Summary

Jira: SAO-18644

When several processes start together, each can find no agent stream or project and try to create it. The server keeps one and rejects the rest. Until now the rejected processes raised from SplunkAOLogger and failed at startup.

• After a failed create, _init_agent_stream() and _init_project() look the stream or project up again and use what the winner made.
• For a project, the look-up repeats up to 5 times, 0.2 s apart, because the winner may not have granted access yet.
• If nothing appears, the original create error is raised, so other failures still show.
• An existing stream or project, and a normal create, behave as before.

The server rejects the loser with HTTP 400 for a stream and HTTP 422 for a project. Older servers returned HTTP 500 after duplicates existed. The change does not depend on the status code.

This pairs with rungalileo/orbit#3159. That PR makes the server reject a duplicate stream name instead of silently creating a second stream. Without this SDK change, SDK processes that start together on a brand-new stream can fail at startup after that PR. This PR does not depend on it. It only changes what happens after a create has already failed, so it is safe on its own and with older servers. Shipping it first or together with the orbit change is recommended, not required.

Test plan

• 11 new unit tests (tests/test_logger_lost_create_race.py): an existing or new stream or project, a lost race, a wait for access, a winner that never appears, a failing second look-up, and a winner of the wrong project type.
• With the look-up disabled, 7 of the 11 fail. The 4 normal-path tests still pass.
• 31 nearby tests in tests/test_logger_otel_egress.py pass.
• ruff check and ruff format --check pass on both changed files.
• Real race against a local orbit API (with the orbit change above): 6 rounds of 8 SDK processes started at the same moment, standalone mode, with about 6 ms of database latency.

Neither run left a duplicate stream or project.

Notes

• mypy passes, but the repo's config excludes splunk_ao.logger, so it did not check this change.
• Other get-then-create spots in the SDK (for example in decorators and experiments) are not changed here.
• CHANGELOG.md has an entry under Unreleased. The branch was merged with main (v0.6.1) before the entry was added. The 42 tests (11 new and 31 nearby) pass on the merged branch.
• Not run: a lab realm, the full test suite.