- Nested fan-out under the fork start method is fixed: a worker that
  performed a second submit() while a sibling Future was still registered
  crashed with "can only join a child process" because under fork no
  pickling occurs, so __setstate__ never ran and the worker inherited the
  parent's selector (and sibling Process objects belonging to an ancestor).
  The selector is now built lazily through a single _lazy_selector()
  accessor that rebuilds a clean per-process selector whenever
  _selector_pid no longer matches os.getpid(), so a forked worker discards
  the inherited selector before reclaim_resources() or submit() can touch a
  non-child process.  A regression test exercises nested fan-out across all
  start methods.

- Pickle-dump error handling is shared rather than duplicated: the
  (PicklingError, AttributeError, TypeError) classification, previously
  spelled out in both _worker_entrypoint and the executor's
  _responses_put_failed with comments asking the two sites to agree, is
  hoisted into a single PICKLE_DUMP_ERRORS constant in _compat so the
  contract is enforced by one definition.

- An EnvItems type alias deduplicates the env signature: the four-line env
  Union previously repeated verbatim across __init__, submit, map,
  _env_coerce, and _map_generate is collapsed into one alias (and
  Optional[EnvItems] where None means "use the instance default") so the
  signatures read uniformly.

- A draft benchmark harness is added under draft/bench.py: a stdlib-only
  program that exercises six semi-realistic CPU-bound workflows (flat,
  nested, callbacks, sleepfn, executor, and cancel) and prints aligned
  tables of wall time, throughput, and speedup versus a serial baseline.
  All kernels are module-level and picklable under spawn/forkserver, it
  supports --quick/--full/--slots/--context/--only, and it runs via
  PYTHONPATH=src.  It lives under draft/ so the CI script does not pay its
  multi-second runtime.