ci: fix both 7502 candidate build failures (macOS XCTest hosted-window flag, Windows NSIS RestartManager template)

[raydocs]

Owner order 2026-10-07 (decision 071): build the 0.0.75 candidates from main with Mac polish A only.

The first 7502 macOS candidate, run 37651992656, failed in the build job's "TonoTests XCTest target" step: 642 tests, 14 failures, all MacUsabilityRenderTests.swift:560 … exact-window native acceptance unavailable; TEST_RUNNER_TONO_HOSTED_WINDOW_DIAGNOSTIC=1 required. The sea render fixtures from #1426 fail closed without that flag; macos-ci.yml sets it on both XCTest steps, macos-release.yml did not. Nothing was signed or packaged.

Change: the release workflow's XCTest step gets the same env as CI. No test is skipped or loosened.

Verification

• Failure evidence: the run above (gh run view 37651992656 --log-failed).
• Proof of the fix is the macOS release run dispatched after merge on the fast-forwarded release/macos; this PR's ci-gate has no called workflow for .github/workflows/macos-release.yml (BUILD_AND_TEST: release workflows need no called workflow) and is not evidence that the step passes.
• Not run: xcodebuild (MacBook does not build natively).

Candidates

• release/macos, release/windows, stability/desktop-0.0.75-20261005 are at f73300c13; after merge they are fast-forwarded to this PR's merge SHA and both platforms are rebuilt from it with release sequence 7503. The paired Windows 7502 run 37651997330 (environment windows-release self-approved per decision 023 at 2026-10-07T16:36Z) is superseded, not used.

Limitations

• macos-release.yml and macos-ci.yml carry two hand-written copies of the XCTest step; folding them into one reusable workflow is a later engineering item.

Record: docs/changelog.d/2026-10-07-macos-release-hosted-window-flag.md (SHIP_PLAN G1/G2 candidates).

🤖 Generated with Claude Code

https://claude.ai/code/session_01TCaYZLcT8rtQDuuTVhSPaC

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[raydocs]

jev-route review record — decision 789d75bb

• Review decision: dual_cross_family via jev, confidence 0.27
• Slots: opus → anthropic (high effort, success); codex → openai (high effort, success); substitutions: none
• Verification: none recorded; confirmed 0, refuted 0, downgraded 0; rechecked: a9401457/opus:F1
• Chain: a9401457 PASSED → 789d75bb PASSED; fix rounds: 1
• Stop rule: qualifies; 0 remaining limitation(s)
• Head e7412163027a99379f4a410566bf9650b66905c6, range origin/main...HEAD

Verification by the merging agent

• Round 1 a9401457 (head ddba57a): 1 confirmed minor (changelog said MacSeaPolishRenderTests fails closed; it uses XCTSkipUnless) → fixed in e741216.
• Round 2 789d75bb (head e741216): clean on both vendors; the NSIS change is the upstream tauri-cli-v2.12.1 template delta (!include "Win\RestartManager.nsh", $INSTDIR\${MAINBINARYNAME}.exe at both call sites).
• Local: node --test apps/windows/app/scripts/windows-packaging.test.mjs 38/38 on the edited template. Not run: xcodebuild, cargo, makensis (hosted only). Proof of both fixes = the 7503 release runs after merge.