refactor(int32): per-model label dtype instead of sticky process-global

[FBumann]

This implements int dtype control per Model instead of through a global config option (global kept).
This resolves all potential issues with the global for little complexity.
I think this is the right approach.

Note

The following content was generated by AI.

Builds on the auto-widening behavior from #822 (now closed; this PR rebased onto master), but moves the label dtype from the sticky process-global option onto the Model itself.

What changed

• The label dtype becomes a Model() argument: Model(dtypes={"labels": np.int64}), default np.int32. It is exposed read-only as Model.dtypes — a types.MappingProxyType mapping (a frozen dict) with a single "labels" field for now, extensible to more fields later (per review feedback). Unknown keys and non-int32/int64 dtypes are rejected. The linopy.options["label_dtype"] setting is removed entirely — embedders forward constructor kwargs instead (PyPSA: n.optimize(model_kwargs={"dtypes": {"labels": np.int64}})), so config.py returns to its pre-perf: default integer arrays to int32 for ~25% memory reduction #566 state.
• A model widens only itself to int64 when its labels would exceed the int32 maximum; the UserWarning recommends passing dtypes={"labels": np.int64} upfront.
• All cast sites read the dtype from their object's model reference; common.save_join takes it as an explicit fill_dtype parameter instead of reading the global.
• read_netcdf restores the widened dtype per model (from the counters, as before). Pickle round-trips preserve it automatically since it is an instance attribute.
• OptionSettings.widen_label_dtype() and its sticky-reset() semantics are removed, along with the test-suite fixture that guarded against global-state leakage.

Why per-model

#822 chose the process-global because "the four float→int cast sites … have no reliable model reference (expressions can be detached)". That premise doesn't hold in the current codebase: BaseExpression.__init__ raises unless model is a Model instance, and every label_dtype cast site is a method on an object with a validated model reference (BaseExpression, Variable, Variables, Constraints, CSRConstraint). The only model-less site, save_join, is called with integer_dtype=True from exactly three container properties that all have self.model in scope.

With a reliable model reference, a per-model monotonic widen gives the same no-narrowing guarantee at the same zero hot-path cost (an attribute read), and removes the global design's edge cases:

• No revert footgun. Under the global design, following the documented revert (options["label_dtype"] = np.int32) while a widened model is alive lets the cast sites silently narrow >2^31 labels — int64 → int32.astype() wraps to negative values without any error. Per-model, there is nothing to revert.
• Widening travels with the model. A widened model shipped to another process via pickle/cloudpickle (dask, multiprocessing) keeps its dtype; the global flag does not cross process boundaries.
• No cross-model contamination.perf(int32): auto-widen labels to int64 on overflow instead of raising #822 documents that one huge model permanently costs every later small model in the process its int32 memory/speed win. Per-model, small models are unaffected.
• One knob in one place: the constructor argument. Nothing needs test isolation.

Cast-site → model-reference verification

Site	Enclosing scope	Model access
expressions.py vars cast in init	BaseExpression.__init__	model param, enforced isinstance(model, Model) (check moved above the cast)
expressions.pysanitize	BaseExpression	self.model
expressions.py simplify (2×)	LinearExpression	self.model
variables.pyffill/bfill/sanitize	Variable	self.model (init enforces isinstance(model, Model))
variables.pyflat remap	Variables	self.model
constraints.py_to_dataset (2×)	CSRConstraint	self._model
constraints.pyflat remap	Constraints	self.model
common.pysave_join	free function	dtype passed by callers (Variables.labels, Constraints.labels, Constraints.vars)

Tests

test/test_dtypes.py: the widen tests assert other models stay int32; the init-arg test drives the dtypes={"labels": ...} mapping (+ invalid-dtype and unknown-key rejection), plus test_dtypes_is_read_only_mapping (mutating m.dtypes raises TypeError), test_widen_applies_to_expressions, and test_auto_widen_survives_pickle; the netcdf test asserts the loaded model is widened. The restore_label_dtype fixture and the option tests are gone. Full suite passes.

🤖 Generated with Claude Code

[FabianHofmann]

generalize label_dtype accessor to dtypes which is property pointing to a frozen dict with fields "labels" at first stage. can be expanded later.

[FabianHofmann]

should only be true if dtype!=int64

[codspeed-hq]

Merging this PR will not alter performance

✅ 173 untouched benchmarks
⏩ 173 skipped benchmarks¹

---

_{Comparing perf/int32-model-label-dtype (e63d92b) with master (c02f5db)}

Footnotes

1. 173 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[FabianHofmann]

great!

[coroa]

This implemented auto-widening the dtypes now?

[FBumann]

This implemented auto-widening the dtypes now?

Yes!
And this led to storing it on the model to avoid issues