DONE: Windows physical-memory budgets for parallel verification (#2035)

[Tairitsu-Aya]

DONE / RELEASED — PR #2035

Exact contribution head: ddd8690874ecab0eea671d4c68de8bbb3ed8b991.

The native Windows physical-memory query now restores the existing scheduler budget. On the tested 32-logical-CPU host, the unchanged base returned unknown RAM and chose 32 workers; the fix detects 31.297 GiB and chooses 10. The patch changes one helper and adds regression tests; no heavy audit workers were launched for this comparison.

Evidence: five new-test failures against the old implementation; all 11 new tests pass on Windows CPython 3.12.9 and 3.14.2, including the actual API. The combined Linux/WSL Python 3.12.3 parallel suites run 17 tests with one Windows-only smoke test skipped and all others passing. Four existing Windows-safe scheduler tests also pass. Source-bound baseline, selected-record integrity and diff checks pass.

Submission overlap: 361 open PR bodies, 676 Discussion records, 362 changed-file inventories and 1,448 local metadata records. Full pagination; no warnings. Text search matches only this claim and no open PR reports the two changed paths. The documented scope/coverage limits remain.

Limitations: this retains total-physical-memory policy and explicit JOBS overrides; it does not reserve free memory or account for container/job-object limits. No mathematical result or full cold replay is claimed. Hosted CI is awaiting maintainer approval before jobs run: https://github.com/CrocSwap/integer-mult-bounds/actions/runs/38610219719.

Claim released. Prepared for Tairitsu-Aya with substantial OpenAI Codex assistance; predecessor credit is in the PR.

---

Original claim:

CLAIM — Windows physical-memory detection for parallel verification workers

Question: Does scripts/run_parallel.py apply its documented three-GiB-per-worker default on native Windows? On CPython 3.12 / win32, os.sysconf is absent, memory_gib() returns None, and the existing runner selects 32 workers on a machine with about 31.3 GiB physical memory. This is a scheduler observation; no heavy worker batch was launched.

Base: c676548.
Track/card: verifier-engineering / portable-replay, priority revision 3; chooser nonce cdb3518b3b43e210c16c478540c21dea; scope key 831cc1a3c8f9ff01b197.
Scope refined to one OS-specific detection failure in the existing runner and its shared isolated-group scheduler.

Method and bounded domain: use the standard-library Windows physical-memory API; preserve POSIX behavior, explicit JOBS overrides and unavailable-API fallback. Add failing-before/passing-after controls for values above 4 GiB, API failure and both worker budgets, plus native Windows and Linux/WSL tests. The donor budget is one 60-minute session; final verification/handoff starts no later than 18:17 UTC.

Nearest work: eumemic's PR154 and the integrated worker validation/snapshot hardening. Merged PR1298 (e760265) fixes Windows onboarding encoding/process cleanup in different scripts. Targeted all-state searches found no matching worker-memory change. The repository overlap helper is still collecting full Discussion and changed-file coverage; I will reconcile any match before submission.

Expected artifact: one focused maintenance PR with a reproducible scheduler regression and cross-platform evidence. No mathematical result, selected frontier or frozen source changes; no full mathematical replay claim.

Claim expires: 2026-10-14 18:32 UTC.
Prepared for Tairitsu-Aya with substantial OpenAI Codex assistance.