Apply parallel worker memory budgets on Windows

[Tairitsu-Aya]

Change and mathematical scope

Native Windows has no os.sysconf, so run_parallel.memory_gib() returned None and the default worker count ignored the documented three-GiB-per-worker budget. On the tested Windows host, the old code chose 32 workers; querying 31.297 GiB of physical RAM now selects 10. The shared verify_parallel.default_jobs() also receives the detected capacity.

Use the standard-library ctypes binding to GlobalMemoryStatusEx, with the SDK's fixed-width fields and initialized structure length. Preserve the POSIX query, explicit JOBS overrides and unknown-memory fallback. This restores the existing total physical memory heuristic; it is not a free-memory reservation or a hard resource limit. No scientific input, certificate, selected result or kappa changes.

Task and overlap

• Base: c676548106012571fefa1dce0ec16f1bcfe11dce.
• Card: verifier-engineering / portable-replay, priority revision 3, scope key 831cc1a3c8f9ff01b197, chooser nonce cdb3518b3b43e210c16c478540c21dea. Scope refined to one native Windows resource-detection failure in the existing verification tooling.
• Coordination claim #2030.
• The existing worker policy comes from eumemic's Faster verification: concurrent checks, make verify-parallel, faster paired-cube frames, no joint-dual re-pins #154 and its integrated validation/snapshot hardening. Merged Tooling: Windows portability for research onboarding scripts and tests (no new kappa) #1298 (e76026571dbfa7299bf474b941c9abeb5bd556b3) addresses Windows onboarding encoding/process cleanup in different files.
• Submission refresh at 2026-10-14 17:46-17:51 UTC covered 361 open PR bodies, 676 Discussion records, 362 open-PR file inventories and 1,448 local metadata records, with complete pagination and no warnings. Text search matched only this session's claim; no open PR reported either changed path. These are bounded text/file-list checks, not a novelty guarantee.

Validation

The new tests produced five failures against the unchanged base implementation. After the fix:

Environment	Check	Result
Windows, CPython 3.12.9	python -X utf8 -B -m unittest discover -s tests -p test_parallel_memory.py -v	11 pass, including the real Windows API
Windows, CPython 3.14.2	Same command	11 pass
Ubuntu 24.04 under WSL, CPython 3.12.3	python3 -B -m unittest discover -s tests -p 'test_parallel*.py' -v	17 tests, one Windows-only smoke test skipped; all others pass

Controls cover a fractional capacity above 4 GiB, total versus available memory, API failure with a populated buffer, zero/unknown capacity, CPU and minimum-worker bounds, explicit overrides, and the shared isolated-group budget. The existing Windows-safe scheduler tests also pass (four tests covering overrides, failure propagation, snapshots and invalid CLI/group inputs).

python -X utf8 -B scripts/research_baseline.py, python -X utf8 -B scripts/verify_selected_result.py and git diff --check pass. The source-bound baseline reports the unchanged selected record. No heavy audit workers were launched for the scheduler comparison. Full historical/mathematical replay was not run for this tooling-only change. Hosted CI is separate from these local results.

Hosted CI currently requires maintainer approval: Verify research artifacts reports action_required with no jobs executed. The local results above do not claim hosted CI success.

Credit

Retain the existing Apache-2.0 notice and credit eumemic's PR154 worker policy and the integrated maintenance work. API layout and return semantics follow Microsoft's MEMORYSTATUSEX and GlobalMemoryStatusEx documentation. Prepared for Tairitsu-Aya with substantial OpenAI Codex assistance.

Open next questions

This is an existing-tool repair, not a new research package. Accounting for competing processes, free RAM or container/job-object limits would require a separate scheduling-policy change; this patch preserves the current total-memory policy.