[rajeeja]

Replaying the notebooks in UXARRAY/uxarray-community
through the server found capabilities a user reaches for that we cannot run.
Looking for the rest found twelve, and no record anywhere of which absences
were decided and which were simply unnoticed.

The inventory

domain/uxarray_surface.py lists every UXarray-specific method on
UxDataArray, UxDataset and the subset accessor, each with either the
operation that reaches it or the reason it is out of scope. The
xarray-inherited surface is excluded on purpose: ~150 methods that are not
UXarray's to change and would drown the signal.

tests/test_uxarray_surface.py holds it to the installed library:

• a new upstream method fails until someone writes a verdict;
• a renamed or removed one fails the day it lands, rather than the next
  time a user follows our advice into a method that is gone;
• every route the inventory claims must exist — an inventory claiming a
  route that was never wired is worse than one with a visible gap.

Deliberately not a count. A test asserting "we expose N of M methods" goes
red on every upstream release for reasons nobody can act on, and the repair is
to edit the number — which teaches people to stop reading it. The weekly floor
job spent six Mondays red for exactly that reason.

Three gaps closed — as new tools, not in run_analysis

Tool	Closes	Notebook uses
subset_advanced	subset.bounding_circle, subset.nearest_neighbor	4 each
regrid_elements	topological_mean and 6 siblings	3

regrid_elements is the step that moves node-centred vorticity onto faces
before anything else can touch it — the cyclone notebook opens with it.

An empty selection is refused with the mesh extent, the rule
test_subset_coverage.py already enforces for bounding boxes. UXarray's own
message says the circle was empty without saying where a circle would not be:

Selection kept no faces: circle centred at (0.0, 0.0) with radius 30.0
degrees. The mesh covers {'lon_min': 40.0, 'lon_max': 46.0, ...}. Move the
centre inside that extent, or widen the radius.

Why separate tools: #89 measures run_analysis at 46 parameters over 25
operations and names that as the problem. Adding a radius, a neighbour count, a
centre and a destination element to it would have made the measured problem
worse in order to fix a different one. Each new tool takes only what its own
operation uses, and both are deferred.

Two honesty fixes

get_capabilities advertised things it cannot run. It lists native UXarray
calls — grid.get_ball_tree(), var.integrate(), grid.subset.bounding_circle()
— all valid UXarray, none executable here, and nothing said so. A model reading
that plans a call that does not exist; the same defect class as a wrong tool
annotation. The result now carries uxarray_capabilities_are_advisory, naming
mcp_server_tools as the executable surface.

The face-centred refusals were a dead end. Variable 't_node' is not face-centered is correct, and the standard repair is topological_mean("face")
— which the server could not run, so the message had nothing to offer. It now
names regrid_elements, and a test follows that advice to confirm it works
rather than landing on a second error.

Verification

• 1538 passed, 1 skipped, 44 deselected, 1 xfailed
• EXPECTED_DEFERRED 33 → 35; that counter guards the namespace plan against
  drift, which two new deferred tools genuinely changes.
• Over stdio on deferred-full: both tools run through call_deferred, the
  circle and nearest selections return the expected face counts, regrid_elements
  moves ['n_node'] -> ['n_face'], and the empty-circle refusal carries the
  mesh extent.

Still open, now recorded rather than unnoticed

weighted_mean, neighborhood, to_raster, and the two
constant_*_interval selections are marked gap with reasons. get_dual,
scalardotgradient, topological_prod and the boolean reductions are
out-of-scope with reasons.