fix: resolve the qwen model catalog from the gateway, not qwen-code

[saucam]

Problem

The model picker rendered empty for the qwen backend, against a gateway serving a dozen models.

The cause is not codeoid's caching. qwen-code's getAvailableModels() is not a catalog fetch — it returns modelRegistry.getModelsForAuthType(currentAuthType), an in-memory registry seeded once at construction from:

• hardcoded QWEN_OAUTH_MODELS (a single entry, coder-model) for qwen-oauth
• the user's modelProviders setting for every other authType

Under authType: "openai" with no modelProviders declared it returns []. Verified against the live SDK:

$ probe against @qwen-code/sdk 0.1.8, authType=openai
RAW: { "subtype": "get_available_models", "models": [] }

qwen-code never contacts the gateway, so there was nothing for codeoid to cache — _cacheModels correctly ignores empty reports. Declaring modelProviders in ~/.qwen/settings.json fixes it, but that is a hand-maintained list that goes stale and has to be repeated on every box.

Change

QwenProvider asks the gateway itself. On the openai path it is an OpenAI-compatible endpoint, so GET /models is authoritative and stays current as the provider adds models.

The qwen-code registry is unioned in rather than replaced: it carries real display labels, and qwen-oauth has a dynamic base URL and no key to present, so HTTP is not an option there. A failed fetch falls back to the registry and logs once — the session still runs, only the picker is poorer.

Everything the endpoint reports is surfaced. The response carries only {id, object, created, owned_by} — no modality field — so filtering out image/audio entries would mean pattern-matching ids, which is the same hardcoded-list problem this removes.

Two supporting fixes:

• normalizeModelCatalog read the wrong field. It looked for name, but the SDK emits label (its handler projects the registry entry down to {id, label, capabilities, contextWindowSize}, also dropping description). Every model rendered as its raw id. name / modelId / availableModels are kept as accepted aliases so a future rename doesn't silently empty the picker again.
• _cacheModels was first-non-empty-wins per daemon lifetime. Providers report on each query-loop build, so a model added to a gateway now appears on the next session instead of after a daemon restart. Empty reports are still ignored so a failed fetch cannot clobber a good catalog, and an unchanged report skips the SQLite write.

Verification

Against the real Bailian token-plan gateway, with modelProviders removed from ~/.qwen/settings.json:

live /models  : 12 → qwen3.7-max, qwen3.7-plus, qwen3.6-flash, glm-5.2,
                deepseek-v4-pro, wan2.7-image, wan2.7-image-pro,
                qwen-audio-3.0-tts-plus, deepseek-v4-flash-0731,
                qwen3.8-max, qwen-audio-3.0-realtime-plus, qwen3.8-flash
unioned       : 12
OK

• bun test — 2411 pass, 19 skip, 0 fail (161 files)
• bun run typecheck — clean
• bun run lint — clean

New coverage: the real SDK 0.1.8 response shape, fetchOpenAiModelCatalog (list shape, trailing-slash base URLs, non-2xx, junk payloads), unionCatalogs (dedupe, label preference, registry-only models, ordering), and both _cacheModels cache semantics.

🤖 Generated with Claude Code