[colbyt]

Summary

Implements Decision 6 in docs/design_decisions.md ("derived artifacts are
generated, not hand-maintained — and kept honest by tests"), specifically its
"still partially built" note that framing math lived in more than one place.

Framing-lumber positions (studs, plates, kings, jacks, headers, cripples,
sills, blocking) now have one home: enumerateMembers() in
web/js/members.js.

• scripts/export_members.mjs (new) bakes its output to a committed
  web/assets/lib/members.json, with a --verify mode mirroring the
  existing gen_wall_instances.py --verify pattern.
• seh_lib/wall_builder.py — shared by generate_wall_library.py and
  every library/modules/<id>/compiler.py — now builds FreeCAD box shapes
  from members.json via a new boxes_from_members(), instead of
  independently recomputing stud/cripple positions, king/jack heights,
  header/sill/subheader/blocking placement. boxes_from_members() is pure
  Python (no FreeCAD import) and is unit-tested without FreeCAD.
• build_lib.py bakes members.json before the freecadcmd geometry step
  (both in-place and in --verify).
• CI (ci.yml) gets a node scripts/export_members.mjs --verify freshness
  check on every push; geometry-verify.yml (FreeCAD BREP-verify, the
  mandatory gate this refactor can't check locally — see below) gets the new
  files added to its path filter plus a setup-node step, since
  build_lib.py --verify now shells out to node.
• The parity tests grow a matching pair for the new artifact:
  tests/members_export.mjs (JS: members.json == live
  enumerateMembers() for every module) and tests/test_members_boxes.py
  (Python, FreeCAD-free: the same acceptance numbers
  tests/enumerate_parity.mjs already asserts on the JS side, re-asserted
  against what boxes_from_members() produces from the identical
  members.json). Object naming (king_stud_left/right,
  jack_stud_left/right, etc.) is preserved verbatim since
  library/modules/*/expect.yaml pattern-matches on those names for
  libtools validate-output.

Deliberately NOT touched: OSB sheathing

The JS/3D side models exterior OSB as several rectangular strips split
around a rough opening; the Python/FreeCAD side still cuts a single full
sheet with a rectangular hole. This is a real, pre-existing geometry
difference between the two implementations — already tracked as
CAD-AUD-005 and recorded as a known_issue on every aperture module's
meta.yaml. members.json still carries sheathing role entries (so JS
consumers are unaffected); boxes_from_members() just skips them. This
refactor does not silently pick a winner between the two OSB approaches —
consolidating OSB generation is separate follow-up work, noted in
docs/design_decisions.md.

No dimensional numbers changed. Every acceptance number in
tests/enumerate_parity.mjs (studs at 0/16/32/46.5in, window king/jack/
header/sill/cripple dimensions, etc.) is unchanged and re-verified from the
Python side in tests/test_members_boxes.py against the same
members.json.

Verified locally (Node 26, Python 3.12, no FreeCAD on this machine)

• All JS tests in ci.yml's "JS test suite" step, including
  tests/parity.mjs and tests/enumerate_parity.mjs (both still pass
  unchanged — 4/4 and 9/9), plus the new tests/members_export.mjs (17/17).
• node scripts/export_members.mjs --verify passes.
• python -m pytest tests/test_blocking_math.py tests/test_members_boxes.py
  (26/26). tests/test_library_drift.py and tests/test_generated_entry.py
  could not run locally — they import libtools, which requires a network
  pip install from git+https://github.com/OpenSourceEcology/vcs-library
  not available in this sandbox; unrelated to this change.
• Static web smoke test: python3 -m http.server over web/, confirmed
  index.html, js/main.js, assets/lib/specs.json, and the new
  assets/lib/members.json all serve 200. web/js/fcstd.js is untouched
  (git diff confirms zero changes to that file).

What CI must still prove (deferred — no FreeCAD on this machine)

geometry-verify.yml's python3 build_lib.py --verify --no-thumbs is the
mandatory gate: it must confirm that regenerating web/assets/lib/*.brp,
volumes.json, and cad_library/*.FCStd from the refactored
seh_lib/wall_builder.py (now reading members.json instead of
recomputing framing positions) produces geometry identical to what's
currently committed, for all 15 modules. Since boxes_from_members()
reproduces the exact same box positions the old inline math computed (same
constants, same coordinate convention, verified above against the JS
acceptance numbers), no .brp/.FCStd changes are expected — but this PR
does not (and, without FreeCAD locally, cannot) regenerate or touch those
binary assets itself. libtools validate-output (also in that job) is the
check that confirms object naming still matches each module's
expect.yaml role patterns.

Not merging

Opening for review — not merging this myself.