Add a pytest suite covering the nhflotools code 09pwnmodel2 uses

[bdestombe]

The boundary conditions and the layer model of the active PWN model were untested. Of the thirteen entry points 01_pwnmodel2.py imports, only interface_elevation and the Tata wells had any coverage; the RIV, DRN, GHB, CHD and WEL boundaries, the chloride initial condition, the layer model and the output loaders had none.

This adds 126 tests over that whole closure, running in about four seconds.

Approach

Tests build their own DISV vertex grid (tests/util.py) rather than calling gridgen, so the unit tests need no executables, no network and no large datasets — and every cell area is an exact float, which is what makes first-principles expected values possible instead of pasted implementation output.

Live services (HHNK, REGIS, RWS, background tiles) are monkeypatched by name with the reason stated; everything else runs the real nlmod code, so the suite doubles as a compatibility canary for nlmod@dev. One mf6-marked test runs MODFLOW on a 3x3x2 model and checks budget closure plus an analytic head field h = (2x + y)/50, with deliberately unequal x and y coefficients so a mirrored or swapped axis in the .grb/.hds path fails where a symmetric field would survive. It fetches the executable via nlmod.util.get_exe_path(download_if_not_found=True), so it works on a fresh runner without a preinstalled binary.

File	Tests		File	Tests
test_nhflodata_contract.py	41		test_pwnlayers3_plot.py	10
test_postprocessing.py	17		test_nhi_chloride.py	8
test_pwnlayers3_layers.py	15		test_major_surface_waters.py	7
test_well.py	13		test_panden.py	7
test_polder.py	4		test_pwnlayers_get_top.py	3
test_mf6_smoke.py	1

Evidence the tests catch bugs

Every file was mutation-checked while being written: a plausible bug was introduced into the source, the test confirmed to fail, and the source restored. Among those caught:

• the panden conductance formula reverted to the pre-83365c7 resistance aggregation — a silent ~1e4x error (caught by two tests)
• the polder fallback conductance zeroed, and the northsea guard removed so the sea gets drains
• numeric_only dropped from the well median, and the split over sec_nput removed
• extrapolate_ds's in-place contract broken, leaving NaNs under the sea
• the IJsselmeer stage-override mask inverted (caught by three tests)
• the get_top_from_ahn nearest-donor metric axis-swapped, and the falsy-constant sea guard introduced
• the layer-split group top misidentified in _compute_thickness_ratios
• the budget threshold inverted and the budget key renamed, and the head field mirrored on the x axis (both via the real MODFLOW run)

Defects found, filed rather than pinned

Ten issues were found while writing these tests and filed as #60–#69, so none is entrenched as expected behaviour by a test asserting it. Notable ones:

• [postprocessing] check_budget_discrepancy inspects only the incremental budget, never the cumulative one #67 check_budget_discrepancy reads only the incremental budget frame, never the cumulative one — the transport model is transient, so a run whose per-step discrepancies each sit under the threshold can still accumulate past it and be reported as fine.
• [pwnlayers3] REGIS-completeness guard raises an inscrutable pandas error in the case it was written for #65 the REGIS-completeness guard raises pandas' Lengths must match to compare in exactly the missing-layer case it exists for, instead of the message naming the fix. Covered by an xfail(strict=True) test that becomes a visible failure once the guard is fixed.
• [major_surface_waters] chd_ghb_from_major_surface_waters always returns ts_sea=None #66 chd_ghb_from_major_surface_waters always returns ts_sea=None (the series does attach to chd, so models are unaffected — the return value and docstring are wrong).
• [postprocessing] concentration_mean is biased toward zero where concentration is NaN #60 concentration_mean is biased toward zero wherever concentration is NaN.

Other changes

• pyarrow becomes a runtime dependency — well.py reads a feather file through pandas, which needs it in production, not just in tests. It was missing.
• A test workflow keyed on the current nlmod@dev and nhflodata@main commits, so those moving branches cannot serve a stale environment (the failure mode that produced the get_isosurface(left=...) breakage).
• A README overview separating the eleven modules 09pwnmodel2 uses from the nine untested ones, so the maintained surface is obvious.
• TEST_PLAN.md recording the design, the mutation evidence, the deviations from plan, and the progress table.
• Removes tests/test_hhnk.py and tests/test_nhflo_utils.py, which contained no tests while suggesting hhnk.py and nhflo_utils.py were covered.

Notes for review

• Three planned tests were dropped rather than written: one was unreachable through panden's real code path (it would have asserted nlmod's own guard), and two would have pinned scipy internals. Recorded in TEST_PLAN.md.
• Branched off main and independent of Stop storing idomain in pwnlayers/pwnlayers3 layer models #59 — the suite was verified green against both main's source and Stop storing idomain in pwnlayers/pwnlayers3 layer models #59's.