fix: reuse duplicate capability extension models

[ShuoRen-TT]

Summary

• enable datamodel-code-generator model reuse during SDK generation
• regenerate capability aliases so repeated constraints reference canonical definitions
• add a dependency-free regression test for the generated compatibility aliases

Motivation

Schema preprocessing expands the inline capability extends union into several structurally identical anonymous schemas. Without model reuse, code generation repeats the same string constraints across the numbered Extends* aliases.

This enables the --reuse-model option recommended in #36. Existing numbered aliases remain importable and preserve their validation behavior; this deliberately does not enable --collapse-reuse-models, which would remove those compatibility names.

Validation

• python -m unittest discover -s tests -p "test_*.py" on Python 3.10, 3.11, and 3.12: 14 tests passed on each version
• regression test fails against the previous generated output and passes after regeneration
• two consecutive ./generate_models.sh 2026-04-08 runs produced the same capability.py SHA-256
• targeted Ruff format and lint checks passed
• bash -n generate_models.sh passed
• runtime validation confirmed the reused string and list aliases retain their Pydantic semantics

Related work

#55 also touches generate_models.sh but addresses minProperties enforcement. There is no functional dependency or overlap beyond that file.

Fixes #36