[ethan-scitix]

Type

• fix — bug fix or alignment correction

Summary

• Every stage and model concurrency limit could run one task over its bound. They are all anyio.CapacityLimiter. Before anyio 4.15, a release woke the next waiter without handing it the token. A task scheduled ahead of that waiter found the wait queue empty and a free slot, so it took the token; then the waiter took one too.
• This is the intermittent assert 5 <= 4 in test_runner.py::TestE2EStageConcurrency::test_stage_concurrency_limit_allows_higher. sieval and the test are both correct.
• Production impact: concurrency_limit, concurrency_limits and a model's concurrency_limit can occasionally admit one extra in-flight request. That is harmless for most endpoints, but can draw a 429 from one rate-limited to an exact concurrency.
• Fix: anyio>=4.15.0. 4.15 adds the waiter to the borrowers before waking it (_notify_next_waiter). Bisected with a deterministic repro: 4.12.0, 4.12.1, 4.13.0 and 4.14.0 all run 2 tasks under a limit of 1; 4.15.0 and 4.15.1 run 1.
• anyio 4.15 requires typing-extensions>=4.16.0, so the lock moves it 4.15.0 → 4.16.0. It is declared in pyproject.toml so the bump is reviewable rather than resolved into existence (check_deps).
• New regression test TestCapacityLimiterHandoff in tests/unit/core/utils/test_concurrency.py pins the handoff deterministically: no timing, one scheduling step.

Related Issues

Found while trimming the unit suite in #222.

Test Plan

Automated

• Lint/format clean (ruff check && ruff format --check, pre-commit incl. taplo + lock-drift)
• Type check clean (ty check on the changed test)
• Unit tests pass: pytest tests/unit tests/integration tests/acceptance -m "not stress and not benchmark" -n 16 with anyio 4.15.1 + typing-extensions 4.16.0 → 9644 passed, 2 skipped, no new warnings
• scripts/check_preflight.py --check check_deps → PASS (2 justified moves)

Manual

• Regression test fails on anyio 4.12.0 (assert 2 == 1) and passes on 4.15.1.

• The runner scenario (concurrency_limits={"infer": 4}, 6 samples), 16 parallel processes × 150 runs:

  anyio	runs	peak 4	peak 5
  4.12.0	2400	2330	70 (2.9%)
  4.15.1	2400	2400	0

Checklist

Required (all PRs)

• PR title follows conventional format (type(scope): description)
• No internal paths, credentials, or personal info in committed files
• AI-generated code has AI-Generated Code - <model> (<provider>) in module docstring
• No new upper-layer dependencies added to core/
• Deleted code verified — no remaining call sites depend on it

If: New Dependency

• Added to correct PDM dependency group: typing-extensions is default, as anyio's own requirement, not a new import
• Justified in Summary

🤖 Generated with Claude Code