GeezerVison

The HuggingFace Hub and Cohere streaming iterators wrapped the entire for chunk in res: ... yield chunk loop in capture_internal_exceptions(), which catches every exception — including provider errors raised mid-iteration. The loop ended normally, the truncated answer looked complete, and nothing was reported.

Fix (following the repo's own _capture_exception + span-finish + reraise pattern from the non-streaming paths): iteration/yield now runs outside capture_internal_exceptions — only Sentry's span-data bookkeeping stays inside it. A try/except around the loop records the provider error, finishes the span with error status, and re-raises so the app sees the same exception as without Sentry.

Verified: 4 new regression tests (test_streaming_chat_iteration_error_propagates x2, test_text_generation_streaming_iteration_error_propagates, test_chat_completion_streaming_iteration_error_propagates) fail on pristine code with DID NOT RAISE (exception swallowed) and pass with the fix. Cohere suite 76/76 green, ruff clean, mypy shows no new errors.

Fixes #7889.