Update the architecture diagram to match the code

CHANGELOG.md

@@ -2,6 +2,7 @@
 
 ## [4.18.17RC] - 2026-06-03 Unreleased in PyPI
 
+- [CHANGED] update the architecture diagram (`doc/img/openTEPES_architecture.svg`) so it matches the code. The old picture used the planned folder names (`io/`, `schema.py`, `solver/`, `solve.py`); it now shows the real flat module names (`openTEPES_InputSchema.py`, `openTEPES_ProblemSolving.py`, and so on), with the five real solver modules. Implemented modules are shaded green and planned ones are left white, with a small legend. Also commit a rendered `doc/img/openTEPES_architecture.png` and point the `README.md` at the PNG instead of the SVG, so the diagram shows on pages that do not display SVG (such as the PyPI project page). Added a short `doc/img/README.md` explaining the SVG is the hand-drawn source (there is no generator script), that the PNG is rendered from it, and how to regenerate the PNG on Windows, macOS and Linux.
 - [FIXED] a binary investment problem (for example `IndBinNetInvest=1`) crashed under the HiGHS solver with `NoDualsError` on the first solve. `ProblemSolving` attaches the `dual` Suffix before the first solve only for a pure-LP model, because a mixed-integer problem has no duals; it then recovers the duals later by fixing the integer variables and re-solving as an LP. The check that decided "is this a pure LP" also required each variable to already have a value, but before the first solve no variable has a value yet, so a model with binary *investment* variables was wrongly treated as an LP, got the Suffix, and crashed when HiGHS was asked for duals it does not have. The check now looks only at whether a variable is integer/binary and unfixed, not at its value. Unit-commitment models did not hit this because their binary variables are given starting values. No change for any model that already worked.
 - [ADDED] a test (`test_binary_investment` in `tests/test_run.py`) for a binary (integer) investment decision, which no other test covered — every other case solves the investment variables as a continuous relaxation. It switches on `IndBinNetInvest` for the `9n` case so the single candidate line becomes a {0,1} build-or-not decision; the cost (254.337 MEUR) differs from the continuous result (252.201 MEUR) because the binary decision forces a full line build. The test also guards the fix above. A companion fixture `case_7d_binary` runs the case from a private temporary copy (so its solver log files do not clash with the other 9n tests on Windows, where a log file can stay open after a solve), applies the 7-day truncation, and overrides one or more columns of the Option file.
 - [CHANGED] split the CI workflow (`.github/workflows/ci.yml`) into two jobs to save time. A `fast` job runs the linter and the tests that do not solve a model, on all three operating systems and all three Python versions (3.11, 3.12, 3.13) — this is where import, packaging and Python-version problems show up. A `solve` job runs the full model test suite (every case, the multi-stage case, and Benders) once per operating system on Python 3.12, since the model results are the same on every Python version. Tests that solve a model are now marked with `@pytest.mark.solve` (registered in `pyproject.toml`); the `fast` job runs `-m "not solve"` and the `solve` job runs `-m solve`. Also added a per-test timeout on the solve job so a stuck solver fails quickly instead of using up the whole job, and turned on dependency caching for `uv`. No test was removed; the same tests still run, just spread across the two jobs.

README.md

@@ -29,7 +29,7 @@ It has been used by the **Ministry for the Ecological Transition and the Demogra
 
 The package is organised in six layers, from input/output (pure pandas, no Pyomo) up to result aggregation. Each module encodes its layer in the file name alongside the other `openTEPES_*.py` modules at the package root — `openTEPES_Input*` for the input-source layer (`openTEPES_InputSchema`, `openTEPES_InputSource`, `openTEPES_InputCSVSource`, `openTEPES_InputDuckDBSource`) and `openTEPES_ProblemSolving*` for the solver layer (`openTEPES_ProblemSolving`, `openTEPES_ProblemSolvingBenders`, `openTEPES_ProblemSolvingDualExtraction`, `openTEPES_ProblemSolvingPersistent`, `openTEPES_ProblemSolvingTuning`) — so concerns are addressable in code and the parallelisation modes (per-case sweep, in-memory overlay, post-build hot-swap) become first-class architectural seams.
 
-![Architecture diagram](/doc/img/openTEPES_architecture.svg)
+![Architecture diagram](/doc/img/openTEPES_architecture.png)
 
 # How to Cite
 

doc/img/README.md

@@ -0,0 +1,30 @@
+# Architecture diagram
+
+Two files make up the architecture diagram shown in the project `README.md`:
+
+- `openTEPES_architecture.svg` — the source. It is hand-drawn SVG (the boxes and
+  text are plain SVG elements) and is edited directly; there is no script that
+  generates it. Open it in any text or vector editor, change what you need, save.
+- `openTEPES_architecture.png` — a raster copy rendered from the SVG. The project
+  `README.md` embeds the PNG, because some pages (for example the PyPI project
+  page) do not display SVG images.
+
+Modules drawn in **green** are implemented (merged upstream); modules drawn in
+**white** are planned and their names are indicative.
+
+After editing the SVG, regenerate the PNG so the two stay in step. Use a vector
+tool such as [Inkscape](https://inkscape.org), which runs on Windows, macOS and
+Linux:
+
+Windows (Command Prompt or PowerShell):
+
+```
+inkscape openTEPES_architecture.svg --export-type=png --export-filename=openTEPES_architecture.png -w 1710
+```
+
+Linux / macOS (Inkscape, or `rsvg-convert` from librsvg):
+
+```
+inkscape openTEPES_architecture.svg --export-type=png --export-filename=openTEPES_architecture.png -w 1710
+rsvg-convert -w 1710 openTEPES_architecture.svg -o openTEPES_architecture.png
+```