fix(serializer): preserve TypeAttribute reference identity

[zhenrongliew]

Summary

Serialization should preserve identity relationships:

Original IR	(De)serialized-roundtrip IR
Same instance referenced twice	One new instance referenced twice
Two distinct but equal instances	Two distinct but equal new instances

Motivation

#711 encodes repeated TypeAttribute instances as one definition followed by
attr_refs. However, the decoder recreated the referenced definition for every
attr_ref, preserving v1's value behavior but losing the identity relationship of the original IR.

As @weinbe58 demonstrated, this can also make decoding exponential:

typ = types.Generic(tuple, typ, types.Vararg(typ))

Each level references the previous type twice. The serialized DAG is linear, but
recreating every reference expands it into a binary tree during decoding.

Fix

Add a TypeAttribute_Lookup instance table to SerializationContext. Each
TypeAttribute definition is decoded once and cached by its serialization ID;
subsequent attr_refs return that same instance.

[github-actions]

☂️ Code Coverage

current status: ✅

Overall Coverage

Statements	Covered	Coverage	Threshold	Status
11889	10725	90%	0%	🟢

New Files

No new covered files...

Modified Files

File	Coverage	Status
src/kirin/ir/attrs/types.py	88%	🟢
src/kirin/serialization/base/context.py	96%	🟢
src/kirin/serialization/base/deserializer.py	84%	🟢
TOTAL	89%	🟢

updated for commit: c85e772 by action🐍

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-09-01 13:20 UTC

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[weinbe58]

Verified the fix separately — the exponential decode is gone (flat ~0.5 ms across depths 8–22, vs ~43 s at depth 22 before), encode is unaffected, and the full suite is green.

One coverage regression worth fixing before merge: rewriting test_nested_type_graph_preserves_v1_value_semantics into test_nested_type_graph_preserves_identity_relationships dropped the only Union/Literal coverage in the serialization tests. The old version built types.Literal(1, leaf) and types.Union(literal, types.String); the new one only exercises Generic. After this PR, types.Union and types.Literal appear in no serialization test anywhere under test/ — on main they're at test_type_attribute_refs.py:190-191.

Worth keeping because those two have the least ordinary construction paths, and they interact with the new TypeAttribute_Lookup caching in ways Generic doesn't exercise: UnionMeta.__call__ does subset-simplification and can return a member instead of a Union, and LiteralMeta interns on (data, datatype). A small separate test that round-trips a nested Union/Literal graph and asserts the shared leaf comes back shared would cover it — no need to restore the old test's shape.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

[zhenrongliew]

Verified the fix separately — the exponential decode is gone (flat ~0.5 ms across depths 8–22, vs ~43 s at depth 22 before), encode is unaffected, and the full suite is green.

One coverage regression worth fixing before merge: rewriting test_nested_type_graph_preserves_v1_value_semantics into test_nested_type_graph_preserves_identity_relationships dropped the only Union/Literal coverage in the serialization tests. The old version built types.Literal(1, leaf) and types.Union(literal, types.String); the new one only exercises Generic. After this PR, types.Union and types.Literal appear in no serialization test anywhere under test/ — on main they're at test_type_attribute_refs.py:190-191.

Worth keeping because those two have the least ordinary construction paths, and they interact with the new TypeAttribute_Lookup caching in ways Generic doesn't exercise: UnionMeta.__call__ does subset-simplification and can return a member instead of a Union, and LiteralMeta interns on (data, datatype). A small separate test that round-trips a nested Union/Literal graph and asserts the shared leaf comes back shared would cover it — no need to restore the old test's shape.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

Restored separate Union/Literal structural round-trip coverage and added a test where a TypeVar is shared between a Literal and a Generic.
The identity test exposed an issue: Literal.deserialize called the normal Literal(...) constructor, whose LiteralMeta cache could return the source Literal. That retained the source leaf instead of the decoded leaf, so the objects were structurally equal but the shared instance was lost.
I updated Literal.deserialize to reconstruct a fresh instance without consulting the process-local interning cache.

[weinbe58]

LGTM

[github-actions]

Backport results for 024f6ce

Succeeded:

• release-0-22: Backport PR #713 to release-0-22 #716