feat: public controller registry lookup, with one precedence rule

[simontaurus]

Makes the controller registry usable from outside oold, and gives resolution one
precedence rule instead of three.

Public lookup

A caller holding an entity could not find the controller that implements it
without knowing the concrete class, so downstream packages reached into
_controller_types directly. Added, for both pydantic layers:

• register_type() / register_controller_type()
• get_registered_type() / get_registered_controller_types()
• registered_types() / registered_controller_types()
• find_controller_type(type_iris, base=None)
• oold.backend.auth.registered_credentials()

One precedence rule

Previously the three accessors disagreed: get_registered_controller_types
returned earliest-first, find_controller_type last-first, and resolve_type
refused to resolve at all once a second controller was registered - which turned
the documented override mechanism into a way of disabling controllers.

• The controller registered last wins, in every accessor.
• A registered controller replaces its pure data model during resolution, in
  from_json / from_jsonld and for inline linked documents, falling back to
  the pure model when the controller cannot be built from the document.
• The two layers never mix: a lookup is answered from the registry of the layer
  the model registry belongs to. resolve_type took the v2 controller registry
  even when resolving a v1 type.

Registration

• __oold_register__ = False keeps a class out of both registries. For a
  controller that reuses another model without being its handler - an in-memory
  buffer over a local-file model - which would otherwise compete for that
  model's type IRI.
• The v1 register_type gained the inherited-IRI guard the v2 path had: a
  subclass that only narrows a field reported its parent's IRI and replaced it,
  so every parent document deserialized as the subclass.
• register_type / register_controller_type flatten nested IRI lists and
  raise on an unusable IRI instead of silently doing nothing.
• Every registration logs at INFO, an override at WARNING. A module reloaded by
  importlib.reload re-registers rather than piling up a second copy.

Notes

• Rebased onto v2.2.0; 800 tests pass, also with OOLD_DESCRIPTOR_BINDING=0.
• tests/conftest.py now restores the controller registries between tests,
  deep-copying the per-IRI lists.
• Two packages depend on this and pin the branch until it is released:
  opensemantic.core-python
  (file controllers) and
  osw-python
  (registered_credentials). Both unpin once this lands.

[github-actions]

Release preview

Merging this PR would release v2.3.0 (current: v2.2.0).

Changelog preview (truncated)

## v2.3.0 (2026-10-15)

### Bug Fixes

- One precedence rule for controllers, isolated per pydantic layer
  ([`2b3f70d`](https://github.com/OO-LD/oold-python/commit/2b3f70dc27d42a508a90ebaa4eeca1dae0fd16a2))

- Quieter registry logging, and controllers for inline documents
  ([`37c578e`](https://github.com/OO-LD/oold-python/commit/37c578e7b81f91075e5c6ef2ce9ed993ad0e4dda))

### Features

- Let a class opt out of the type registries
  ([`e501f01`](https://github.com/OO-LD/oold-python/commit/e501f0103e488d2354cd3b34a0561e1fd146c125))

- Public lookup for the controller type registry
  ([`e777b70`](https://github.com/OO-LD/oold-python/commit/e777b7091484f6aad4edbe3cc3e0419deccc23df))

- Public registry lookup for the pydantic v1 models
  ([`84b76cb`](https://github.com/OO-LD/oold-python/commit/84b76cb003e9560dc1c6c2d2262c0e363dfcc85b))

- **auth**: Expose the registered credentials
  ([`12e979b`](https://github.com/OO-LD/oold-python/commit/12e979b65e2acfe1a5342c0118df6ef6206108b2))

_{Preview via python-semantic-release and conventional commits.}

[github-actions]

📊 Benchmark Results

Click to see benchmark comparison

📊 Benchmark Comparison (threshold: 1.3x)
============================================================

➖ Unchanged (within threshold):
  ➖ test_simple_dict_document_store: 0.0020s → 0.0020s (-0.5%)
  ➖ test_sqlite_document_store: 0.0021s → 0.0021s (+0.1%)
  ➖ test_local_sparql_store: 0.0418s → 0.0396s (-5.2%)
  ➖ test_oneof_subschema: 0.0636s → 0.0630s (-0.9%)
  ➖ test_enum_docstrings: 0.0542s → 0.0527s (-2.8%)
  ➖ test_subclass_inheritance: 0.0591s → 0.0584s (-1.2%)
  ➖ test_class_hierarchy: 0.0555s → 0.0561s (+1.1%)
  ➖ test_core[v1]: 0.0384s → 0.0384s (+0.0%)
  ➖ test_core[v2]: 0.0456s → 0.0470s (+3.0%)
  ➖ test_schema_generation[v1]: 0.0017s → 0.0017s (-0.4%)
  ➖ test_schema_generation[v2]: 0.0038s → 0.0038s (-0.7%)
  ➖ test_simple_json: 0.0007s → 0.0007s (+4.6%)
  ➖ test_complex_graph: 0.0017s → 0.0016s (-2.7%)

============================================================
Summary: 0 regressions, 0 improvements, 13 unchanged
============================================================

✅ No significant performance regressions

Threshold: 1.3x (30% slower triggers a regression warning)

Note: Benchmarks are informational only and won't fail the build.

_{💡 Tip: Download the benchmark-results artifact for detailed JSON data}

[codecov]

Codecov Report

❌ Patch coverage is 87.64706% with 21 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/oold/model/_descriptor.py	65.5%	6 Missing and 4 partials ⚠️
src/oold/model/v1/__init__.py	82.5%	3 Missing and 4 partials ⚠️
src/oold/model/__init__.py	94.2%	2 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!