Build the network model result on mikeio1d's topology layer

[jpalm3r]

Phase 2 of the move recorded in ADR-013. mikeio1d now owns the layer that reads
network result files and builds the graph; modelskill keeps the model result, the
observations, the matching and the MIKE+ lookup, and consumes what mikeio1d
produces.

Phase 1 was mikeio1d #247, where the
module was built and checked against snapshots of this loader's output.

What goes

src/modelskill/network.py (1205 lines), model/adapters/_res1d.py (189) and
_inp.py (109), with the tests that covered them. The extension table goes too,
and with it the test that failed our CI whenever a mikeio1d release added a
format.

What changes

NetworkModelResult takes a path or a mikeio1d.network.Network. A path goes to
Network.open, which picks the reader and finds the EPANET companion files.

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

Observations are addressed by the name the model uses, or by (reach, distance)
for a break point. at=<int> is gone: which integer a location gets depends on
the network and the mikeio1d that built it, so nothing a user writes should
depend on it. That signature shipped in the 1.4.0a3 alpha only.

Extraction reads the dataset and consults the topology only to explain a failure.
Node lookups go through Network.find, so there is one chainage tolerance rather
than a second copy here, and a failed lookup names the near misses.

The extra is renamed networks → network, matching mikeio1d's.

Bugs fixed along the way

Each has a test that fails on main.

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

Behaviour changes

• A location whose column is entirely NaN now counts as having no data on a
  reach. It used to be selected and produce a match with no points, so the
  observation quietly disappeared. Another break point on the reach is used
  instead.
• Break points with an unknown distance now take part in reach matching. That is
  EPANET read without its .inp, where no reach has a length.
• mr.network is the caller's object, not a deep copy.
• Chainage tolerance and every format refusal message are mikeio1d's now.

Does this block a lazy source?

No. Logged as mikeio1d #250.

matching.py uses m.name and m.extract(obs) on a NetworkModelResult and nothing else,
so a lazy source replaces the constructor and extract and leaves the rest of this PR alone.
at already takes a node name or (reach, distance). The location travels into the matched
dataset as scalar coordinates taken from the network, with node_index beside them as
provenance nothing reads back, so a source that never mints integers omits it.

NetworkModelResult.nodes is removed here. It published the integers to_dataset() hands
out, which ADR-013 says users should not handle, and it was the only member that could not be
answered without building the rectangle. Nothing in src/ or the docs read it. .data,
.time and .sel_items stay; DfsuModelResult.data is already a file handle, so they
promise nothing about materialisation.

One cost does not go away: a ReachObservation reads every break point series on its reach,
to check they agree before picking one.

Draft: two things block merging

1. mikeio1d has not released the module. [tool.uv.sources] points at its
   main branch. Both network entries carry a TODO to swap that for a version
   floor. ADR-013 holds modelskill 1.4.0 until mikeio1d releases.
2. Five open PRs to close. Speed up from_* and add a quantities filter #685, Promote reach-end gridpoints to breakpoints, remove boundary #694 and Reach EPANET link quantities via duplicated end breakpoints #695 touch almost only files this
   deletes, and their code already shipped in mikeio1d Plots directional axis #247. Locate network observations with a MIKE+ database #698 (MIKE+ lookup)
   and Build a NetworkModelResult straight from a result file #699 (path constructor) are carried here — Locate network observations with a MIKE+ database #698 cherry-picked, Build a NetworkModelResult straight from a result file #699
   rewritten onto Network.open.

Checks

769 passed, 5 skipped. ruff check src, ruff format --check, mypy and the
metrics doctests are clean. The user guide renders. The notebook runs end to end,
which CI does not check — it is in SKIP_LIST.

tests/regression/test_regression_rose.py fails, on main as well: the lockfile
pins matplotlib 3.10.9 and the baseline image was made with a newer one.

Notes for review

• The user guide went from 587 lines to 197. What stays is the skill assessment
  workflow and the MIKE+ section.
• _mikeplus.py was folded into obs.py as one class, _MikePlusStationResolver,
  next to its only caller. Its 23 tests were rewritten to go through
  NodeObservation.from_multiple(db=...).
• tests/testdata/node_comparer_1.4.0a3.nc is a file the alpha actually wrote. It
  pins that such a comparer still loads.
• The fixtures for formats whose tests moved are kept. Nothing here reads them
  except network.res1d.

🤖 Generated with Claude Code

[jpalm3r]

The MIKE+ station lookup is coming out of this PR and onto its own branch: #706.

It is unrelated to the mikeio1d move (ADR-013, Phase 2), and its schema assumptions —
the locationtype codes, the resitemname encoding, the file-name matching — have never
been read against a database MIKE+ wrote. #706 lists what to settle first.