Build the network model result on mikeio1d's topology layer

.github/workflows/docs.yml

@@ -27,7 +27,7 @@ jobs:
     - name: Set up uv
       uses: astral-sh/setup-uv@v6
       with:
-        python-version: "3.13"  # Need to use 3.13 for docs build due to pythonnet dependency not yet supporting 3.14
+        python-version: "3.14"  # mikeio1d needs Python < 3.15, so the docs build cannot go higher
         enable-cache: true
 
     - name: Set up Quarto
@@ -36,7 +36,7 @@ jobs:
         version: "1.8.27"
 
     - name: Install dependencies
-      run: uv sync --group dev --group docs --group networks
+      run: uv sync --group dev --group docs --group network
 
     - name: Build documentation
       run: just docs

.github/workflows/full_test.yml

@@ -38,7 +38,7 @@ jobs:
           enable-cache: true
 
       - name: Install dependencies
-        run: uv sync --group test --group networks --no-dev 
+        run: uv sync --group test --group network --no-dev 
 
       - name: Install pandas 2.x
         if: matrix.pandas-version == 'pandas2'
@@ -57,7 +57,7 @@ jobs:
       - name: Test
         run: just test
 
-  build-no-networks:
+  build-no-network:
     runs-on: ubuntu-latest
 
     steps:
@@ -71,7 +71,7 @@ jobs:
           python-version: "3.12"
           enable-cache: true
 
-      - name: Install dependencies (without networks)
+      - name: Install dependencies (without network)
         run: uv sync --group test --no-dev
 
       - name: Test

.github/workflows/notebooks_test.yml

@@ -24,7 +24,7 @@ jobs:
         python-version: "3.14"
         enable-cache: true
     - name: Install dependencies
-      run: uv sync --group test --group notebooks --group networks --no-dev
+      run: uv sync --group test --group notebooks --group network --no-dev
     - name: Test notebooks
       run: |
         uv run pytest tests/notebooks/

adr/010-optional-domain-dependencies.md

@@ -56,7 +56,10 @@ Installation: `pip install modelskill modelskill-network`
 
 **Open Questions:**
 - Should `modelskill[all]` install all optional model types?
-- How to handle version constraints for optional dependencies?
+- How to handle version constraints for optional dependencies? Answered for network
+  support by [ADR-013](013-network-topology-in-mikeio1d.md): the `network` extra names a
+  minimum mikeio1d, because the topology layer ships there. Network support requires
+  whatever Python that release requires.
 - Should optional dependencies be tested in CI for every commit or separately?
 
 ## Status Notes

adr/012-network-format-constructors.md

@@ -1,9 +1,21 @@
 # ADR-012: One Network Constructor per Modelling Product
 
-**Status**: Draft
+**Status**: Accepted, narrowed by [ADR-013](013-network-topology-in-mikeio1d.md)
 
 **Date**: 2026-08
 
+## Narrowed by ADR-013
+
+The constructors, the companion arguments, the extension tables and the coverage test are
+mikeio1d's. It replaced `from_mike` and `from_epanet` with one `Network.open` that reads
+the extension. Naming a constructor after the product that wrote the file is still the
+rule, and mikeio1d applies it.
+
+EPANET reach lengths now come from the `.res`, so no `.inp` is read.
+
+`NetworkModelResult` hands a path to mikeio1d. The refusal messages for `.out`, `.resx`
+and the formats without a fixture are written there.
+
 ## Context
 
 `Network` is built from result files read through mikeio1d, whose single `Res1D` class opens nine extensions across five products — MIKE 1D (`.res1d`), MIKE 11 (`.res11`), MOUSE (`.prf`, `.crf`, `.xrf`), EPANET (`.res`), SWMM (`.out`), Water Hammer (`.whr`), and `.resx`, which is shared by the last three. There is no per-format reader and no per-format constructor argument, so from mikeio1d's side all nine look alike. modelskill's constructor was named `from_res1d`, and its extension guard was briefly widened to accept everything mikeio1d could read — making the name promise one format while reading nine.

