-`views/canvas3d/` (**PyVista/VTK**, default) and `views/canvas_plotly/` (plotly.js in a `QWebEngineView`) are two backends for the same central 3D view. Both satisfy the `CanvasBackend` protocol in `views/canvas_base.py`, share one `SelectionState` owned by `MainWindow`, and are swapped live via **Options → Canvas Backend** (persisted in `QSettings` under `canvas/backend`). `MainWindow._activate_canvas` keeps both widgets in a `QStackedWidget` — never destroy a canvas mid-session (VTK leaves dangling make-current callbacks). Backend-specific gaps are declared by `CanvasCapabilities` (e.g. Plotly has no force-diagram overlay or video export yet); gate UI on `canvas.capabilities`, never on the backend name. `canvas_plotly/trace_builder.py` is pure (no Qt, no pyvista) and unit-tested headless. `PlotlyCanvas` pushes with `Plotly.react`, which resets scene attributes the layout omits: data-only pushes pass `preserveView` so `html.py` carries the live camera and padded axis ranges forward, and the hover/snap marker is restyled only when its visible state changes (a restyle per hover event redraws per mouse move until the stack blows). Regression tests: `tests/gui/test_plotly_view_preservation.py`, `tests/gui/test_plotly_hover.py`.
-`views/canvas3d/style.py` (`RenderStyle`) is the single source of truth for colours/sizes on **both** backends; renderers must read it rather than hard-coding a colour. The `STYLE_FIELDS` subset is user-editable via **Options → Plot Properties…** (`views/dialogs/plot_properties.py`), previews live through `MainWindow.set_plot_style(..., persist=False)`, and persists as JSON in `QSettings` under `plot/props`.
- Rules: public functions need type hints + docstring; new domain entities go through Pydantic validation; ops >50 ms run off the GUI thread (`AnalysisWorker` in QThread, cancel via `isInterruptionRequested()`, results cross threads as lightweight `ResultsHandle` to HDF5).
Runner emits OpenSeesPy commands in fixed order (`docs/architecture.md`): `wipe → model → node → fix → material → section → geomTransf → element → timeSeries → pattern/load → recorder → system/numberer/... → analyze`. Never reorder.
## Verify (in this order)
```bash
ruff check src tests
ruff format src tests # line-length 100, E501 ignored
mypy src/otko/core src/otko/services
pytest -m "not slow" # CI gate: lint → this, on 3.10/3.11/3.12 × ubuntu/windows/macos
```
Focused runs: `pytest tests/unit` (pure logic, ms), `pytest tests/gui -k <name>` (pytest-qt, needs display), `pytest tests/integration -k <name>` (real OpenSeesPy runs). Single test: `pytest tests/unit/test_project.py::test_name -q`. Markers: `gui`, `slow`. Linux GUI tests need `xvfb-run -a pytest ...` plus system Qt libs (see `ci.yml` apt list).
Notes: `tests/conftest.py` auto-`ops.wipe()`s the OpenSees domain between tests (lazy import so core tests stay Qt/OpenSees-free). Coverage omits `views/`. Commit style: Conventional Commits (`feat:`, `fix:`, ...). `pre-commit install` runs ruff + mypy (mypy hook scoped to `core|services|viewmodels`).
- Web UI / HTTPS clone: `https://smill-home.ddns.net/forgejo/smill/otko.git` (note the `/forgejo` subpath — SSH URLs don't have it).
- Forgejo SSH runs through the system sshd on **port 22** (`SSH_PORT = 22` in the server's `app.ini`). If the site ever advertises a `:2222` SSH URL again, that config regressed — push/pull with port 22 anyway and flag it.
- Auth is by SSH key (Forgejo → Settings → SSH Keys). Test: `ssh -T git@smill-home.ddns.net` should answer `Hi there, <user>! ... Forgejo does not provide shell access.`
- Default branch `main`. Work on short-lived branches (`fix/<topic>`, `docs/<topic>`), open a PR against `main`. Never commit directly to `main` from an agent session.
- Repo root is `otko/` itself. (On the dev laptop there is also an empty, commit-less parent repo at `Sync/coding/` — ignore it; all git work happens inside `otko/`.)
-`examples/*.py` are source of truth; `examples/*.osmodel` are generated artifacts (checked in). Never hand-edit `.osmodel` — change the script and regen: `python examples/cantilever.py` (each script saves, reloads, asserts clean round-trip).
- Projects persist as single Pydantic-validated JSON `.osmodel` (diffable); analysis output goes to `*.osresults.h5` (HDF5, one group per case).
- Quick smoke: open `examples/cantilever.osmodel` → run `Tip-Load` → M3 peaks 50 kN·m at fixed end.
