fix(ai): refresh subscription models and harden OAuth flows

[bobleer]

Summary

Subscription model discovery could show an outdated fixed Antigravity list, hide Codex subscription-only models, mix OpenCode offerings, and prevent retries after failures. Fetch account-scoped catalogs, add explicit refresh and failure states, and fix OAuth expiry, polling, and Nous authentication across the five subscription providers.

Type and Areas

Type: bug fix

Areas: AI adapters, subscription OAuth, Web UI model settings, documentation and locales.

Motivation / Impact

• Antigravity uses authenticated fetchAvailableModels with endpoint fallback and preserves current server model IDs instead of returning a static seven-model list. An old display name alone does not imply an outdated model, and published models are not assumed to be available to every OAuth account.
• Codex retains subscription-only entries marked supported_in_api: false, respects hidden entries, and reports catalog errors instead of silently using another CLI's local cache.
• OpenCode refreshes offerings by Go/Zen plan and request format, clears obsolete optional catalogs, and propagates explicit refresh errors.
• Codex and xAI respect the earliest JWT/stored expiry. OpenCode, xAI and Hermes share deadline-aware device polling with persistent slow_down backoff.
• Nous Messages/model discovery uses Bearer authentication for its trusted HTTPS endpoint.
• Model pickers support retry/refresh, discard stale responses after editor/device switches, and do not present preset subscription models as account-discovered results. Browser/peer Antigravity login reports the local desktop requirement.

Compared against OpenCode 7c2199d84a5830f70a8250731a42ff958145b4d6 and Hermes Agent 245e48008fa814b3251f50755eb656bd9fb86cb1.

Verification

• cargo test -p openbitfun-ai-adapters --features subscription-auth --lib — 352 passed.
• pnpm --dir src/web-ui exec vitest run src/infrastructure/config/components/modelDiscoveryCoordinator.test.ts src/infrastructure/config/components/subscriptionLoginCoordinator.test.ts src/infrastructure/config/components/ModelSettingsPage.presentation.test.ts — 16 passed across 3 files.
• pnpm run fmt:rs, pnpm run check:core-boundaries, pnpm run i18n:audit — passed.
• pnpm run check:web — passed (canvas, Appearance/theme governance, and TypeScript); generated API types with pnpm --dir src/web-ui run gen:types as build preparation.
• pnpm run lint:web — passed with 0 errors (11 existing warnings). Effect cleanup captures its coordinator instance; the 16 focused frontend tests also pass after this correction.
• git diff --check upstream/main...HEAD and node scripts/check-git-object-sizes.mjs --base upstream/main --head HEAD — passed after rebasing onto the latest upstream main.

Reviewer Notes

AI-assisted; testing level: lightly tested end to end, with focused automated regression coverage. Tests use local HTTP mocks and synthetic credentials; no live subscription login or inference was performed. Real SSH workspaces, remote control, Peer Device Mode and Detached Dispatch were not exercised. Surface-switch invalidation and device polling are covered locally. No desktop build or authenticated UI screenshots were produced.

Manual verification: rebuild the desktop app, sign into each provider, refresh the model picker, verify offerings against that account, and send a short request. Repeat discovery after a transient failure and after switching the active device. Existing saved model choices remain intact; no data migration or credential reset is required.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.