[simontaurus]

pydantic>=2.0.0 becomes pydantic>=2.0.0,!=2.13.*, plus a regression guard.

pydantic 2.13 routes ModelPrivateAttr.get_default() through resolve_default_value, which calls inspect.signature on the factory twice per private attribute per construction, uncached. LinkedBaseModel declares two factory-backed private attrs, _links and _extra_iris (src/oold/model/_descriptor.py:1319, :1322), so the cost lands on every model in the stack. Fixed upstream in 2.14.

Measured on one machine, same source, min of interleaved rounds:

pydantic	2 x PrivateAttr(default_factory=dict), bare BaseModel	LinkedBaseModel construction
2.12.0	5.62 us	-
2.13.4	90.19 us	43.01 us
2.13.5	48.75 us	-
2.14.3	3.28 us	4.90 us

Profile on 2.13.4: init_private_attributes is 80.6% of construction cumtime, with 80000 inspect.signature calls for 20000 constructions.

Changes:

• exclude the series rather than raise the floor, so 2.12 and earlier stay supported; CI has no minimum-version job
• uv.lock regenerated, resolves to 2.14.3
• tests/test_construction_cost.py asserts construction stays under 8x plain pydantic. Absolute timings on this machine drift by tens of percent between runs; the ratio holds within 3.38x to 3.53x. Measured 3.4x on 2.14 and 17.8x on 2.13, so the budget sits in the gap. Verified it fails on 2.13.4 (24.3x) and passes on 2.14.3
• the guard imports the descriptor binding directly, so it runs under either value of OOLD_DESCRIPTOR_BINDING; the legacy binding constructs an order of magnitude slower by design and would need a threshold that measures nothing

801 passed under the descriptor binding, 785 under OOLD_DESCRIPTOR_BINDING=0. The pydantic.v1 shim is still present in 2.14, so the v1 binding is unaffected.

Reworking _links and _extra_iris into lazy dicts was the alternative. It buys 3.6x to 4.9x on real classes against 9x here, and it breaks __eq__ at src/oold/model/_descriptor.py:1532, which compares __pydantic_private__ directly: a fresh instance holds None, one that has been read holds {}, and the __iris__ getter alone is enough to flip it.

[github-actions]

Release preview

Merging this PR would release v2.3.1 (current: v2.3.0).

Changelog preview (truncated)

## v2.3.1 (2026-11-15)

### Bug Fixes

- Exclude pydantic 2.13, which is ~9x slower to construct
  ([`48c405c`](https://github.com/OO-LD/oold-python/commit/48c405c00f90b2349de5198b5d355cb1e7cf3fde))

### Documentation

- **controller**: Name controller members for what they do
  ([`648018f`](https://github.com/OO-LD/oold-python/commit/648018fc480934b5b5185c6d0702860392e90c25))

### Testing

- Guard construction cost as a ratio of plain pydantic
  ([`084ca45`](https://github.com/OO-LD/oold-python/commit/084ca45820289c3228b53091106586189cc22fbc))

- Raise the construction budget to 15x
  ([`2a9f89b`](https://github.com/OO-LD/oold-python/commit/2a9f89b234842cead44dca827b3d8136d55bbc09))

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

[github-actions]

📊 Benchmark Results

Click to see benchmark comparison

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

✅ Performance Improvements:
  ✅ test_subclass_inheritance: 0.0638s → 0.0383s (-39.9%, ratio: 0.60x)

➖ Unchanged (within threshold):
  ➖ test_simple_dict_document_store: 0.0015s → 0.0013s (-9.7%)
  ➖ test_sqlite_document_store: 0.0016s → 0.0014s (-12.6%)
  ➖ test_local_sparql_store: 0.0286s → 0.0285s (-0.6%)
  ➖ test_oneof_subschema: 0.0441s → 0.0428s (-3.1%)
  ➖ test_enum_docstrings: 0.0394s → 0.0359s (-8.9%)
  ➖ test_class_hierarchy: 0.0401s → 0.0361s (-10.0%)
  ➖ test_core[v1]: 0.0255s → 0.0256s (+0.4%)
  ➖ test_core[v2]: 0.0322s → 0.0316s (-1.7%)
  ➖ test_schema_generation[v1]: 0.0013s → 0.0013s (+0.9%)
  ➖ test_schema_generation[v2]: 0.0028s → 0.0026s (-9.1%)
  ➖ test_simple_json: 0.0005s → 0.0005s (+0.4%)
  ➖ test_complex_graph: 0.0011s → 0.0011s (+1.5%)

============================================================
Summary: 0 regressions, 1 improvements, 12 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

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

📊 Benchmark Results

Click to see benchmark comparison

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

➖ Unchanged (within threshold):
  ➖ test_simple_dict_document_store: 0.0014s → 0.0011s (-18.0%)
  ➖ test_sqlite_document_store: 0.0015s → 0.0012s (-16.9%)
  ➖ test_local_sparql_store: 0.0281s → 0.0278s (-1.3%)
  ➖ test_oneof_subschema: 0.0433s → 0.0414s (-4.3%)
  ➖ test_enum_docstrings: 0.0393s → 0.0350s (-11.1%)
  ➖ test_subclass_inheritance: 0.0389s → 0.0407s (+4.7%)
  ➖ test_class_hierarchy: 0.0375s → 0.0380s (+1.3%)
  ➖ test_core[v1]: 0.0297s → 0.0300s (+0.9%)
  ➖ test_core[v2]: 0.0327s → 0.0302s (-7.7%)
  ➖ test_schema_generation[v1]: 0.0011s → 0.0011s (+0.0%)
  ➖ test_schema_generation[v2]: 0.0027s → 0.0024s (-10.7%)
  ➖ test_simple_json: 0.0004s → 0.0004s (+0.3%)
  ➖ test_complex_graph: 0.0011s → 0.0011s (+3.8%)

============================================================
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}