fix(streaming): use parse_obj fallback to preserve nested event types (fixes #941)

[Echolonius]

Problem

When Pydantic's union discriminator silently returns the raw dict instead of a typed model, accumulate_event raises TypeError at the second isinstance check:

Unexpected event runtime type, after deserialising twice - {'type': 'content_block_delta', 'index': 0, 'delta': {'type': 'text_delta', 'text': '...\}} - <class 'dict'>

Why existing PRs (#1167, #1287, #1448, #1467) do not fully fix it

PRs #1287, #1448, and #1467 all add a type-map fallback, but use model_construct(**raw). That method skips Pydantic validation and leaves nested fields as raw dicts — so event.delta inside a content_block_delta event remains {"type": "text_delta", "text": "..."}. The moment accumulate_event accesses event.delta.type, a secondary AttributeError is raised and the text is never accumulated.

PR #1167 fixes the root cause in _models.py but changes core model-construction logic in a way that is hard to verify will not have side effects elsewhere.

This fix

Replace model_construct with parse_obj() — the SDK's existing Pydantic v1/v2 compatibility helper (in _compat.py). parse_obj runs full validation, so nested union fields like delta are properly resolved to their typed objects (TextDelta, InputJSONDelta, etc.), not left as raw dicts. If parse_obj itself raises, the original TypeError is preserved.

The same fallback is applied symmetrically to _beta_messages.py.

Changes

• src/anthropic/lib/streaming/_messages.py — add _RAW_EVENT_TYPE_MAP, import concrete event classes and parse_obj, apply fallback in accumulate_event
• src/anthropic/lib/streaming/_beta_messages.py — same for the beta path
• tests/lib/streaming/test_messages.py — three regression tests for the non-beta path (single delta accumulates, multiple deltas accumulate, unknown type still errors)
• tests/lib/streaming/test_partial_json.py — three matching regression tests for the beta path (TestBetaRawDictFallback)

Test plan

• test_accumulate_event_raw_dict_text_delta_is_fully_typed — passes a raw dict content_block_delta through accumulate_event and asserts snapshot.content[0].text == "hello", which only passes if event.delta.type was accessible (i.e. delta is a typed TextDelta, not a dict)
• test_accumulate_event_raw_dict_multiple_deltas_accumulate — three raw dict deltas build up "Hello there!"
• test_accumulate_event_raw_dict_unknown_type_still_raises — unknown type string still raises (TypeError, RuntimeError)
• Beta variants of all three in TestBetaRawDictFallback (in test_partial_json.py)

---

Written with Claude Code

[Echolonius]

Comparison with #1467

#1467 uses construct_type_unchecked(type_=RawContentBlockDeltaEvent, value=event) as the fallback. That converts the outer event into a BaseModel, so the first isinstance check passes — but construct_type_unchecked calls model_construct internally, which skips Pydantic validation and leaves nested fields as raw dicts.

Concretely, after #1467's fix, for a raw-dict content_block_delta event:

event.type        # "content_block_delta" ✓
event.delta       # {"type": "text_delta", "text": "hello"}  ← still a dict
event.delta.type  # AttributeError: 'dict' object has no attribute 'type'

accumulate_event then hits this line:

if event.delta.type == "text_delta":   # AttributeError here
    content.text += event.delta.text

So the text is never accumulated — the crash just moves from TypeError to AttributeError.

This PR uses parse_obj(target_cls, raw) instead. parse_obj runs full Pydantic validation, so nested union fields are resolved to their typed objects:

event.type        # "content_block_delta" ✓
event.delta       # TextDelta(type="text_delta", text="hello") ✓
event.delta.type  # "text_delta" ✓  — no AttributeError
content.text += event.delta.text  # works correctly

The regression test test_accumulate_event_raw_dict_text_delta_is_fully_typed specifically catches this: it asserts snapshot.content[0].text == "hello", which only passes if event.delta was a typed object (not a dict) when accumulate_event ran.