fix(ci): green the Windows leg, cap openai<3, surface desktop failures

[agentforce314]

Diagnosed from the actual run artifacts (junit XMLs pulled anonymously for the last four PR runs; the deterministic windows-latest failure set was stable across all of them, plus one cross-platform break that appeared this morning). Three red jobs, three causes:

1. openai 3.0.0 (released 2026-08-12, hours before today's runs) broke BOTH
   pytest legs. The repo floats openai>=1.109.1, and v3's httpx2 migration
   moves requests off the httpx layer we are coupled to on both sides of the
   SDK boundary: http_client=httpx.Client(verify=False) stops carrying the
   transport, and the stream-abort guard closes an httpx Response that no
   longer exists behind SDK streams. Verified in a fresh floating venv:
   chat() under a patched httpx.Client.send sails straight past the spy
   onto the real network — exactly the CI failure in
   test_xai_requests_go_to_chat_completions ("no HTTP request was
   attempted"). Capped openai>=1.109.1,<3 (requirements.txt + pyproject,
   same precedent as the deliberate mcp<2 cap); with the cap (2.54.0) the
   test passes. Lift only with a real httpx2 migration.

2. windows-latest pytest: the stable 8-test failure set plus one flake.

   • Bare ["bash", "-lc", ...] argvs resolve to the WSL launcher on the runner (exits immediately, no distro): supervisor + reaper tests now spawn via shell_platform.bash_argv / popen_tree_kwargs, the way production does (test_bash_timeout_vs_esc, test_kill_shell_for_agent).
   • test_bash_timeout_vs_esc keeps the pre-port < 3.0s kill bound on POSIX; on Windows taskkill /T cannot reach the MSYS2 sleep grandchild, so elapsed is bounded by deadline + one _KILL_REAP_TIMEOUT_S drain and the timed_out/interrupted flags carry the regression guard there.
   • activate_conditional_skills_for_paths: os.path.relpath raises ValueError cross-drive on Windows (runner workspace on D:, TEMP on C:); a path we cannot express relative to cwd is outside it — skip it like the ../absolute guards (src/skills/loader.py; the D:-drive condition is unreachable on single-drive dev boxes, exercised directly).
   • test_reaper_terminal_state_is_eligible spawned with cwd=Path("/tmp"), which only works where C:\tmp happens to exist — the runner throws NotADirectoryError (WinError 267). Now tempfile.gettempdir(), and the wait loop requires notified+evict_after (the reaper commits those in a second registry update; breaking on terminal status alone raced it).
   • test_user_migration_preserves_symlinks_as_symlinks compared os.readlink literally against a target that Windows returns with a ?\ prefix; compare resolve() on both sides, keep is_symlink().
   • test_failure_retries_with_exponential_backoff asserted wall-clock gaps between real asyncio sleeps; Windows' ~15.6ms timer granularity ate the 75ms-floor-under-80ms margins. Now captures the delays the uploader schedules (instance-attribute _sleep shadow) and pins the exact 20/40/80 schedule — deterministic on every platform, and unlike a lower bound it also catches delays that are too long.
   • test_stream_watchdog: _on_timeout sets fired under the lock but closes the response after releasing it — deliberate ordering the consumer relies on — so a fixed sleep then close.assert_called() races that window under runner load (the observed flake). Added a bounded _await_fire poll and applied it to the three same-pattern siblings.

3. Desktop (windows-latest) has been red since the job landed and its
   failures were invisible: vitest job logs need an authenticated API call
   and the job uploaded no artifact. The run step now writes a junit XML
   that the existing Test Results publisher picks up automatically (it
   globs every artifact), so the next red names its tests. Best-evidence
   fix applied alongside: the electron vitest project gets the same
   testTimeout: 30_000 the ui project already documents for Windows
   cold-start starvation — its tests shell out for real (git clone/worktree
   fixtures, spawned helpers), which a 2-core Defender-scanning runner
   starves past the 5s default. Not reproducible locally: full vitest is
   green here both against the working tree and a pristine worktree +
   fresh npm ci.

Verification: full pytest suite on Windows 10,000 passed / 0 failed; all touched test files re-run green after every edit; xai test verified failing on 3.0.0 and passing capped in a CI-equivalent venv; desktop typecheck + full vitest green twice; junit reporter flags validated against vitest 4.1.10. POSIX semantics preserved by construction (relpath never raises there, bash_argv is -lc identical, tight timing bound restored, watchdog polls assert the same fired→close contract).

All Submissions:

• Notify authors: * @agentforce314, @ericleepi314

New Feature Submissions:

• Have you added an explanation of what your changes do and why you'd like us to include them?

Description

Fixes # .

Changes proposed in this pull request:

[github-actions]

Test Results

     4 files     966 suites   27m 16s ⏱️
14 608 tests 14 591 ✅ 17 💤 0 ❌
29 894 runs  29 839 ✅ 55 💤 0 ❌

Results for commit 6616d70.

♻️ This comment has been updated with latest results.