TokenUsageExtractor drops mapping-based usage metadata

[KXHXK]

Description

TokenUsageExtractor.extract_from_response() silently drops all token counts when a provider exposes usage or usage_metadata as a mapping instead of an attribute-based object.

The shared streaming wrapper explicitly accepts chunks with usage_metadata, and the existing test fixture uses a dictionary shape, but _extract_from_usage_object() reads every field with getattr(). As a result, dictionary-backed usage produces a TokenUsage whose fields are all None, so the span receives no token usage attributes.

Minimal reproduction

from types import SimpleNamespace

from agentops.instrumentation.common.token_counting import TokenUsageExtractor

response = SimpleNamespace(
    usage_metadata={
        prompt_tokens: 10,
        completion_tokens: 5,
        total_tokens: 15,
    }
)

print(TokenUsageExtractor.extract_from_response(response))

Current output on main:

TokenUsage(prompt_tokens=None, completion_tokens=None, total_tokens=None, cached_prompt_tokens=None, cached_read_tokens=None, reasoning_tokens=None)

Expected behavior

Dictionary and attribute-based usage containers should be normalized consistently, including the cache and reasoning token fields already supported by TokenUsage.

Suggested fix

Use a small mapping-aware field accessor in _extract_from_usage_object() and add regression coverage for both usage and usage_metadata mappings. The existing object behavior should remain unchanged.