…ze, block save on errors

From the GRAPH_EDITOR_REVIEW + the robustness plan's P1:
- Removed the top-level window keydown listener that leaked one anonymous
  handler per component instance (multi-tab); folded the Esc-exits-poly-mode
  into onWindowKeydown (registered + cleaned up in onMount). (review #1)
- refreshCallArgs now goes through finalize() (newly exported from
  composition-graph) instead of a raw node-spread, so graph.edges/imports
  rebuild after a drift-refresh (no stale phantom wires / wrong orphan
  detection). (review #2)
- Save button disabled when emitted.validationErrors.length > 0 — can no
  longer save a graph with broken node refs that won't reload. (plan 1e)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Archive the 2026-06-01 K.63 tree-editor survey verbatim
(cad_authoring_patterns.archived-2026-06-01.md) and write a current
successor that (a) reconciles which 2026-06-01 recommendations we kept,
dropped, or REVERSED — notably §F3 'don't add a 2D node-canvas', which the
shipped GraphEditorPane inverts — and (b) answers 'what is our possibility'
for smooth/curved geometry: cadtrain is an OpenSCAD-class mesh-CSG /
polygon-sampling kernel (NOT BREP), so Tier-1 dense continuous-function
sampling is reachable today, Tier-2 sketch+spline+fillet needs one
sketch_repeat engine fix, Tier-3 true BREP is explicitly out of class.

Also add docs/plans/wavy-star.md — the concrete g_wavy_star (#2) +
g_star_smooth (#1) build plan that the smoothness investigation produced.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…mode works

calculateNormals(3, crease) returns an ALL-ZERO normal buffer on welded/revolved
meshes (g_shaft etc.), so flat shading worked (face-derived) but smooth shading
(flatShading:false → vertex normals) had nothing to shade with — "no shading".

- builder.ts: detect degenerate normals (normalsDegenerate / bakedNormalsDegenerate)
  and recompute. Indexed paths (manifoldToGeo) computeVertexNormals directly; the
  NON-indexed color-by-source + cutVC paths build smooth normals from the indexed
  topology first (indexedSmoothNormals) then scatter — a plain computeVertexNormals
  on non-indexed geo would be FLAT. Winding is outward-consistent (Manifold
  canonicalises it), so the recomputed normals face outward (verified: g_shaft
  side normals 100% radial-outward, 0 inward).
- bake-worker-core: bump KERNEL_VERSION (+nrm2) so stale zero-normal meshes in the
  client IndexedDB cache re-bake. (Server bake cache: cleared via /api/cache/clear.)

Verified: g_shaft renders smooth radial shading in smooth mode (client + server),
correctly lit (not inverted). Fixes TODO PARKED #1/#2.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Gradient-id collision across N /primitives PrimitiveSvgView instances;
fixed by per-instance id prefix (6736f50), surfaced by the
svg-emit/svg-camera modularization (474e8f5). TODO PARKED bug #2 done.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The engine already had it (compileSketch scaleX/scaleY + setSketchScale mutator +
emit that drops sx/sy at 1/1 → byte-identical); this adds the missing UI — a ⚙
button in the sketch tool rail + a popover with independent X(r)/Y(z) inputs +
Reset 1×1, writing literal ArgValues via setSketchScale. 2D canvas stays in design
space; scale flows downstream to bake. Build+test green; verified Y=2→sketch(...,1,2).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

User-supplied research consolidated INTO our stack (not a parallel system):
compact TS tool schemas (local-model lever, cloud already prompt-cached),
Claude-synthetic prompt→call pairs seeding the #27 corpus, few-shot injection via
RAG-P3 embeddings now + optional LoRA fine-tune (Qwen2.5-1.5B #2 / Mistral-7B)
later, Neo4j/Chroma parked (Rule 1). Ties to #27/#2/#1.
Research doc docs/research/local-fncall-synthetic-data.md + TODO #28.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…bLLM+MLC+XGrammar)

Consolidate user research on running a FINE-TUNED small model fully in-browser
into the AI research docs. New "In-browser deployment (WebLLM + MLC + XGrammar)"
section in local-fncall-synthetic-data.md (#28, the deploy half of the fine-tune
endgame): Qwen2.5-Coder-0.5B (4-bit <350MB, offline via Cache-Storage), MLC-LLM
AOT compile + host + custom-URL CreateMLCEngine, XGrammar CFG constrained to our
EDITOR_TOOLS/ArgValue syntax, zero-system-prompt-after-fine-tune, output ->
dispatchEditorTool on the in-memory Graph (no server). Honest risks flagged
(XGrammar = syntax not semantics; MLC compile is a real build step; ~350MB
hosting; 0.5B accuracy = conditional-GO spike; few-shot is the near-term win).
Cross-linked from web-llm-functionary.md (#2). TODO #2 + #28 pointered.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…l + data-residency

Merge worktree-agent-aef8bf37 — extends #28 with the "In-browser deployment
(WebLLM + MLC + XGrammar)" section (Coder-0.5B, MLC compile→wasm, XGrammar CFG
from EDITOR_TOOLS → dispatchEditorTool, zero-prompt-after-fine-tune, honest
risks) + TODO #2/#28 pointers. Plus a data-residency banner: prod restricts
sending prompt data out → local runtime is a REQUIREMENT, cloud = dev/build-time
only (memory ai_data_residency_local_first).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… / future)

SVTC-style backend picker in AiMenu, pluggable via ge-assist's postTurn seam: a
registry {id,label,kind,postTurn,needsKey} — Claude (cloud, key) / WebLLM (local,
no key = data-residency) / future entries. Keys handled SECURELY (Rule 15 —
env/Railway/settings, never pasted-to-disk); local WebLLM needs none. Component
of the AI master plan; extends #2. docs/plans/ai-backend-selector.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… reconcile

