3.1.1

modern-di 3.1.1 — containers stop being garbage the refcounter can't free

A single-line fix in Container.__init__ with a disproportionate effect. Every
container stored itself inside its own scope map, which made it a reference
cycle: no container — root or child — could be freed by reference counting, so
each one waited for a generational GC pass. An application that builds a
container per request was handing the collector work at its request rate.

No API changes. No integration needs updating.

Fix

• Containers are no longer reference cycles. Container.__init__ seeded
  _scope_map with {scope: self}, so the container referenced a dict that
  referenced the container. The map is now seeded from the parent instead — it
  holds ancestors only. find_container already returned self for its own
  scope before consulting the map, so the self-entry was never read and
  resolution is unaffected.

  Measured on an Apple M4 / CPython 3.14.6:

  	before	after
  100 closed REQUEST children	792 objects reclaimable only by the GC	0
  100 closed root containers	2300	0
  100 full APP→SESSION→REQUEST→ACTION→STEP chains	5500	0
  build a child and drop it, GC enabled	~849 ns	~542 ns

  Isolated construction is unchanged within noise (~500 vs ~506 ns with the GC
  off); the gain is the collector no longer reclaiming what reference counting
  now frees. On the comparative request-lifecycle benchmark the median fell from
  232.7 µs to 194.8 µs per 100-request batch, and the standard deviation from
  123.0 µs to 7.9 µs — the tail that scenario carried was GC, and it is gone.

Behavior change

• Container.scope_map no longer contains the container itself. A root's
  map is now empty and a child's holds only its ancestors. The property is
  documented as an internal surface with no stability guarantee and already
  emits a DeprecationWarning; a survey of all 12 sibling modern-di-*
  integration repositories found no references to it. find_container is
  unaffected on both its own-scope and ancestor paths.

Internal

• The benchmark suite was rebuilt for measurement fidelity. Nine defects
  found by audit, none of them affecting library code. The largest: the
  request-lifecycle scenarios entered the asyncio event loop once per timed
  iteration, and that entry costs ~27 µs regardless of the body — so ~93% of
  every published figure was a shared constant that compressed the real
  differences between frameworks toward parity. Scenarios are now batched, the
  per-framework setup asymmetries are levelled, iterations is pinned so no
  cell sits on the platform timer's ~42 ns grid, and the published comparison
  table is generated by just bench-report rather than hand-assembled.

  This changed a published claim: modern-di is not level with dishka on the
  request lifecycle, as the old numbers implied. See
  Performance.

• The reference-cycle fix above was found while investigating why modern-di's
  request-lifecycle measurement was the least stable cell on that page.

Downstream

No action needed. No API changed, and no integration reads the affected
property.

Internals

• 450 tests, 100% line coverage, Python 3.10–3.14 including the free-threaded
  3.14t build.