RFD 1060 details: the second move, why a real move, what stays, paths, verification

The symlink attempt, and why it failed

The first attempt at moving src/library/ left a symlink at src/library. A working session found that symlink broke relative imports inside the moved files. Vite resolves a symlink to its real path before applying a relative import. So thirdparty/library/aiModelsCatalog.js’s import '../core/domain/catalog.js' resolved against thirdparty/core/, which does not exist, instead of src/core/, which does. resolve.preserveSymlinks: true in vite.config.js and vitest.config.js fixed it once. Then a direct instruction reversed the symlink call: move the real files.

The second move, in full

weftspun_studio/ becomes the repository root. Everything that was at the root — src/, package.json, vite.config.js, index.html, public/, native/, tests/, and the JS toolchain’s own scripts/*.mjs files — moves to thirdparty/3d_studio/. decisions/, .github/, and a small, named set of root-level scripts/ stay at the root, alongside weftspun_studio/’s promoted lib/, test/, config/, mix.exs, Dockerfile, and deploy/.

New work happens at the root: in lib/, in thirdparty/3d_studio/src/core/, or in thirdparty/3d_studio/src/chain/. It composes with thirdparty/3d_studio/ through a port, per RFD 1023, or it does not touch it.

Why a real move, not a rewrite of every import

RFD 1057 says a sweep is the wrong shape for a large, low-value change. “Convert one when its model is next worked on, and not in a sweep,” in RFD 1057’s own words. RFD 1023 says the same rule for src/library/. “A module moves when a port covers it, and not before.”

Neither RFD 1023’s per-file rule, nor RFD 1057’s no-sweep rule, required rewriting relative imports inside the moved trees. git mv preserves history. Every existing import between files that moved together keeps its shape. Those files moved as one tree, not as 139 and then several hundred independent edits.

What stays at the root, and why

decisions/ — direct instruction. The RFD index documents both weftspun_studio and the browser client. Splitting it would break RFD 1000’s DRY policy.

.github/ — GitHub Actions only discovers workflows at the true repository root. It cannot move.

.devcontainer/ — RFD 1056’s dev container, for the Elixir side. devcontainer.json sets privileged: true and --cgroupns=host specifically so Quadlet units can start inside it. Removed in this session, ahead of a rebuild.

Six scripts/ files, at the time of this move — ci.sh, deploy-weftspun-quadlet.sh, studio-test.sh, check-elixir-parses.exs, ste-lint-decisions.py, check-model-images.py. .pre-commit-config.yaml named four of these by path, and pre-commit hooks run from the repository root.

Every other file under the old scripts/ moved to thirdparty/3d_studio/scripts/ instead. That set is the JS build helpers, the DGX sync scripts, and the XR proxies. Those files read and write that tree, not this one. RFD 1063 later deleted ste-lint-decisions.py, leaving five.

LICENSE, TRADEMARKS, .formatter.exs, .gitattributes, .pre-commit-config.yaml — repository-wide, not app-specific. .formatter.exs now merges two input globs that used to be two files: decisions/**/*.{ex,exs} (from the root’s old file) and {mix,.formatter}.exs, {config,lib,test}/**/*.{ex,exs} (from weftspun_studio/’s).

Paths this move touched

WeftspunStudio.JsCatalog.@default_path, test/catalog_parity_test.exs, and test/fact_store_test.exs read thirdparty/3d_studio/src/library/aiModelsCatalog.js now, in place of ../src/library/aiModelsCatalog.js.

deploy/quadlet/weftspun.build and weftspun-crdb.build read /opt/weftspun/src/Dockerfile and /opt/weftspun/src/deploy/Dockerfile.crdb now, in place of the weftspun_studio/ prefix RFD 1058 wrote. scripts/deploy-weftspun-quadlet.sh still syncs the whole repository to /opt/weftspun/src, and that tree’s root is this repository’s root.

package.json’s lint:ste script read ../../scripts/ste-lint-decisions.py and ../../decisions/0*/README.md at the time of this move, because package.json lives two levels under the tree those two paths were actually in. RFD 1063 later deleted that script and that npm entry.

.pre-commit-config.yaml’s studio-test hook matches ^(lib|test|config)/.*\.exs?$|^mix\.exs$ now, in place of ^weftspun_studio/.*\.exs?$.

Verified

mix compile and mix test ran clean from the new root, at the time of this move: 100 of 105 tests passed, 5 excluded. The 5 failures were a pre-existing architecture mismatch, unrelated to this move. deps/taskweft_nif/priv/libtaskweft_nif.so was an ARM aarch64 binary, and this box is x86_64. file on the .so confirmed it, and the failure was identical with or without the swap. RFD 1057 records the later fix. All 105 tests pass now.

npx vitest run from thirdparty/3d_studio/ resolves src/library/aiModelsCatalog.js’s import '../core/domain/catalog.js' correctly, now that library/ is a real directory under src/ again, and not a symlink elsewhere. src/core and prepareGlbForApiUpload.test.js together: 54 passed, the same 5 pre-existing glbCompress.js export failures RFD 1023 already records, and no import-resolution failure.

What this does not do

It does not fix the prepareGlbForApiUpload gap RFD 1023 already records. That gap sits inside thirdparty/3d_studio/src/library/glbCompress.js now, the same as it sat inside src/library/glbCompress.js before.

It does not touch RFD 1023’s src/core/ / src/chain/ split. Both trees moved together, under thirdparty/3d_studio/src/, with every import between them intact.