fix(experiments): reject nonpositive concurrency limits

[linjiaye929]

What does this PR do?

Reject nonpositive max_concurrency values before scheduling an experiment.

With nonempty data, max_concurrency=0 creates a zero-permit semaphore and every item waits indefinitely. No task can enter the semaphore and release a permit. Negative values already raise ValueError from asyncio.Semaphore; they retain that exception type with a more actionable message.

The validation runs before experiment IDs, item tasks, or observations are created. The client and dataset docstrings explicitly require a value greater than zero, including for empty datasets. Positive limits and empty datasets with positive limits retain their behavior.

Type of change

• Bug fix

Verification

Python 3.12.14; dependencies installed from the frozen lockfile. socksio==1.0.0 was added only to the local test environment so httpx could initialize with the execution environment's SOCKS proxy; dependency files are unchanged. Proxy settings were preserved for the final verification. Test commands use OTEL_TRACES_SAMPLER=always_on because the execution environment otherwise samples only 1% of spans; this only affects the test process and its in-memory exporter.

• OTEL_TRACES_SAMPLER=always_on .venv/bin/pytest tests/unit/test_experiment.py::TestExperimentConcurrency -q: 10 passed. Uses local asyncio scheduling and the repository's in-memory Langfuse fixture, with dummy credentials and no real provider or Langfuse requests.
• Before the fix, the same 10 cases produced 6 failures and 4 passes. The public zero-concurrency call hit the 5-second pytest timeout; the internal nonempty zero-concurrency call hit a 1-second asyncio.wait_for timeout and was cancelled. Empty data with zero concurrency did not raise, and negative inputs raised the old semaphore ValueError message.
• OTEL_TRACES_SAMPLER=always_on .venv/bin/pytest tests/unit/test_experiment.py -q: 30 passed. The pristine base at bf11ec121479145923ff27b32ec2ad1d06fd2d51 passes its 20 existing experiment tests with the same test-only sampler setting. Initial runs had four missing-span failures on both base and fixed source because of the inherited 1% sampler; those failures disappear with deterministic test sampling. This is not a full-suite or CI pass.
• .venv/bin/ruff check .: passed.
• .venv/bin/ruff format --check langfuse/_client/client.py langfuse/_client/datasets.py tests/unit/test_experiment.py: passed.
• .venv/bin/mypy langfuse --no-error-summary: passed.
• git diff --check: passed.

The full unit suite, E2E suite, and live-provider tests were not run.

Checklist

• I self-reviewed the diff using code_review.md.
• I added tests for the behavior change.
• I updated the affected parameter documentation.
• I did not hand-edit generated files.
• I did not commit secrets or credentials.

AI assistance: OpenAI Codex inspected the source, implemented the change, and ran the local checks described above.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}