[BUG] Langfuse tests fail on Python 3.14 — langfuse SDK uses pydantic v1 internals incompatible with 3.14

[afarntrog]

Strands Version: N/A (dependency issue)

Strands Evals Version: current main (post-1dafdfb)

Python Version: 3.14.0

Operating System: macOS (Darwin 24.6.0)

Installation Method: pip

Steps to Reproduce:

1. Install strands-agents-evals with langfuse extra on Python 3.14
2. Run: hatch test tests/strands_evals/providers/test_langfuse_provider.py
3. All 29 langfuse provider tests fail at setup — the from langfuse import Langfuse
   import crashes because the installed langfuse (<3) uses pydantic.v1.BaseModel
   for its internal API models (e.g. AnnotationQueue), and pydantic v1's
   _set_default_and_type() cannot infer types under Python 3.14's new annotation
   evaluation (PEP 649).

Error: pydantic.v1.errors.ConfigError: unable to infer type for attribute "description"

Stack trace originates in:
langfuse/api/resources/annotation_queues/types/annotation_queue.py
→ class AnnotationQueue(pydantic_v1.BaseModel)
→ pydantic/v1/fields.py:576 ConfigError

Expected Behavior:
All langfuse provider tests pass on Python 3.14, as they do on Python ≤3.13.

Actual Behavior:
7 FAILED, 22 ERROR — every test in test_langfuse_provider.py fails. The provider fixture cannot patch
langfuse_provider.Langfuse because importing the module crashes. Tests on other Python versions are
unaffected (631 passed).

Additional Context:

• The langfuse warning confirms it: Core Pydantic V1 functionality isn't compatible with Python 3.14 or
  greater.
• Current dependency constraint: langfuse>=2.0.0,<3
• langfuse 4.0.1 (latest) has moved off pydantic.v1.BaseModel to pydantic.BaseModel (v2) via
  UniversalBaseModel. It requires pydantic>=2,<3 and no longer ships pydantic v1 model definitions.
• However, langfuse 4.x introduces a breaking change: observations.get_many switches from page-based
  pagination (page/meta.total_pages) to cursor-based pagination (cursor/meta.cursor). The trace.list API is
  unchanged.

Possible Solution:

1. Bump dependency to langfuse>=4.0.0,<5
2. Update LangfuseProvider._fetch_observations to use cursor-based pagination (the new observations.get_many
   uses cursor param and returns ObservationsV2Meta with .cursor instead of MetaResponse with .total_pages)
3. Update _fetch_all_pages or add a _fetch_all_cursor helper for the observations endpoint
4. Update test mocks to reflect the new pagination shape
5. _fetch_traces_for_session and its tests need no changes — trace.list still uses page-based pagination with
   MetaResponse