`__repr__` strings name a private module and are not evaluable

[tehrengruber-ai]

__repr__ strings name modules that don't exist and can't be evaluated.

The convention in _atlas4py.cpp is to prefix reprs with _atlas4py. and print a spec dict, e.g. (all line refs against master, a8b3365):

type	repr	ref
Projection	_atlas4py.Projection({...})	_atlas4py.cpp:312
Domain	_atlas4py.Domain({...})	_atlas4py.cpp:320
Grid	_atlas4py.Grid({...})	_atlas4py.cpp:336
Spacing	_atlas4py.Spacing({...})	_atlas4py.cpp:342
Config	_atlas4py.Config({...})	_atlas4py.cpp:407
Metadata	_atlas4py.Metadata({...})	_atlas4py.cpp:603
Gmsh	_atlas4py.output.Gmsh()	_atlas4py.cpp:633
PointLonLat	_atlas4py.PointLonLat(lon=..., lat=...)	_atlas4py.cpp:300
PointXY	_atlas4py.PointXY(x=..., y=...)	_atlas4py.cpp:307

Ideally a repr should be a valid expression that reconstructs the object, or — where that isn't possible — should follow a convention that makes clear it isn't one. Currently neither holds, for three separate reasons.

1. The _atlas4py. prefix names a private module

__init__.py does from ._atlas4py import *, so the name a user actually has is atlas4py.Config. _atlas4py.Config(...) only resolves if you have separately imported the private extension module, which nothing tells you to do. The prefix should be atlas4py. (or omitted).

2. _atlas4py.output.Gmsh() names nothing at all

The class is registered as plain "Gmsh" (_atlas4py.cpp:626), so there is no output submodule and no _atlas4py.output anywhere. The correct expression is atlas4py.Gmsh(path).

3. eckit.Configuration is not reachable by attribute access

_atlas4py.cpp:376,378 pass dotted names to nb::class_:

nb::class_<eckit::Configuration>( m, "eckit.Configuration" );
nb::class_<eckit::LocalConfiguration, eckit::Configuration>( m, "eckit.LocalConfiguration" );

nanobind builds the full name as "_atlas4py" + "." + "eckit.Configuration" (nb_type.cpp:1353) and then splits it at the last dot (nb_type.cpp:929), so the class ends up with __module__ == "_atlas4py.eckit" — a module that does not exist — and __qualname__ == "Configuration". It is registered with setattr(scope, "eckit.Configuration", ...) (nb_type.cpp:1661), i.e. the module dict key literally contains a dot, so:

atlas4py.eckit.Configuration            # AttributeError
getattr(atlas4py, "eckit.Configuration")  # the only way to reach it

Since the key doesn't start with an underscore, import * also copies that unusable name into the atlas4py namespace. Either give these classes plain names (Configuration, LocalConfiguration), hide them (_Configuration), or create a real submodule with m.def_submodule("eckit").

This one matters now because #35 is the first change to put these names into user-visible strings (_atlas4py.eckit.Configuration(...)); on master the two classes are bare registrations with no methods.

4. The dict payload isn't accepted by any constructor

Even with the prefix fixed, none of the spec-dict reprs round-trip:

• Config({'a': 1}) → TypeError; Config only has nb::init().
• Projection, Domain, Spacing, Metadata have no public constructor at all.
• Grid does have nb::init<const std::string&> (_atlas4py.cpp:330), but the repr prints the spec dict rather than the name, so Grid(...) still doesn't work.

PointLonLat / PointXY are the only two that would evaluate correctly today, once the module prefix is corrected.

Proposal

Pick one of two forms per type and apply it consistently:

• Constructible types get an eval-able repr using the public path: atlas4py.Grid('O32') (from g.name()), atlas4py.Gmsh('out.msh'), atlas4py.PointXY(x=..., y=...). For Config, a kwargs form — atlas4py.Config(a=1, nested={'x': 2}) — round-trips exactly if kwargs construction is reachable from __init__ rather than only from the Config.from_kwargs static added in Improve bindings for configuration types #35.
• Non-constructible types (Projection, Domain, Spacing, Metadata, and the eckit bases) use the standard non-eval form, <atlas4py.Metadata {...}>. The angle brackets are the conventional signal that the string is descriptive rather than an expression, which is more useful than a string that looks constructible and isn't.

A few doctest-style assertions (eval(repr(x)) == x for the first group, a prefix check for the second) would keep this from drifting again.

Note that apart from point 3 this is all pre-existing on master — #35 only renames toPyObject to atlas4py::make_object inside the repr bodies — so this is a separate cleanup rather than something to fold into that PR.

[wdeconinck]

I addressed what makes sense in the #35 and made another PR #42 to address remaining points.