v0.1.0-alpha.4

Experimental pre-release. Second published release of the AIENG Workbench
line, following v0.1.0-alpha.3 (2026-08-10). The Python packages carry
version 0.1.0a3.

What changed

This release is almost entirely defects found by using the product, not by
reading it. Four rounds of dogfooding drove the documented agent paths end to end —
assembly authoring, the no-MCP fallback scripts, topology optimization, and
parametric editing — and every one of them was broken in a way the test suite
could not see. Thirteen real defects, plus the CAE pre-processing rewrite that
started the round.

The pattern worth stating up front, because it shaped the whole release: a
path the docs advertise but no test exercises is probably rotten. In each case
CI was green throughout.

Say the physics instead of hand-translating it

Setting up an analysis used to mean reading a digest of every face's normal and
area, picking ids by eye, and hand-writing four JSON patches with NSET names,
DOF ranges and direction vectors. Now it is one call in engineering language:

cae.setup_static {
  material: "Al6061-T6",
  fix:  "bottom",                                    # or "bolt holes" / "base_plate bottom"
  load: { at: "rib_main top", force_n: 500, direction: "-Z" }
}

Chinese wording works (底面, 螺栓孔, 向下). What makes it safe rather than
merely convenient: it echoes what it actually bound — face pointer, surface
type, area, normal, owning part — so a mis-pick is visible immediately; ambiguous
wording is refused with the real candidates listed, never guessed; a sloped
face resolves but is reported as inclined 32° from top; and force_n: 0 is
refused outright, because it would converge and report zero stress as a result.

The load case can also be recorded as a requirement (cae.author_load_case)
with acceptance criteria. It is resolved against the geometry at authoring
time, so unpinnable wording is caught while rewording is still cheap, and it is
executable (cae.apply_load_case) so the recorded requirement and what was
actually solved cannot drift apart. Criteria land in the package's existing
task/design_targets.yaml and come back as pass / fail / unknown — a
criterion the run could not measure never silently passes.

Thirteen defects found by dogfooding

Assembly authoring

• A bonded tie between faces 20 mm apart stayed solver-enabled and
  load-transferring — stiffer than reality, in the non-conservative direction,
  with needs_user_input: []. A joint across a gap cannot exist at any scale;
  it is now invalid and the existing disable gate fires.
• aieng.agent_context — the tool an agent reads every session — reported
  nothing about assemblies. It now carries a compact block, and a refused
  joint reaches the top-level warnings.
• 4 of 4 correctly-authored interfaces warned (ok: 0). Both rules were
  satisfied by construction rather than by defect. Now judged by area coverage.

The no-MCP fallback path

• Dead on command one — the runner re-implemented the backend's placeholder
  substitution and had been failing with a NameError since a second
  placeholder was added. Nothing exercised these scripts; only the doc that
  advertises them to agents with no other way in.
• --data-root, a documented flag, raised TypeError.
• A failed require() leaked an internal marker and a temp-file traceback
  instead of the promised structured design_rule_violation.

Topology optimization

• The chain could not see a setup authored by the key-free path, and did not
  fail — it substituted a textbook cantilever preset under status: "ok",
  so a real 500 N bracket was posed as someone else's beam and would have been
  written back as its geometry. Both dimensions now refuse honestly.
• The 2D idealization cannot represent plate bending at all (the projection
  plane is spanned by the two largest dimensions, so the load is always
  out-of-plane). It now says exactly that and points at 3D.

Parametric editing

• A constant that dimensions the plate and positions the rib was reported as
  scope: "local" — "the safe single-part edit" — because the stored feature
  graph predated the binder that would have caught it. The scope-risk gate reads
  the same graph, so the edit skipped confirmation and resized the plate. Scope
  is now re-checked against the live source and only ever widened.

Windows / MCP transport (from the same practice, earlier in the cycle)

• Real CAD/CAE through the stdio server was fully broken: a lazy heavy
  C-extension import deadlocked the DLL loader, and any child inheriting the
  protocol pipe as stdin blocked at startup.
• A 500 N load silently became 0 N (0.0 mm / 0.0 MPa reported as a result);
  a second ccx spawn site hung until the client's 1800 s timeout.
• The server now stays answerable during long calls — a liveness ping 2 s into a
  16 s build was answered at +14.12 s, now +0.01 s.

Accuracy and honesty

• Mesh accuracy is measured through the thinnest solid — the wall a bending
  gradient must resolve through — not the model bounding box. On the canonical
  plate-plus-rib bracket the old reading was 6× optimistic, in exactly the
  non-conservative direction quadratic elements exist to prevent.
• The solver preflight stopped blocking correct setups, and reads the bound
  physics back in engineering language so you can answer "what is set up here?"
  without running anything.

Install

Unchanged from alpha.3 — Docker image, or uvx from a tag pin; see
MCP_SETUP.md. PyPI publication is
still pending owner setup (#273); the git-pin and container paths are the
supported ones.

Honesty boundaries (unchanged)

Linear static / modal / buckling / steady-state thermal only. Assembly
connections are simplified proxies — no nonlinear contact, no bolt preload, no
friction. Topology optimization 2D is the solid path and 3D is experimental,
producing a mesh proxy rather than production CAD. Heuristic manufacturability
rules, not certification. Every output is review material for a qualified
engineer, and the workbench says so in its own responses.