Windows: make grok + codex implementer lanes work headless

[tr3lane]

• run_capped: cross-platform hard wall-clock; Windows tree-kills via taskkill //T //F on the winpid; *nix uses validated GNU coreutils timeout (never trusts system32 timeout.exe). Foreground-only to avoid orphaned marker-poll loops.
• grok --permission-mode auto (was acceptEdits, which silently drops every write headless on Windows).
• codex --sandbox read-only on Windows (no OS sandbox there): runs as a read-only reviewer returning a git-apply-able patch (STATUS: proposal); grok is the Windows writer. macOS/Linux keep workspace-write.

*nix behavior unchanged. Tested headless on Windows 11 / Git Bash.

[tr3lane]

Re-tested this today against a newer codex-cli, since "just update codex" is the obvious objection to the read-only branch in this PR.

Still broken on codex-cli 0.144.6 (2026-07-21). Trivial write task, workspace-write requested explicitly:

$ codex exec --sandbox workspace-write --skip-git-repo-check --cd <scratch-dir> -
OpenAI Codex v0.144.6
--------
workdir: <scratch-dir>
model: gpt-5.6-sol
approval: never
sandbox: read-only          <-- workspace-write was what got passed
--------

No file was written. Same behavior previously confirmed on 0.143.0 and 0.144.1, so this isn't a transient regression: workspace-write depends on OS-enforced sandboxing (Seatbelt/Landlock) that Windows doesn't provide, and codex declines to run unsandboxed rather than silently writing outside a sandbox. Bumping the CLI does not fix it — which is the case for keeping a Windows branch in the lane rather than waiting it out.

Also worth flagging that #2 and #5 hit these same bugs independently from different setups: #2 makes the same acceptEdits -> auto change (headless file writes are otherwise silently dropped — exit 0, success narrated, no file on disk), and #5 addresses the codex sandbox denial with a fail-loud + opt-in fallback instead of this PR's read-only-reviewer route. Three separate reports of the same two failures suggests it's worth picking a resolution for each, whichever shape you prefer.

Can rework this along #5's lines, or split the wall-clock/timeout fix into its own PR, if either is easier to review independently.