fix(windows): patch platform._syscmd_ver to survive non-UTF-8 output under PEP 540

[webtecnica]

Fixes #69413

Description

On Windows, the OpenAI SDK generates platform headers by calling platform.platform(). This triggers:

platform.platform()
→ platform.win32_ver()
→ platform._syscmd_ver()
→ subprocess.check_output("ver", text=True, encoding="locale", shell=True)
→ subprocess.Popen._readerthread

When PYTHONUTF8=1 (PEP 540 UTF-8 mode) is active — which Hermes sets on Windows — locale.getencoding() returns "utf-8". However, the Windows ver command outputs text in the system's active ANSI code page (e.g. cp1252 on US-English). Bytes that are valid in cp1252 but not in UTF-8 — such as 0xe9 (é in cp1252, an invalid UTF-8 start byte) — raise UnicodeDecodeError inside the subprocess reader thread.

Fix

Added _patch_platform_syscmd_ver() to hermes_bootstrap.py that wraps platform._syscmd_ver with a try/except UnicodeDecodeError on Windows. When the ver command output can't be decoded as UTF-8, the patched function gracefully returns its input defaults.

Applied automatically on module import (every Hermes entry point already imports hermes_bootstrap first). No-op on POSIX.

Changes

File	Change
hermes_bootstrap.py	Added _patch_platform_syscmd_ver() function + module-level call
tests/test_hermes_bootstrap.py	Added TestPatchPlatformSyscmdVer class (5 tests)

Test results

All 23 tests pass, 5 Windows-only skipped on Linux.

Closes #69413

[isak-ialogics]

Triage evidence only (not a merge decision): this PR currently contains commit 47bad20, which is the complete head commit of the separate #69521. As a result, the diff here includes unrelated changes to agent/chat_completion_helpers.py, run_agent.py, and tests/agent/test_non_stream_stale_timeout.py in addition to the Windows fix described by this PR. Concrete next action: rebase this branch onto current main and retain only fde3e142 (or otherwise drop 47bad20), then rerun CI so #69522 is scoped to hermes_bootstrap.py and its regression tests.

[teknium1]

Closing after a deep verification pass — the diagnosis deserves a detailed response because the issue you chased is partially real.

What's correct: on Python 3.11.0/3.11.1, platform._syscmd_ver lacks CPython's encoding="locale" fix (added in 3.11.2), so under PEP 540 UTF-8 mode the cmd /c ver output is strict-utf-8 decoded and raises on OEM-code-page bytes. That crash path was live in processes that don't run hermes_cli.main (slash workers, tui_gateway entry, run_agent).

Why this PR couldn't merge as-is: (1) the wrapper's supported_platforms=None default is forwarded into the original, which does sys.platform not in supported_platforms → guaranteed TypeError on every no-arg call — platform.win32_ver() invokes it exactly that way, so on Windows the patch breaks platform.platform() outright (the tests mock past this); (2) for Python ≥3.11.2, locale.getencoding() explicitly ignores UTF-8 mode, so the PR's core claim doesn't hold there; (3) the branch bundles an unrelated commit (#69521's stream-timeout changes).

What shipped instead (PR #71014): the existing full-stub suppress_platform_ver_console() now runs from hermes_bootstrap, covering every entry point — win32_ver() falls back to sys.getwindowsversion(), fixing both the decode crash on the old micros and the console flash, with a regression test pinning the no-arg-call contract. Thanks for surfacing the coverage gap — the bootstrap move exists because your PR pointed at the right set of exposed processes.