fix(onboarding): surface gateway wizard terminal-error status instead of silent complete

[indierawk2k2]

Symptom

During first-run onboarding, the gateway wizard can return a terminal-error payload:

{"done": true, "status": "error", "error": "Error: Please enter a valid URL (e.g. http://...)"}

WizardPage.ApplyStep ignored both status and error and collapsed any done:true into the "complete" lifecycle state. The user saw a deceptive green ✅ "Setup complete — click Next to continue" screen with no buttons and no error text — the gateway's error message was silently dropped on the floor.

The user could only escape via the outer Next button, which advanced them through Permissions → Ready → Finish with a broken configuration. At Ready, OnboardingWindow.OnWizardComplete then observed setupStillRequired=true and silently chose not to open the Hub window, leaving the user staring at the desktop with only a tray icon and no UI — perceived as "the app crashed."

Log evidence (Mike Harsh's session, 2026-05-13)

09:22:22.927  Wizard response payload raw={"done":false,"step":{"type":"text","message":"API Base URL",
                "placeholder":"https://api.example.com/v1",...}}
09:22:31.383  [GatewayClient] Sending frame: wizard.next        ← user submitted non-URL value
09:22:31.389  Wizard response payload raw={"done":true,"status":"error",
                "error":"Error: Please enter a valid URL (e.g. http://...)"}
09:22:34.245  [OnboardingApp] Advancing pageIndex 2→3, next route=Permissions
09:22:36.935  [OnboardingApp] Advancing pageIndex 3→4, next route=Ready
09:22:36.961  Settings saved
09:22:38.581  [OnboardingWindow] OnWizardComplete skipping chat launch;
                route=Ready, setupStillRequired=True

No exception, no error UI, no Hub launch afterward. Tray process stayed alive (verified PID 28224, started 09:13:47, still receiving NODE health events through 09:25). No WerFault, no .NET Runtime event log entries.

Fix

src/OpenClaw.Tray.WinUI/Onboarding/Pages/WizardPage.cs — the done:true branch of ApplyStep now inspects status and, when it equals "error", routes into the existing "error" lifecycle state with the server-supplied error string (falling back to the existing Onboarding_Wizard_StepError localization key when the gateway sends status:"error" with no message).

The "error" render branch already renders a ❌ title + errorMsg body + Retry/Skip controls. No new lifecycle state, no new resource key, no gateway-side change. Pattern exactly matches the five other setErrorMsg / setWizardState("error") / SaveState("error", …) triples already in the file.

Deferred (filed separately by the coordinator)

• Wedge B — OnboardingWindow.cs:644-653: when finishedFromReady && setupStillRequired, the window currently closes silently with no replacement UI. This is the perceived "tray crash" moment. Intentionally not included here to keep the diff narrow and the navigation surface untouched.
• Back-navigation gap (Mattingly's Gap A) — the server-driven wizard has no wizard.back RPC, so the user cannot recover from a wrong branch (e.g. picking the wrong provider). The terminal-error wedge fixed here is the immediate cause of Mike's session; back-nav is the structural cause and warrants its own design pass.

Tests

No unit-test harness exists for WizardPage.ApplyStep — the method is bound to React-style UseState hooks inside the WinUI component and is not directly testable. The parallel WizardStepParser service is unit-tested (tests/OpenClaw.Tray.Tests/WizardStepParsingTests.cs) but is not invoked by ApplyStep, so adding a parser-level test would exercise a different code path. Creating a WizardPage harness from scratch was explicitly scoped out of this fix; filing as a follow-up.

Manual validation: the new branch was exercised under the same payload shape that produced Mike's wedge.

Validation per AGENTS.md (Reporting Standard, post-fix)

• Build: ✅ all four projects (Shared, Cli, WinNodeCli, WinUI) succeeded.
• Shared: 1547 / 0 failures / 28 skipped (1575 total) — duration 9 s.
• Tray: 1134 / 0 failures / 0 skipped (1134 total) — duration 45 s.

Recent baseline from PR #310 work was Shared 1455/0/25 and Tray 961/0/0; counts have grown on master since.

Scope note

Strictly orthogonal to PR #310 (uninstall) — branched off master. No overlap with LocalGatewaySetup/, PS scripts, or installer.iss. Single file changed, ~12 lines added.