Build the network model result on mikeio1d's topology layer

[jpalm3r]

Phase 2 of ADR-013. mikeio1d reads network result files and builds the graph. modelskill keeps the model result, the observations and the matching. Phase 1 was mikeio1d #247, released in mikeio1d 1.4.0.

Removed

• src/modelskill/network.py (1200 lines), model/adapters/_res1d.py (189) and _inp.py (109), with their tests.
• The extension table and its CI test. That test failed whenever a mikeio1d release added a format.
• NetworkModelResult.nodes, .data and .time. .nodes published graph integers, which ADR-013 says users should not handle. .data and .time needed the whole network read up front. .period gives the time span from the file header.
• at=<int> on NodeObservation. It shipped in the 1.4.0a3 alpha only. An integer now raises and points at network.to_networkx().nodes[<int>]["address"].
• NodeModelResult(data, node=..., name=..., item=...). It takes only a dataset that already carries its location, which is what extract returns.
• The epanet.inp fixture. mikeio1d reads EPANET reach lengths from the .res, so nothing reads an .inp.

API

mr = ms.NetworkModelResult("model.res1d", item="WaterLevel")
obs = ms.NodeObservation(data, at="node_A")
cc = ms.match(obs, mr)

• NetworkModelResult takes a path or a mikeio1d.network.Network. A path goes to Network.open, which picks the reader and finds companion files where the format needs them.
• ms.model_result() builds a NetworkModelResult from either, by type or by .res1d, .res11 or .res extension.
• A NodeObservation names its location by node name or by (reach, position). position_tol sets how far a position may be from a break point; by default it absorbs rounding only (mikeio1d's 1e-3). from_multiple takes either form as a key, and position_tol.
• ReachObservation.from_multiple is new.
• NodeObservation, ReachObservation, NodeModelResult and Comparer have .at: a node name, a (reach, position) break point, or a reach name. Comparer.at is None outside a network. Comparer also has .reach and .position.
• A comparer sits at its observation's location, and so do its raw model results. Where a model was read is mr.extract(obs).at.
• The break point coordinate is renamed distance → position. node, reach and position are reserved names.
• Observation, model result and comparer reprs print the location. A quantity with no unit prints without empty brackets.
• The extra and the dependency group are renamed networks → network, matching mikeio1d's. Both require mikeio1d[network] >= 1.4.0.

Reading

Building a NetworkModelResult reads no timeseries. extract resolves the observation's location with Network.resolve and reads that location only.

A node observation is resolved with the model's item, so a position snaps onto a break point that carries it. A location that exists but lacks the item gets a "has no data for quantity" error, not "not found".

A ReachObservation reads the break points on its reach that carry the quantity, in one read. They must agree. The one at the lowest position is used.

Chainage tolerance, location lookup and format refusal messages are mikeio1d's.

Bugs fixed

Each has a test that fails on main.

• Comparer.load coerced the node coordinate with int(). A comparer built from at="node_A" could not be reloaded.
• A reach comparer was saved without its raw model data. Loading it raised NotImplementedError: Unknown gtype: reach.
• NodeObservation(df, at=("r1", 24.5)).to_dataframe() raised. It dropped a node coordinate a break point does not have.

An observation built from data that already carries a location now raises if at or reach names another. Before, the argument was dropped silently.

Stacked PRs

Closed unmerged. Each change landed here or in mikeio1d:

PR	Where it is now
#685 quantities filter, shared empty frame, once-per-node read	mikeio1d #247: Network.open(quantities=), _EMPTY_DATA and the node_data cache in network/_res1d.py. Pinned by the res1d_one_quantity snapshot and TestQuantityFiltering
#694 reach-end gridpoints promoted, boundary removed	mikeio1d #247: _build_reach_breakpoints, and the clamped zero-length boundary=True edge in network/_graph.py. Pinned by TestAReachThatDoesNotStartAtZero
#695 EPANET link quantities	mikeio1d #247: the synthetic gridpoint duplicated to both ends, Address = str | tuple[str, float], and .resx reach quantities merged onto break points. Its two modelskill-side guards are superseded by _extract_reach and Network.resolve
#698 MIKE+ lookup	Split. The MIKE+ half is on the mikeplus-station-lookup branch. ReachObservation.from_multiple and the address keys on from_multiple are here. The to_dataset quantity fix is moot: NetworkModelResult takes its quantities from Network.quantities and no longer calls to_dataset
#699 path constructor	Here, rewritten onto Network.open. The extension policy, companion discovery, the cp1252/latin-1 name repair and the CI coverage test are mikeio1d's _policy.py and _companions.py

mikeio1d's code answers both review threads on #695. #679, #680 and #599 are closed separately. #682, #683 and #684 moved to mikeio1d.

Checks

791 passed, 5 skipped, 1 failed locally. ruff check src, ruff format --check, mypy and the doctests in metrics.py and types.py are clean. CI is green. The user guide renders. The notebook runs end to end against mikeio1d 1.4.0. CI does not run it, since it is in SKIP_LIST.

The failure is tests/regression/test_regression_rose.py, which fails on main too. The lockfile pins matplotlib 3.10.9, and the baseline image came from a newer one.

The network tests and the docs build run on Python 3.14. mikeio1d requires < 3.15.

Notes for review

• The user guide went from 587 lines to 161. It keeps the skill assessment workflow, and shows where a reach comparer sits and what position_tol does.
• The MIKE+ station lookup lives on mikeplus-station-lookup, which is this branch plus that feature. from_multiple(db=...) is public API and would otherwise ship with 1.4.0.
• Loading a network comparer saved by 1.4.0a3 is not supported, and its fixture is gone.
• The fixtures for formats whose tests moved are kept. network.res1d and the EPANET .res/.resx are read here.
• Closes Hide internal network integer IDs from public API where possible (future improvement) #639
• Closes Include tolerance as parameter in find #640

🤖 Generated with Claude Code

[ecomodeller]

Nice work. Moving the topology and format handling to mikeio1d removes about 1,500 lines here, and what is left is small and easy to follow: building a NetworkModelResult reads nothing, extract reads only the locations the observations name, and users only ever handle node names and (reach, position) pairs. Reading location from the coordinates in one place (network_gtype, NETWORK_LOCATION_COORDS) fixes the save/load bugs where they came from, and the tests now run against a real result file instead of a hand-built network.

One correction to the description: Collection_systems_network is now skipped only on Python ≥ 3.15, so the notebooks workflow (Python 3.14) will run it, either on the push to main or through a manual run of that workflow. It has not run on this PR yet.

Optional follow-ups, none of them needed for this PR:

• network_location and _at_from_coords could become one function in timeseries/_coords.py that coerces to str/float. Then comparison and model/network would no longer import a private helper from obs.
• The node branch of NetworkModelResult.extract could move to _extract_node, next to _extract_reach.
• network_cali.res11 and swmm.out are no longer read here and could be removed in a separate PR.
• test_init_fails_with_unsupported_type can assert TypeError alone now.