatial, but devide it up between spatial and temporal. for interiors you need lots, at least 16k. For 16k samples would go like this: 256 Spatial and 64 Temporal for example.
Also make sure you do NOT have the deferred renderer in your render settings. that also messes up things