adr/013-network-topology-in-mikeio1d.md

@@ -0,0 +1,60 @@
+# ADR-013: The Network Topology Layer Belongs to mikeio1d
+
+**Status**: Accepted
+
+**Date**: 2026-08
+
+## Context
+
+`modelskill.network` had grown to roughly 630 lines of topology: the abstract node/reach/breakpoint types, a `Res1D` adapter, one constructor per modelling product, the `.resx` and `.inp` companions,
+tables of which extensions we refuse and why, a networkx graph carrying reach lengths and boundary edges, an alias map, and `find`/`recall`/`to_dataset` on top. `NetworkModelResult` uses five members
+of `Network`, two of them private, and never traverses the graph. mikeio1d's `experimental.to_networkx` converts the same files in 25 lines and ignores gridpoints.
+
+That leaves us on the far side of the line ADR-001 drew for mikeio, where we call `mikeio.read()` and stop, modelling no dfsu geometry and policing no format list. `Res1D` reads nine extensions across
+five products. Our tables decide which of the nine we accept, and a test fails our CI when a mikeio1d release adds a tenth. The fixtures those tables are checked against are copies of mikeio1d's own:
+`network.res1d`, `network_cali.res11`, `epanet.res/.resx/.inp`.
+
+## Decision
+
+mikeio1d gains an optional network module that builds and owns `Network`. modelskill requires it and consumes what it produces.
+
+| Owner | Pieces |
+|---|---|
+| mikeio1d | abstract types and `BasicNode`/`BasicReach`, the `Res1D` adapter, `Network.open`, the `.resx` companion, the extension policy tables, graph construction with its length and boundary semantics, the alias map, `find`, `recall`, `to_dataframe`, `to_dataset` |
+| modelskill | `NetworkModelResult`, `NodeModelResult`, `NodeObservation`, `ReachObservation`, matching |
+
+`NetworkModelResult` takes a `Network` the upstream module built, or a path it hands to that module. The module is an extra there, carrying networkx and xarray, so `to_dataset()` ships with the class.
+modelskill's `network` extra requires a mikeio1d release new enough to contain it.
+
+Original IDs become the only identifier a user handles: `NodeObservation.at` takes a node name or a `(reach, position)` pair, not an integer. The alias integers stay an internal index,
+because the ID space mixes names and break points and a tuple cannot be an xarray coordinate value. A saved comparer records only the original ID, so reloading
+does not depend on the numbering the installed mikeio1d handed out.
+
+The loader's output over six fixture loads was recorded before anything moved — graph edges with their lengths and boundary flags, the alias map, the dataframe, and every answer `find` and `recall`
+give. Those snapshots are the upstream module's acceptance test. Phase 1 landed as mikeio1d [#247](https://github.com/DHI/mikeio1d/pull/247), merged 2026-08-19. The snapshots pass there unchanged
+twice: against the code moved verbatim, and again after the two product constructors collapsed into `Network.open`.
+
+mikeio1d 1.4.0 is the first release carrying the module, and modelskill 1.4.0 requires it.
+
+## Alternatives Considered
+
+**Keep the layer here.** Defensible while the API is private. Costs a format matrix, an EPANET `.inp` parser and a graph contract for traversals we never perform.
+
+**Move only the constructors and companions**, leaving the graph and the abstract types here. Splits the format knowledge from the topology it produces, and leaves `Res1DReach` here as the single
+adapter for a plug point with no second implementation.
+
+**A separate `modelskill-network` package.** Rejected in ADR-010 for fragmenting the install. It would still own format knowledge that belongs with mikeio1d.
+
+**Ask mikeio1d to guarantee stable node numbering** instead of keeping integers out of our API. Puts a promise on someone else's release process, to protect a number users should not be handling.
+
+## Consequences
+
+- ADR-012 is narrowed: the constructors, the companion arguments, the extension tables and the coverage test become mikeio1d's. Naming a constructor after the product that wrote the file is still the
+  rule, and mikeio1d applies it.
+- The EPANET `.inp` companion and its parser did not move. mikeio1d reads EPANET reach lengths from the `.res`, so no `.inp` is read.
+- ADR-010's open question about version constraints for optional dependencies is answered for this feature: the `network` extra pins a minimum mikeio1d, and network support requires whatever Python
+  that release requires.
+- A hand-built network needs mikeio1d installed, since `BasicNode`/`BasicReach` move too. That costs a .NET dependency for users who touch no MIKE file, which only matters for tests and for a backend
+  nobody has written.
+- `at` refuses an integer with an error that points to the name. Accepting one in 1.4.0 would make the internal index part of the API, and hard to withdraw later.
+- Releases become coupled in one direction: a fix to network file reading ships on mikeio1d's schedule. A format mikeio1d adds no longer breaks our CI.

adr/README.md

@@ -30,7 +30,8 @@ Each ADR follows this structure:
 - [ADR-009](009-factory-pattern.md) - Factory pattern for type detection
 - [ADR-010](010-optional-domain-dependencies.md) - Optional dependencies for domain-specific model types (Draft)
 - [ADR-011](011-vertical-pre-extracted-columns.md) - VerticalModelResult ingests pre-extracted columns
-- [ADR-012](012-network-format-constructors.md) - One Network constructor per modelling product (Draft)
+- [ADR-012](012-network-format-constructors.md) - One Network constructor per modelling product (narrowed by ADR-013)
+- [ADR-013](013-network-topology-in-mikeio1d.md) - The network topology layer belongs to mikeio1d
 
 ## Contributing