fix(gemini): tolerate missing usage metadata on final response - #4724

Fixes #4721.

Problem

Gemini sometimes sends a complete final response without usage metadata. Strands finishes the answer, then tries to read token counts from an object that is not there and crashes.

Fix

One condition: if event becomes if event and event.usage_metadata is not None. If the counts object is absent, the stream finishes normally without inventing a zero count. If the object exists with empty counts, existing behavior is unchanged. No change to token calculations, finish reasons, formatting, requests, or provider calls.

Tests

Three cases use real google.genai GenerateContentResponse objects parsed from JSON: two omitted-usage cases (STOP and SAFETY finish reasons) checking preserved answer text, expected stop reason, normal completed iteration, and no invented metadata; one empty-but-present usage object case protecting existing zero-count behavior. The client's response source is mocked only through the existing test fixture; the actual editable-checkout GeminiModel.stream and formatter run normally.

Before: 2 failed, 1 passed (the focused 3 cases). After: the full existing + new test_gemini.py passes, 78 tests in 1.22s. Python 3.10.12, pytest 9.1.1. Ruff check/format and git diff --check clean. No API requests, credentials, models, GPU, or broader integration-suite runs are claimed.

Credit: @abdes reported the issue.

Prepared with AI assistance (Instinct assisted).