v8.2.0

Changed — snapshot/checkpoint boundary isolation (backlog #8)

• Pause checkpoints are now deep-copied at creation. getCheckpoint() /
  the PausedResult.checkpoint previously embedded the LIVE SharedMemory
  context as sharedState (the alias only detached at the next commit —
  post-pause, never), and executionTree nodes referenced live diagnostic
  bags. The docs said "store the checkpoint in Redis" while handing out live
  references. Now the whole checkpoint is detached via one structuredClone
  at pause time: mutating a checkpoint you hold cannot corrupt engine state,
  and a later same-executor resume cannot mutate a checkpoint you already
  persisted. Edge-case consequence of the (pre-existing) JSON-safe contract:
  a pauseData value that is not structured-cloneable (e.g. contains a
  function) now throws a descriptive contract error at pause time —
  naming the offending checkpoint field, with the raw DataCloneError as
  error.cause — instead of silently surviving in-process and breaking on
  real persistence.
• Non-cloneable diagnostics never abort a pause (clone resilience).
  $debug/$error/$metric/$eval accept ANY value at write time without
  cloning, so a function logged in any stage of a pausing run would have
  made the whole-checkpoint structuredClone reject with a naked
  DataCloneError — swallowing the human-in-the-loop pause. Now, on clone
  failure, the checkpoint's executionTree diagnostic bags
  (logs/errors/metrics/evals) are sanitized — non-cloneable values
  become '[non-serializable: function]'-style marker strings in the
  CHECKPOINT only (live engine diagnostics keep the raw value) — and the
  pause succeeds. If the clone still fails after sanitization, the
  violation is in consumer-owned data (realistically pauseData; functions
  in shared state, e.g. via the initialContext option, already reject at
  stage entry when the transaction buffer clones the context) and the
  descriptive contract error above is thrown. A naked DataCloneError
  never escapes the executor. Ingress points documented in
  docs/guides/execution-model.md.
• resume() clones the checkpoint in. checkpoint.sharedState and
  checkpoint.subflowStates are deep-copied before seeding engine runtimes
  (mergeContextWins copies only the top level), so the engine never holds
  a live reference into the caller's checkpoint object — caller mutations
  can't reach the resumed run, and engine writes can't bleed back.
• Dev-mode mutation guard on getSnapshot(). When enableDevMode() is
  on, getSnapshot().sharedState is a deep-frozen CLONE — consumer mutation
  throws a TypeError instead of silently corrupting engine state.
  Production behavior is unchanged: sharedState remains the zero-copy live
  view — treat it as read-only (documented in
  docs/guides/execution-model.md).
  Clone-always in production is deferred pending the Phase-3 bench
  (~4 ms per structuredClone of a 1 MB state, Node 22/M-series).