feat(desktop): align connection tests with the generate invoke contract

[Lxr-max]

Summary

Connection testing was a separate /models probe, so it could pass when generate would fail (wrong wire/auth/baseUrl) or fail when generate would work (/models missing). This PR extracts a shared effective invoke contract and makes connection:v1:test-active / connection:v1:test-provider a low-cost validation of that same contract.

• test-active uses resolveActiveModel() (including snap-to-canonical-active) plus resolveCredentialForProvider() — the same path as generate, including ChatGPT Codex getValidAccessToken() refresh.
• test-provider resolves the named provider with the same auth, header, canonical baseUrl, wire, and keyless rules.
• GET /models is defined as optional discovery. Runtime generate never calls it. Missing /models + working inference is degraded / degraded-discovery. /models success + failed inference is ok: false with compatibility: "diverges" so the renderer can tell “reachable but not the real invoke path”.
• Generate, memory, and complete() now canonicalize baseUrl with the same helper the connection test uses.

Type of change

• Bug fix
• New feature
• Refactor (no behavior change)
• Documentation
• Build / CI / tooling
• Breaking change

Linked issue

Fixes OpenCoworkAI#216

Refs OpenCoworkAI#214

Does not implement sibling OpenCoworkAI#213 (capability-aware diagnostics beyond invoke-parity).

Checklist

• I checked the linked issue / relevant context before starting
• pnpm lint && pnpm typecheck && pnpm test passes locally (pre-push hook: typecheck + lint + full unit suite)
• Added/updated tests for the change
• Added a changeset (pnpm changeset) if user-visible
• Updated docs if behavior changed (i18n strings for the new toast states)

Principles 5b

• Compatible: existing IPC ok / probeMethod / compatibility fields remain; new invokeParity, invokeContract, and diverges are additive.
• Upgradeable: schemaVersion-bearing config is unchanged; connection-test payloads stay v1.
• Lean: no new shipped dependencies; one shared contract module instead of a second probe stack.
• Elegant: one resolveEffectiveInvokeContract + inferenceEndpointUrl instead of parallel credential/URL logic.

Tests

Regression coverage includes:

1. test-pass / runtime-fail: /models 200 with a dead invoke URL (including openai-responses vs chat-only gateway) reports compatibility: "diverges" instead of Connection OK.
2. test-fail / runtime-pass: /models 404 with working /chat/completions still reports degraded success (degraded-discovery).
3. Parity: ChatGPT Codex OAuth (getValidAccessToken, no /models HTTP probe), keyless proxy (empty bearer), OpenAI-compatible gateway (canonical /v1 + Bearer).

Screenshots / recordings (UI changes)

Settings now uses structured toasts: success, degraded-info, and a distinct “reachable but generate would fail” error. No layout change; not verified in Electron (no desktop session in this environment) — covered by connection-test-toast.test.ts.

  

[Lxr-max]

Superseded by the upstream PR: OpenCoworkAI#421 (Fixes #216). Closing this fork-local PR so review happens on OpenCoworkAI/open-codesign.