Make reconstruction deterministic and harden LoD 1.1 fallback

[rijos]

Kadaster observed run-to-run variance in roofer while generating the national 3DBAG. The same input could produce different output across runs.

• Pointcloud density check now uses an absolute threshold. Previously the minimum density was a calculated relative value, which itself varied between runs. This caused buildings with sufficient point density to be skipped non-deterministically. The check now compares against a fixed absolute value.
• Region growing no longer enforces a per-footprint time limit. The previous limit_n_milliseconds cutoff meant a footprint could be abandoned simply because the scheduler hadn't granted the process enough CPU time, leading to starvation-driven variance.
• Solve run to run variance by doing a proper pointcloud_insufficient.clear(); after each tile. Preventing invalid lod11 fallbacks.

Before merging the pull request, I'd like to confirm the following:
We've set the floor to 1 point/m², Roofer has mainly been validated on AHN (~10 points/m²), so a higher floor like 5 points/m² may be more appropriate.

We're dropping limit_n_milliseconds entirely. How should we handle existing configs that set it, show a warning or show a configuration error?

[rijos]

Possible regression when processing individual footprints, we see possibly unjustified LOD 1.2 fallbacks in our national run. Will investigate further.

[Ylannl]

We've set the floor to 1 point/m², Roofer has mainly been validated on AHN (~10 points/m²), so a higher floor like 5 points/m² may be more appropriate.

A default of 5 seems reasonable. And maybe we should make this optional, as to not have 'hidden' filters enabled by default that may be confusing to new users (as we have now).
I also want to add an option for the minimum footprint coverage fraction. Together this covers the cases for which the relative density offset was originally implemented.

We're dropping limit_n_milliseconds entirely. How should we handle existing configs that set it, show a warning or show a configuration error?

I agree it would be best to drop this feature given the problem that you see with it. We can simply remove it, and users will see an error if they still try it (don't expect there to be many users that would see this).

[rijos]

Remove the deprecated notification.

[rijos]

Scrap

[rijos]

Hi @Ylannl, we've completed a test run using this version of roofer, because we're confident that this version works, I have the following requests:

• Let's keep the abs guard, even though it was proven that it should be technically impossible, it now acts as an additional guard.
• Dropped lod11-fallback-time as requested.
• Keep min_building_density at 1.0, as without it, we won't have a graceful lod11_fallback. I think it can be any positive float, but it's really data dependent.

The new min_building_density should also help with:

• Add a scale parameter, to better handle low density pointclouds #169