fix(instrumentation): guarantee span ends when Anthropic stream manager exit raises

[Aftabbs]

Summary

When MessageStreamManager.__exit__ or AsyncMessageStreamManager.__aexit__ raises an exception, post_hook_exit_impl / post_hook_aexit_impl was never called, so span.end() was skipped and the OTel span leaked. Every subsequent request then ran under a stale unclosed parent span.

Root Cause / Motivation

In both WrappedMessageStreamManager.__exit__ and WrappedAsyncMessageStreamManager.__aexit__, the telemetry cleanup was called sequentially after the underlying manager's exit:

# before (sync)
def __exit__(self, exc_type, exc_val, exc_tb):
    self._manager.__exit__(exc_type, exc_val, exc_tb)  # if this raises...
    post_hook_exit_impl()                               # ...this never runs → span leaks

When the underlying stream manager raises (e.g., network timeout, server error during stream teardown), Python propagates that exception before post_hook_exit_impl() is reached, leaving the OTel span permanently open.

Changes

File	What changed
src/judgeval/instrumentation/llm/llm_anthropic/messages_stream.py	Wrapped _manager.__exit__ and _manager.__aexit__ calls in try/finally so span cleanup always runs; added clarifying comment to the intentional bare except in token-extraction
src/tests/instrumentation/llm/anthropic/test_messages_stream.py	New test file: 5 tests covering span creation, token recording, and the span-leak fix for both sync and async paths

Testing

• Existing tests pass (7 / 7 in test_messages.py)
• New test: test_span_ends_even_when_manager_exit_raises — verifies span is exported even when __exit__ raises (was asserting False before fix)
• New test: test_span_ends_even_when_manager_aexit_raises — same for async path
• New tests: test_span_is_created and test_span_records_token_usage for both wrappers
• All 12 tests pass

Impact

Users who experience stream teardown errors (network issues, server-side errors during streaming) previously accumulated leaked OTel spans that were never closed. This caused traces to appear as perpetually-in-progress in Judgment and could inflate trace timing metrics.

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.