Expected Topological Model

My basic answer is 3) "do nothing", such cases are quite commmon in OSM and therefore need to be considered by data consumers.

Longer answer below.

There is no need in OSM mapping for a single polygon reflecting landuse/landcover in any given area. The reasons for this are many:

• Woods, lakes, meadows, residential areas all exist in university, school and hospital campuses. These are either not, or unlikely to be, errors. I would regard the JOSM validation rules to be incorrect in these situations.
• Individual polygons will be created by different mappers at different times with different levels of knowledge of OSM, time available to add data, expectations of how the data will be used, quality of imagery at the time the polygon was mapped (which may be vastly different now).
• Small polygons with a different class within a larger polygon are very common and entirely reasonable way of mapping the data. The main OSM map produced using CartoCSS makes a number of assumptions based on area and tag to decide which ones should be rendered on top of others. These algorithms work very well. (One exception is in the area you mention, where a small wood overlaps a larger feature, see https://github.com/gravitystorm/openstreetmap-carto/issues/2641 for discussion).

If you want a tesselated layer consisting of a contiguous set of polygons you need to post-process OSM data. You may need something like the CartoCSS algorithm, although in practice I have found a simple hierarchy of tags can work well (slide 16 here). My approach (slide 24 in previous link) involved progressive spatial unions and clipping according to the hierarchy I used. The main complication was dealing with invalid geometries at all stages of the process.