Addresses everything but:

For just minor_roads we'd also backfill missing surface tags as paved to promote more merging (rather than leaving it null). But for other road kinds we'd leave null as null.

Because this also will affect tile size, let's compare a few tiles before/after (even reusing the same tiles from the earlier PR?).

For just minor_roads we'd also backfill missing surface tags as paved

Yup, I had second thoughts about this, so didn't initially implement it. I've added it (although disabled) in 1f1b9a3.

I ran a small sample of tiles in a zoom sequence roughly over Brooklyn:

Note that the sizes here are for roads layer only (although roads are often the largest single layer at mid zooms).

Keeping the simplified surface property has a slightly positive effect at zooms 8 & 9 because we're applying the simplification to all road kinds, whereas before we were only dropping surface on minor_road. Zooms 12 through 14 see a small (<2%) increase in layer size due to the reduced opportunities for merging. Zoom 15 and up is unchanged, as we still keep the full range of values.

Backfilling the default surface: paved at zoom 12 has the expected effect: it recovers some of the lost ground by providing better merging opportunities. At zoom 13, however, it makes the layer slightly bigger because it takes a byte each for the index of the key and value that we're adding to those features' properties.

For the basic PR without backfilling, the property indices take up 11,927 bytes across 561 minor_road features. With backfilling we're able to merge down to 548 features - but the property indices are now 12,629 bytes. The average bytes per feature increased by almost 2, as we'd expect, but this pushes the overall size unexpectedly higher.

It looks like the benefits of backfilling might be minor at best. We'd get better results if we dropped surface: paved and made that an explicit default. However, dropping properties that exist feels much more invasive than backfilling defaults, and I'm not sure we want to do it. Perhaps in v2 we can switch to having an explicit default of paved.

I've filed #1724 for the v2 breaking change.

Agree we should not implement the surface backfilling right now in this PR. Take it out entirely versus commenting it out and then we're good to merge?

Take it out entirely versus commenting it out and then we're good to merge?

Sounds good. Removed in c842849.