ENH support subclasses of list, set, dict, object for circular pointers

[adrinjalali]

Follow up from #549 and fixes #553

This was flagged while working on #549 but kept it separate for an easier review.

[Copilot]

Copilot review overview

🟡 Changes recommended

Empty custom state, slotted subclasses, and circular defaultdict instances remain incorrectly handled.

Review effort: Balanced
Findings: 3 · 1

Open (4)

• Empty getstate dict incorrectly suppresses setstate · New
• Slotted subclass state tuple is not restored correctly · New
• defaultdict cycles remain unsupported without DefaultDictNode · New
• Set subclass construction differs from pickle despite documented equivalence · New

What changed in this PR

Adds circular-reference and attribute persistence for built-in container subclasses while retaining protocol 0–2 compatibility.

Changes:

• Constructs container subclasses with __new__, fills them in place, and restores state.
• Adds legacy container-node loaders and compatibility tests.
• Expands round-trip and circular-reference coverage.

File	Description
skops/​io/​_general.py	Implements container state persistence and reconstruction.
skops/​io/​_utils.py	Expands circular-reference support detection.
skops/​io/​old/​_general_v2.py	Adds legacy container loaders.
skops/​io/​tests/​test_persist.py	Tests subclass state and circular references.
skops/​io/​tests/​test_persist_old.py	Tests legacy protocol behavior.
docs/​changes.rst	Documents the persistence changes.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🔵 Needs a closer look

Defaultdict subclasses remain downcast, and overridden set mutators can break loading.

Review effort: Balanced
Findings: 1

Open (1)

• defaultdict cycles remain unsupported without DefaultDictNode

Resolved since last review (3)

• Slotted subclass state tuple is not restored correctly
• Empty getstate dict incorrectly suppresses setstate
• Set subclass construction differs from pickle despite documented equivalence

Previously missed (2)

In code that hasn't changed since last review

Preserve defaultdict subclasses and instance state during loading

skops/​io/​_general.py:208

singledispatch routes subclasses of collections.defaultdict to defaultdict_get_state, but this loader always creates the built-in defaultdict. Such subclasses are therefore silently downcast and their instance attributes are discarded, despite this PR's support for dict subclasses. Preserve the serialized class and state here (with an old-protocol reader for the prior schema), or explicitly exclude these subclasses from the stated support.

Use the built-in set mutator when loading set items

skops/​io/​_general.py:304

Calling the virtual instance.update can execute an override on an object whose __init__ was deliberately skipped. A valid set subclass whose overridden update depends on initialized state—or simply rejects direct calls—will fail to load, whereas pickle bypasses that override for set items. Invoke the built-in mutator directly so filling cannot depend on subclass initialization.

[Copilot]

Copilot review overview

🟢 Approval recommended

The implementation is coherent, backward-compatible, and thoroughly covered by targeted tests.

Review effort: Balanced
Findings: 1

Open (1)

• defaultdict cycles remain unsupported without DefaultDictNode