fix: map reasoning controls onto Zhipu BigModel's generation-specific thinking API

[HUAN2022A]

Problem

Structured ingest against a Zhipu BigModel endpoint (open.bigmodel.cn/api/paas/v4) with a GLM thinking model fails with the app's own diagnostic:

Analysis failed: Model produced 15,233 characters of reasoning / chain-of-thought, but no actual response content. This usually means the endpoint hit a thinking-token limit, the model didn't transition from thinking to answering, or the endpoint is misbehaving.

Two causes stack up:

1. Wire layer: buildOpenAiCompatibleBody has no BigModel adapter — thinking.type=disabled is sent for Xiaomi MiMo and DeepSeek only, so reasoning: { mode: "off" } (every structured ingest call) is silently dropped and the GLM model thinks by default.
2. Capability layer: resolveReasoningCapabilities buckets open.bigmodel.cn as a generic custom gateway (auto-only), so a saved off is normalized away before the wire is built — the user has no way to fix it from the UI.

Root cause: Zhipu's thinking API is generation-specific

Per Zhipu's docs (docs.bigmodel.cn/cn/guide/capabilities/thinking, .../guide/models/text/glm-5.3), verified empirically against glm-5.3-flash on /api/paas/v4:

request	glm-5.3-flash behavior
nothing (server default)	~938 chars reasoning_content before content
thinking: {type: "disabled"}	HTTP 200 + error 1210: 该模型始终思考，不支持关闭思考；请使用 low、high 或 max
reasoning_effort: "low"	0 chars reasoning, answers in ~24 completion tokens
thinking: {type: "enabled"} + reasoning_effort	works, effort respected

So a straight copy of the DeepSeek/MiMo adapter would turn ingest failures into hard 400s on GLM-5.3+.

Fix

reasoning-capabilities.ts

• isBigModelEndpoint(endpoint) (shared, endpoint-string form; llm-providers.ts now uses it for its image-support checks too)
• parseGlmVersion / isGlmAtLeast — parse the generation out of model ids (glm-4.7-flash → 4.7, glm-5v-turbo → 5.0, glm-z1-* / non-GLM → null)
• BigModel capability branch: pre-4.5 → auto-only; 4.5..5.1 → auto/off (TOGGLE_LEVELS); 5.2+ → auto/off/low/high/max (GLM_EFFORT_LEVELS; medium and custom budgets are not representable in Zhipu's lineup)

llm-providers.ts — adaptBigModelBody:

• off → thinking: {type: "disabled"} on 4.5..5.2; reasoning_effort: "low" on 5.3+ (which rejects disabled with error 1210)
• low/high/max → thinking: {type: "enabled"} + reasoning_effort on 5.2+
• auto / unsupported modes → body untouched (server default)

Testing

• New unit tests in llm-providers.test.ts (wire shapes per generation, incl. the 1210-avoidance path and generic-gateway isolation) and reasoning-capabilities.test.ts (mode sets per generation, version parsing)
• npx vitest run (mock suite): 132 files / 1884 tests pass
• npm run typecheck and npx vite build pass
• Live endpoint check: the exact body the fixed app sends for ingest (reasoning_effort: "low", temperature: 0.1) returns normal content with zero reasoning content on glm-5.3-flash

Notes

• Scope is the OpenAI-compatible wire (/api/paas/v4), matching the existing Zhipu preset; the Anthropic-compatible endpoint is untouched.
• The endpoint gate is host-based (open.bigmodel.cn), consistent with the DeepSeek/MiMo branches; a GLM model id through a generic gateway keeps auto-only capabilities.