diff --git a/NOTEPAD.md b/NOTEPAD.md
new file mode 100644
index 0000000..63a53b2
--- /dev/null
+++ b/NOTEPAD.md
@@ -0,0 +1,41 @@
+# Notepad
+
+Leftovers, deferrals and things noticed in passing. Nothing here blocks a
+release; the phase plan lives in the app repo at
+`docs-private/reference/roadmap.md`.
+
+## Open
+
+- [ ] **CHORE/COVERAGE — the bound surface is inventoried from the app alone,
+  but `assay` is a consumer too.** `tools/inventory.py` parses
+  `app/backend` for `from OCP.<mod> import ...`, which is why
+  `BRepPrimAPI_MakeSphere` is not bound: the app never builds a sphere.
+  `assay/scripts/gen_flow_fixtures.py` does, so it has failed at import
+  against the app's venv since the 10C cutover (2026-08-10) and nobody
+  noticed — its fixtures predate the swap. Either teach `inventory.py` a
+  second source root or keep a small hand-maintained addendum for
+  non-app consumers, then bind what it turns up. Logged from the assay side
+  in `assay/NOTEPAD.md`.
+
+- [ ] **DOC/GIL — the STEP/IGES rationale predates OCCT 8.0's thread-safety
+  contract and should cite it.** `docs/design.md`'s Inc 3 note explains why
+  the XSTEP readers and writers hold the GIL. 8.0 added a contract that
+  sounds like it changes this and does not: it is *one reader or writer per
+  thread, default parameter set*, which is a per-instance guarantee that
+  explicitly excludes mutated `Interface_Static` — and `app/backend/n3xd/
+  cad/import_.py` sets `xstep.cascade.unit` process-wide. IGES read is still
+  documented non-thread-safe. Releasing the GIL would also buy nothing: the
+  app parses imports in a resource-bounded subprocess (`cad/import_pool.py`)
+  with `import_max_concurrency = 1`. Worth a few lines so the question is
+  not re-opened from the release notes alone.
+
+- [ ] **CHORE/BUILD — `TKHelix` is compiled by the builder image but not
+  linked.** `occt/Dockerfile` builds ModelingAlgorithms, so
+  `libTKHelix.so.8.0.1` and the `HelixGeom_*`/`HelixBRep_*` headers are in
+  the image; `CMakeLists.txt`'s toolkit list simply does not name it. One
+  line plus a module whenever the thread-feature spike (roadmap 10E) wants
+  it — no image rebuild.
+
+## Closed
+
+_None yet._