better_effect 0.4.0

Published on pub.dev: https://pub.dev/packages/better_effect/versions/0.4.0

• Added the repository-level normative semantic contract and an independent
  cross-package conformance suite mapping stable lifecycle/outcome rules to
  executable rule IDs.
• Semantic compatibility changes now require the matching conformance scenario,
  affected package changelog entries, and a semantic migration ledger entry.
• Added Runtime.fork for feature-length child environments with local-first
  resolution and parent fallback.
• Child Module resources survive multiple executions, close exactly once with
  the child Runtime, and are coordinated before parent resources during parent
  shutdown.
• Runtime observer contexts now expose runtimeId, parentRuntimeId, and
  runtimeLabel so feature and execution boundaries can be correlated.
• Nested execution overlays preserve the complete parent fallback chain.
• Added typed-failure Effect.retry with none, fixed, linear, and
  exponential policies, optional full jitter, maximum delays, and
  explicit attempt-count semantics.
• Each retry attempt owns a child Scope that closes before policy
  evaluation or delay; cleanup defects retain Runtime precedence and
  prevent a broken typed-failure attempt from retrying.
• Added optional EffectClock and EffectRandom contracts with host,
  seeded, and manual test implementations; no service is installed
  implicitly.
• Added observable RetryEvent decisions, deterministic tests, package
  and website guides, and a retry benchmark.
• Added lazy Effect.all and Effect.forEach collection composition with
  sequential-by-default and positive bounded concurrency.
• Added explicit Effect.allUnbounded and Effect.forEachUnbounded APIs so
  unbounded fan-out cannot happen through an omitted limit.
• Collection results preserve input order and are returned as unmodifiable
  Lists, while scheduling remains FIFO.
• Typed failure selection is deterministic by lowest started input index; no new
  work starts after failure or interruption, and already-started operations
  remain physically owned until completion.
• Collection defects stay defects, worker resources retain normal Scope cleanup
  semantics, and sequential/bounded/unbounded benchmarks are included.