[ty] Index TypedDict identities when building unions

[charliermarsh]

Summary

Unions containing only TypedDicts retain distinct dictionary types by identity, but adding each member still scans the existing members for duplicates. We now maintain an identity set alongside the ordered member vector, avoiding repeated scans as these unions grow.

The set is allocated only after a second distinct dictionary is added and is discarded when a non-TypedDict member requires ordinary union simplification. Flattened unions and expanded aliases preserve first-occurrence order and recursion metadata.

Against e579b84aba, the existing Pydantic CodSpeed walltime benchmark measured 2.708 s before and 2.591 s after this change alone. This is a 4.3% reduction in elapsed time in this sample. These local measurements use one process pair with six samples per process and two iterations per sample.

[astral-sh-bot]

Typing conformance results

No changes detected ✅

Current numbers

The percentage of diagnostics emitted that were expected errors held steady at 97.84%. The percentage of expected errors that received a diagnostic held steady at 96.45%. The number of fully passing files held steady at 128/145.

[astral-sh-bot]

Memory usage report

Memory usage unchanged ✅

[astral-sh-bot]

ecosystem-analyzer results

No diagnostic changes detected ✅

Full report with detailed diff (timing results)

[codspeed-hq]

Merging this PR will degrade performance by 4.85%

❌ 2 regressed benchmarks
✅ 138 untouched benchmarks
⏩ 84 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Simulation	ty_micro[inherited_recursive_protocol_diagnostic]	7.8 s	8.2 s	-4.9%
❌	Simulation	ty_micro[inherited_recursive_protocol_constructor]	7.8 s	8.1 s	-4.8%

Tip

Investigate this regression with the CodSpeed MCP and your agent.

---

_{Comparing charlie/codex-ty-typed-dict-union-index (f670cde) with main (e579b84)}

Footnotes

1. 84 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. ↩