refactor(output): make collective HDF5 observer-lane owned (ADC-683)

[wolf75222]

Stack

This draft PR is stacked on #643 at commit 0ac4942. It preserves three separate thematic commits:

• 40a0b07 — require observer-owned HDF5 communication
• c32d76e — fence observer-owned HDF5 lanes
• 836f2f2 — document duplicated HDF5 observer lanes

What changes

• removes collective-HDF5 overloads that accepted WorldCommunicator;
• requires an explicit CommunicatorView in C++ and an active duplicated ObserverMpiLane in Python;
• prevents the Python writer from accepting the world communicator fallback;
• makes parallel HDF5 capability reporting independent from process-global MPI_Initialized discovery;
• updates integration and architecture tests to fence the ownership contract;
• documents that HDF5 collective output owns its observer lane.

The retained HDF5 mutex is intentional encapsulation of HDF5 library-global/thread-safety behavior. It is not communicator discovery or borrowed simulation state.

The touched HDF5 C++ implementation was normalized by clang-format, which accounts for most of the line-count churn; the semantic change remains limited to communicator ownership and capability discovery.

Validation

• architecture batch: 11 passed
• HDF5 unit subset: 6 passed, 87 deselected
• manifest/CI architecture subset: 4 passed
• Ruff: passed
• Python bytecode compilation: passed
• C++ header syntax-only check: passed
• clang-format dry-run: passed
• docs lint: passed (41 Markdown files and 11 docmap files; one pre-existing stale-docmap warning)
• git diff --check: passed

No shared build or install was modified.

Explicit non-claims

• this does not close ADC-683;
• this is not a complete audit of every remaining process-global device, MPI, or I/O state;
• no native MPI/HDF5 build matrix was run in this source-only tranche;
• no GPU or 3D support claim is made;
• the C++ source surface intentionally removes the WorldCommunicator overloads, while the ABI descriptor schema is unchanged.