Retain label configuration and generation in policy versions

[terjekv]

Retain nullable label identifiers and engine generations in PolicyVersion, authorization responses, and version responses. Older servers default missing fields to None and 0; frozen model equality includes all state dimensions. Generation is validated as an unsigned 64-bit integer, rejecting booleans and out-of-range values, and is documented as local to an engine instance.

The initial migration repeated version validation and construction for each response item, producing six CodSpeed regressions. Reuse immutable policy versions through a bounded 256-entry cache keyed by the model class, every state field, and runtime types; booleans cannot alias cached integers. Avoid repeated helper calls for absent legacy fields. Custom subclasses remain uncached so their constructors and independent state preserve existing behavior. Pass subclass fields by keyword, and retain the two-keyword constructor contract when parsing legacy responses. Cache-isolation tests cover every state dimension, and added benchmarks cover complete modern metadata and continuously changing generations.

Same-runtime local comparisons versus main improve version-response parsing from 4.93 to 3.88 microseconds, one-item brief parsing from 8.24 to 6.97 microseconds, 100-item brief parsing from 514 to 451 microseconds, and 50-item detailed parsing from 336 to 298 microseconds. The final CodSpeed comparison reports 10.41% improvement, one improved benchmark, 30 untouched benchmarks, and no regressions. Refreshing the main baseline also removed the runtime-environment warning.

Exercise the integration suite against REST v0.0.16 and retain v0.0.12 coverage. The current-server test asserts that loaded label configuration is returned with its SHA-256 identifier.

Verification

73 unit tests, all 29 integration cases against both published REST v0.0.12 and v0.0.16, Pyright, Basedpyright with zero warnings, and Markdown lint pass. All 33 benchmark cases pass locally. The subclass constructor regression tests pass; the guard also passes the full unit suite, both type checkers, and same-runtime performance comparisons; the final CodSpeed comparison is green and has no runtime-environment warning. Container tests use isolated Podman storage.

[codspeed]

Merging this PR will improve performance by 10.41%

⚡ 1 improved benchmark
✅ 30 untouched benchmarks
🆕 2 new benchmarks

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_parse_brief_authorization_response_128	7.3 ms	6.7 ms	+10.41%
🆕	test_policy_version_with_changing_generation	N/A	84.5 µs	N/A
🆕	test_policy_version_with_label_metadata	N/A	49.9 µs	N/A

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing agent/core-0.0.24-migration (9b76ee3) with main (3c53d38)}