Release v1.1.0

This release resolves a full code audit of v1.0.2 — concurrency, lifecycle,
packaging, and API-boundary issues — and removes a class of intermittent CI
failures caused by un-awaited coroutines in the test suite.

Added

• Public API: ScheduleResult and QueuedRequest are now exported from
  the top-level adaptive_rate_limiter package and from
  adaptive_rate_limiter.scheduler. They are the return type of
  submit_request in INTELLIGENT and ACCOUNT modes.
• Config: RateLimiterConfig gained max_retries, batch_size,
  stale_entry_ttl, and max_tracking_entries fields, each validated in
  __post_init__.
• CI: a packaging job builds the wheel and verifies all six bundled Lua
  scripts ship with it, guarding against a regression that would break the
  [redis] install.

Changed

• ACCOUNT mode: the concurrency limit is now read from the correct
  config field, max_concurrent_executions. Previously the strategy looked
  up a non-existent max_concurrent_requests attribute and silently fell
  back to the hardcoded default of 10.
• Reservations & streaming: staleness and abandonment detection now use
  a monotonic clock (time.monotonic()), making it immune to wall-clock and
  NTP adjustments. The effective stale-reservation age is derived from
  request_timeout, so cleanup can no longer reclaim a still-running
  reservation.
• Removed the empty, unused _internal/ package. _-prefixed names were
  never part of the public API, so this is not a breaking change.

Fixed

• Scheduler loop: the INTELLIGENT mode loop no longer terminates
  permanently when an unexpected exception is raised inside it.
• Queue race: request enqueue and the _queue_has_items flag are now
  mutated under the per-queue lock, preventing a lost update that could
  leave a non-empty queue unscheduled.
• Reset watcher: no longer raises RuntimeError: Set changed size during iteration when a watcher fires during reset-time calculation.
• Shutdown: stop() now cancels and awaits in-flight execution tasks
  (INTELLIGENT and ACCOUNT modes) before the backend shuts down, so
  capacity-release cleanup completes and request coroutines are not left
  un-awaited.
• Cancellation orphan: a reservation orphaned by cancellation between
  capacity reservation and executor hand-off is now released immediately
  instead of being held until stale cleanup.
• MemoryBackend: streaming token refunds are reconciled against header
  syncs via a per-key state version, preventing capacity over-admission when
  a server header update superseded a reservation.
• Streaming: on a successful wrap, reservation ownership is handed to the
  iterator so stale-reservation cleanup cannot reclaim a live stream.

What's Changed

• fix: resolve confirmed audit issues (concurrency, lifecycle, packaging, API boundary) by @sethbang in #4
• fix: eliminate un-awaited AsyncMock coroutine leaks (CI flake) by @sethbang in #5
• ci: bump codecov/codecov-action from 5 to 6 in the actions group across 1 directory by @dependabot[bot] in #1
• deps: update nox requirement from <2026.0.0,>=2025.11.12 to >=2025.11.12,<2027.0.0 by @dependabot[bot] in #2
• deps: update mypy requirement from <2.0.0,>=1.19.1 to >=1.19.1,<3.0.0 by @dependabot[bot] in #3
• chore: add main-branch protection ruleset by @sethbang in #6
• chore: release 1.1.0 by @sethbang in #7

New Contributors

• @sethbang made their first contribution in #4
• @dependabot[bot] made their first contribution in #1

Full Changelog: v1.0.2...v1.1.0