Status: all three halves are wired. Compression is proven on CI. The q3map2 build is still iterating through configure-time dependencies.

.map.options is argv for the wrapper, not a q3map2 command line, and the difference is not cosmetic: -bsp/-light/-vis/-minimap are mode selectors choosing which phase's flag list follows; -sRGB expands to -sRGBtex -sRGBcolor on bsp and adds -sRGBlight on light; -scale <n> drives a separate pass over the built .bsp and a rename. Verified by watching real invocations: the wrapper turned -bsp + -light + -vis + -minimap + -sRGB into a light phase carrying -lightmapsize 1024 … -sRGBtex -sRGBcolor -sRGBlight, and boil's -scale 0.8 into a -scale pass followed by boil_s.bsp → boil.bsp.

An earlier attempt read the options as q3map2 flags. A dry-run showed the cost: erbium's custom flags silently vanished and -sRGB became a phantom stage that would have run q3map2 with no phase at all. Same reasoning as not forking q3map2 for PNG support — the existing tool already does the job.

The wrapper fabricates a Xonotic-style layout instead of requiring one: it chdirs into the map's directory and symlinks <tmpdir>/data → the map's parent, then hands q3map2 two -fs_basepath entries. Verified by instrumenting a stub compiler:

basepath <tmpdir>            -> exposes textures/exx, textures/map_stormkeep, scripts  (i.e. sources/)
basepath $VORTEXARENA_ROOT   -> exposes data/core.pk3dir as a pack

So the symlink half is entirely inside this repo and needs nothing else. The second basepath is a genuine cross-repo dependency, but a small one: measured across all 31 maps, exactly one of 54 referenced texture sets lives only in the game's core content (domination, 1.4 MB), plus core's scripts/ at 124 KB. q3map2 reads .pk3dir as a pack (netradiant tools/quake3/common/vfs.c:240), so pointing a basepath at a VortexArena checkout is enough — no unpacking. Both repos are public, so CI clones it with actions/checkout anonymously: no token, no secret, no submodule. The workflow takes a blobless sparse checkout of data/core.pk3dir and caches it, so 31 matrix jobs do not each re-clone.

Note it needs working symlinks, so it runs on the Linux runner. On a Windows dev box it wants developer mode or an elevated shell.

q3map2 produces a map's geometry, not its art. A shippable archive needs three inputs, and missing any one of them fails quietly rather than loudly:

sources/**.{png,jpg}  ──[compress-textures.sh]──►  dds/textures, dds/models   ─┐
sources/**.map        ──[compile-map.sh -> wrapper -> q3map2]──►  .bsp, lm_*, mini  ├──► publish.py ──► archives
sources/{scripts,env,sound,models,gfx}  ──(copied verbatim)────────────────────┘

• Geometry only would ship a map that loads with no textures.
• Geometry + textures would ship a map with textures and no .shader files to apply them with — which is why publish.py takes --art-from as well.
• The classifier needs no new rules for any of it: it already keys on map_<name>/ anywhere in a path, so dds/textures/map_boil/* goes to boil's archive and dds/textures/exx/* to the shared one.

compress-textures.sh wraps the vendored cached-converter.sh, which does the part worth not reimplementing: for each texture it tries several DXT formats and picks by measured PSNR rather than by a fixed rule. Its cache is content-addressed, so editing one texture recompresses one texture. DDS rather than PNG is deliberate — the game hands these to the GPU still compressed, and re-encoding would enlarge them on disk and multiply VRAM by four to eight (restructure section 4.3).

env/ is excluded on purpose: skyboxes ship as JPEG, matching upstream's pack, which has dds/textures and dds/models and no dds/env.

compress-texture's nvcompress path mapped dxt3 -> -bc3 and dxt5 -> -bc5. The correspondence is DXT1=BC1, DXT3=BC2, DXT5=BC3, and -bc5 is ATI2/3Dc — a two-channel format for normal maps, so a dxt5 request produced a file with no usable colour. The PSNR picker would normally reject that in favour of dxt1, which is exactly why it stayed hidden: it degrades quality silently instead of failing. Fixed in the vendored copy and marked VORTEX DIVERGENCE; it is the only edit in that file.

Verified locally: file discovery (3,229 images), the --only filter, the wrapper's real q3map2 invocations through a stub, compile-map.sh's product handling, and publish.py merging all three inputs (994 files -> 638 shared, 7 per-map, 349 prefab sources correctly dropped).

Compression: verified on CI, and the numbers validate it. The textures job succeeded on run 30506435276: 3,207 textures compressed, 810 MB on disk, 486 MB as the uploaded artifact. That count is the validation worth noting — Xonotic's shipped xonotic-20230620-maps.pk3 contains exactly 3,207 .dds files at 794 MB raw. Independently reproducing upstream's DDS set to the file is a strong signal the pipeline is right. The PSNR picker selected dxt1 for the sampled cases, and the content-addressed cache populated under ~/.xonotic-cached-converter/.

Still iterating: the q3map2 build. CMake stops at the first REQUIRED miss, so each run surfaces one more configure-time dependency — GTK (fixed with BUILD_RADIANT=OFF), then OpenGL via tools/heretic2 (fixed with libgl1-mesa-dev). Expect one or two more rounds. Nothing downstream is blocked: maps-2026.07 (produced by the retired bootstrap) covers the frozen 0.8.6 set.