Keep runtime GPU context handles out of retained callable history

[trissim]

Summary

Fix the cannot pickle '_thread._local' failure reproduced from a retained,
unpublished memory-decorated callable in OpenHCS history (OpenHCSDev/openhcs#169).

Keep GPU context storage on the existing importable ThreadGPUContext runtime
owner, not in decorator function globals that Dill persists by value. Delete
the old global handle/forwarder and call ThreadGPUContext.current() directly.
No filtering of history, permissive pickle handling, alternate registries,
GPU-framework import or change to the durable history format.

Evidence

• Before: fresh unpublished @numpy callable and real ObjectState history
  export both raise TypeError: cannot pickle '_thread._local' object.
• After: restored callable executes correctly; the runtime thread-local handle
  remains identical and outside the serialized declaration.
• Focused source test run: python -m pytest -q -o addopts= tests/test_durable_decorator_context.py: 3 passed (0.15s), using OpenHCS's
  existing environment with source imports from the isolated ArrayBridge checkout.
• Changed-source Ruff and diff checks pass.
• Paired OpenHCS continuous native declaration/history save, fresh-registry
  restore, historical undo and return to head also passes with an unpickleable
  runtime handle retained. This uses the real document/ObjectState owners.

Integration/remaining scope

Paired OpenHCS gitlink/regression/diagnostic draft:
OpenHCSDev/openhcs#208 (references #169 and #131).
This draft is a source fix, not installed/live desktop acceptance. Coordinator
owns merge/install and serialized continuous real capture/restore validation.
No old H001 export or history is modified or reconstructed. No Closes claim.
Both PRs remain draft pending source/live journey and catalog-pattern review.
The paired validation note documents the concrete BOUND-2/BOUND-1 and MEMB-2
diagnostic revision plus its new-case ownership test, not just skill reads.
New heavy validation is held for Euler's serialized H003 run; existing source
checks have finished and released the shared lock. No services are restarted.

[trissim]

Retained callable serialization: parent source review

Reviewed ArrayBridge PR1 head
06837ec267e3ca0b1734461d09919ff65cb1b1fe against remote main
fddd9857bcd21468d20bc8547a769e0c6c9f1906. Paired OpenHCS PR208
is d69d65adbd56633e1d1fd2a3b65678728cd6ada8. Linnaeus retains
the implementation tree; parent owns integration, installation and live checks.
Neither implementation tree was edited during this review.

Ownership and durable state

The existing importable ThreadGPUContext owns its thread-local runtime handle
and current-context lookup. The global handle and forwarding function are
deleted. Decorated unpublished callables no longer pull the thread-local
handle through function globals into durable history. No history filtering,
permissive serialization, alternate registry, compatibility reader or stored
history format change is introduced. The GPU stream/device implementation is
unchanged. New cases use the existing memory-type owner and context mechanism;
there is no second roster or copied framework decision in this patch.

Evidence and its limits

• Parent independent run: four tests PASS in 3.030 seconds, zero failures,
  errors or skips; a 15-second process bound. XML:
  retained-callable-parent-source-tests-20260930.xml.
• Three ArrayBridge cases exercise per-thread context identity and Dill
  round-trips of unpublished NumPy/CuPy decorated callables. Restored NumPy
  invocation executes and returns [4, 8]; CuPy execution is NOT tested.
• The fourth test uses real ObjectStateRegistry, PipelineObjectStateBinding
  and PipelineDocumentAuthority: save current declarations and full history,
  clear/load the registry, navigate to a prior unpublished callable, execute
  [7, 14], return to head, and preserve the unpickleable runtime handle.
  This is a same-process fresh-registry journey, NOT a fresh application
  process or actual live GUI restart.
• Original packaged structural ratchet, remote ArrayBridge main to reviewed
  head: PASS, 38 projected metrics, all deltas zero, one production source
  path. This is a structural screen, not a complete NRA proof.
• Production diff reviewed in full: decorators.py, plus a development-only
  Dill dependency and the tests. Both source working trees are tracked clean.
• Remote main protection disabled and rulesets empty at review. Optional
  hosted CI is deferred per owner, without an administrative override.

Ship the reviewed ArrayBridge source checkpoint. OpenHCS PR208, issues169/131,
current-main integration, installed capture/restore and MCP retention/slope
measurements remain unfinished. No merge/install/live or biological acceptance
is implied by this source review. Current resource assertion exits2: swap14.6
GiB, RAM18.5 GiB available; no new heavy/native/parallel job is launched.

Merge evidence will be appended after verifying remote authoritative state.