[ericapisani]

Summary

Converts the 8 live sentry_sdk.start_transaction(...) calls in test-openai to sentry_sdk.traces.start_span(...).

test-openai already uses trace_lifecycle="stream". Under streaming, sentry_sdk.start_transaction(...) (and the top-level sentry_sdk.start_span(...)) both return a NoOpSpan — only sentry_sdk.traces.start_span(...) creates a real span. This PR updates every project entrypoint to use the real streamed span API.

Files changed:

• main.py
• main_async.py
• main_custom.py
• main_streamed.py
• main_truncation.py
• main_conversation.py (2 calls)
• main_conversation_async.py

Each conversion:

• Adds import sentry_sdk.traces (not re-exported from sentry_sdk/__init__.py)
• Replaces sentry_sdk.start_transaction(name=X) with sentry_sdk.traces.start_span(name=X)

All 8 calls were simple with sentry_sdk.start_transaction(name=...): blocks — no op=, no bound span variable, and no set_data/set_tag/set_status usage, so no further attribute conversion was needed.

Left untouched, on purpose: the 3 @sentry_sdk.trace(...) decorators in main_custom.py (lines 11, 19, 24) use template=SPANTEMPLATE.AI_TOOL / AI_CHAT / AI_AGENT. The streaming decorator (sentry_sdk.traces.trace) only accepts name/attributes/active — there is no streaming equivalent for template/op yet, so converting these would silently drop the AI span templates. This is left as a follow-up pending SDK support.

The before_send_span callbacks in main_streamed.py, main_conversation.py, and main_conversation_async.py are unchanged.

Test plan

• python -m py_compile on all 7 edited files
• grep -rn "start_transaction" test-openai returns no matches
• git diff confirms no @sentry_sdk.trace decorator line changed

[linear-code]

PY-2758