One self-improving, local-first tool-calling assistant: one registry (source of
truth) generates cloud schema / prompt / local XGrammar CFG, feeds one corpus
(parts+docs+simplified+feedback) into one multi-shot loop, grows from judged
turns, targets a local WebLLM runtime under data-residency. Sequenced P0-P4
roadmap (shipped-vs-todo marked). Reconciles #1/#2/#27/#28/#29 as phases:
plan-header banners, TODO umbrella + annotations, /plan task 905 (stale->done,
residual->P1), fix-errors sink superseded by #27, path/tool-count fixes.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

#1/#2/#27/#28/#29)

North star docs/plans/ai-master-plan.md: one function-library registry → generates
cloud/prompt/local-CFG schemas → one multi-shot loop on the live Graph ← one corpus
(parts+docs+simplified+feedback) → feedback grows it → local WebLLM runtime, under
data-residency. 5 AI plans banner'd as detail-sheets under it; TODO #0 umbrella +
#1/#2/#27/#28/#29 annotated as phases; stale /plan task 905 fixed (open→done);
fix-errors sink marked superseded by #27. Cross-linked the backend-selector (P3)
+ #31 stays a separate editor track.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…peat (#1)

/api/tf/compile re-fetched every composite dep on EVERY edit, each a
prod-proxied /api/primitives/source round-trip (the ~300ms 'compile' segment).
Dep sources are param-invariant, so cache them by id (module-level, 60s TTL,
negative entries too). A 🔄 rebake sends bust:true → bypasses the cache for a
fresh resolve. Measured on bw_open_hole: cold 378ms → warm 1-2ms → 🔄 312ms.
TODO #2 (parallelize the BFS frontier) noted in the endpoint header.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… invalidation

Logs the two TF-compile perf follow-ups after #1 (dep-source cache) shipped:
(a) parallelize buildCompositeMap's BFS frontier (N×latency → depth×latency),
(b) invalidate SRC_CACHE on part Save so a dep-source edit is instant across
parts (with the local-dev prod-proxy wrinkle documented).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…0ms → ~300ms (#2)

resolveDepColors (part-colors.ts) re-fetched every composite dep with a RAW
uncached fetch on EVERY /api/primitives/compile (colours computed outside the
script cache) — the ~940ms 'still being compiled' latency on the hidden 3D bake.
Now routes through the loader's exported fetchDepSource (30s TTL cache). GET
compile measured 0.94s → 0.30s warm; the client POST path skips the root fetch
too. TODOs logged: only-bake-active-tab (stop the hidden 3D-bake on the TF tab)
+ TF_WASM tab (C++→WASM builders for concealment).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…mic page

Files the exploration under the local-first AI umbrella, where the binding
data-residency constraint and the existing web-llm spike (#2) already live.

Framed as a spike, not a build. Three things the entry pins down so the work
doesn't start from the wrong end:

- LiteRT.js COMPLEMENTS #2 rather than replacing it — it is the right runtime for
  small task-specific models (classifiers, embedders, regressors), where an
  LLM-shaped runtime is the wrong shape and far too heavy.
- "Dynamic page" was unscoped, so the entry lists candidate surfaces best-first
  (semantic part search over the RAG corpus; /vocab inference triage; bake-cost
  prediction; wells element classification) and explicitly rules OUT silhouette
  retrieval: the cold-classification baseline already scored 17/18 with no
  retrieval, and CLIP embeddings collapsed on 12/18 default-param renders.
  Re-run that baseline before adding any model.
- Names the real blockers up front: app-wide COOP/COEP require-corp already broke
  a main-thread WASM fetch once (Manifold on /wells), weights must ship
  same-origin under the strict CSP, and it must be default-OFF, lazy, and gated
  on the active tab.

Deliverable is a written go/no-go with measured latency + bundle cost against the
non-ML baseline it has to beat. The API surface is recalled, not verified —
the entry says so, and says to check current docs first.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PsYaPm3zaeuWL3KrPdCR8V

… undo + pure part/kind/intent maps (#42b-B)

Decision #2: WSON is the single source of truth for well editing (in-place
mutation keeps doc identity stable, avoiding bake thrash). This lands the pure,
headless-tested core of that layer; the Svelte editor wiring is deferred.

WellEditCore — add session-batched undo (beginEdit/endEdit/session):
  a burst of edits in one focus session coalesces to ONE undo entry (SVTC's
  beginEditSession parity). Only the first mutation in a session snapshots; later
  ones mutate live (onChange still fires per edit) but add no history. Nestable
  (outermost pair bounds it); undo/redo/reset force-close the session. In-place
  restore + root-object identity preservation are unchanged.

well-part-map.ts (new) — baked-part-id → WSON reverse maps:
  - layerForPartId(id) → WellLayerFlags key, accepting BOTH the emit var name
    (OH_/CEM_/CSG_/COMP_, the real WellPart.id) and the bw_* registry id.
  - parsePartId + inspectorElementForPartId — 3D pick → inspector cross-link,
    resolving the 1-based var-name index back to the source element.

well-edit-intent.ts (new) — editor-intent glue:
  - kindToArrayKey(kind) → WSON array key (inspector onEdit routing).
  - addElementIntent(type) — tool-rail Add → concrete default element + target
    array; completion categories map to a registry tool_comp that resolves to
    that category. applyAddElement routes it through the undoable core.

Tests: +20 core session-batch cases, +33 across the two new maps (all mapping
helpers pinned against real samples/*.wson + the live registry). Full wells
suite 311 green; bun run build green; tsc clean on the changed files. No Svelte
files touched (browser-only, out of scope).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PsYaPm3zaeuWL3KrPdCR8V