Rationale from @gravitystorm@m.gravitystorm.co.uk as an issue comment:

We need to be careful with these labels, for a few reasons. Using a single point to represent an ocean is quite an oversimplification! These points are also arbitrarily placed, so mappers could get into endless edit wars about where to put them. Many mappers will use them as "labelling positions" rather than ensuring the position has some kind of geographic basis.

The Arctic Ocean label is a good example. From the overpass screenshot I assume the node is outwith the range we render. Do we want just the bottom half of the label showing?

Normally oceans and sea labels are "hand placed" by cartographers, since the challenge of automating the label placement is so high. But we can solve the technical challenges here; while doing so lets remember not to end up rendering "labelling nodes" by mistake.

Another good reason why it's a complex topic:

The problem with mapping oceans and seas is verifiability - mapping them as nodes is the simplest way to have the tags, in particular the names in various languages, in the database (which are verifiable). The coordinates are essentially meaningless.

Summary here:

I think we have already established in the above discussion and in #2345 that we do not want to render sea and ocean labels based on mapper placed geometries from the OSM database (either nodes or polygons) and thereby have the map painted by mappers based on subjective preferences and the specifics of the labeling style here and the mercator projection.

IMO there are two decent options:

• render by combining the tagging from place nodes in OSM with hand placed labeling geometries based on cartographic considerations for this style (like done by Label oceans and large seas from a static file #2345). This is feasible for oceans but not for seas due to the number of features. Special consideration should be given to avoiding vandalism in combination with slow low zoom update cycles.
• auto-generate labeling geometries based on place nodes and the (simplified) coastline geometry. This is more challenging but would scale better.

In both variants i would only use place nodes to specifically discourage mappers from pointless polygon drawings. There is simply no case where for a sea or ocean a polygon is the most suitable way to map in OSM.